Merge mainline - Aug 12 2024 (#17)

* Merge mainline

* Fix after merge

* Remove CI check

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>

164 files changed, 8874 insertions, 3237 deletions

diff --git a/.devops/llama-server.Dockerfile b/.devops/llama-server.Dockerfile
index b631d580..513e33b3 100644
--- a/.devops/llama-server.Dockerfile
+++ b/.devops/llama-server.Dockerfile
@@ -16,7 +16,7 @@ RUN make -j$(nproc) llama-server
 FROM ubuntu:$UBUNTU_VERSION AS runtime
 
 RUN apt-get update && \
-    apt-get install -y libcurl4-openssl-dev libgomp1
+    apt-get install -y libcurl4-openssl-dev libgomp1 curl
 
 COPY --from=build /app/llama-server /llama-server
 
diff --git a/.devops/nix/package.nix b/.devops/nix/package.nix
index 911c42ec..a87423c7 100644
--- a/.devops/nix/package.nix
+++ b/.devops/nix/package.nix
@@ -126,16 +126,9 @@ let
     ++ optionals useMetalKit [ MetalKit ];
 
   cudaBuildInputs = with cudaPackages; [
-    cuda_cccl.dev # <nv/target>
-
-    # A temporary hack for reducing the closure size, remove once cudaPackages
-    # have stopped using lndir: https://github.com/NixOS/nixpkgs/issues/271792
-    cuda_cudart.dev
-    cuda_cudart.lib
-    cuda_cudart.static
-    libcublas.dev
-    libcublas.lib
-    libcublas.static
+    cuda_cudart
+    cuda_cccl # <nv/target>
+    libcublas
   ];
 
   rocmBuildInputs = with rocmPackages; [
diff --git a/.gitignore b/.gitignore
index 7c7dee0c..5ae03020 100644
--- a/.gitignore
+++ b/.gitignore
@@ -50,6 +50,7 @@ build*
 !docs/build.md
 /libllama.so
 /llama-*
+/vulkan-shaders-gen
 android-ndk-*
 arm_neon.h
 cmake-build-*
@@ -78,7 +79,6 @@ models-mnt
 !models/ggml-vocab-*.gguf*
 
 # Zig
-
 zig-out/
 zig-cache/
 
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 7b95095a..0723e071 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -143,7 +143,8 @@ set(LLAMA_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location o
 # determining _precisely_ which defines are necessary for the llama-config
 # package.
 #
-get_directory_property(GGML_DIR_DEFINES DIRECTORY ggml/src COMPILE_DEFINITIONS)
+get_target_property(GGML_DIRECTORY ggml SOURCE_DIR)
+get_directory_property(GGML_DIR_DEFINES DIRECTORY ${GGML_DIRECTORY} COMPILE_DEFINITIONS)
 get_target_property(GGML_TARGET_DEFINES ggml COMPILE_DEFINITIONS)
 set(GGML_TRANSIENT_DEFINES ${GGML_TARGET_DEFINES} ${GGML_DIR_DEFINES})
 get_target_property(GGML_LINK_LIBRARIES ggml LINK_LIBRARIES)
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
index b688f78e..a9e000e5 100644
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -5,6 +5,7 @@
   - Execute [the full CI locally on your machine](ci/README.md) before publishing
 - Please rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs.
   - The PR template has a series of review complexity checkboxes `[ ]` that [you can mark as](https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/about-task-lists) `[X]` for your convenience
+- Consider allowing write access to your branch for faster review
 - If your PR becomes stale, don't hesitate to ping the maintainers in the comments
 
 # Pull requests (for collaborators)
diff --git a/Makefile b/Makefile
index f77da312..18860cad 100644
--- a/Makefile
+++ b/Makefile
@@ -19,6 +19,7 @@ BUILD_TARGETS = \
 	llama-imatrix \
 	llama-infill \
 	llama-llava-cli \
+	llama-minicpmv-cli\
 	llama-lookahead \
 	llama-lookup \
 	llama-lookup-create \
@@ -528,10 +529,21 @@ ifndef GGML_NO_ACCELERATE
 	endif
 endif # GGML_NO_ACCELERATE
 
+ifdef GGML_MUSA
+	CC := clang
+	CXX := clang++
+	GGML_CUDA := 1
+	MK_CPPFLAGS += -DGGML_USE_MUSA
+endif
+
 ifndef GGML_NO_OPENMP
 	MK_CPPFLAGS += -DGGML_USE_OPENMP
 	MK_CFLAGS   += -fopenmp
 	MK_CXXFLAGS += -fopenmp
+	ifdef GGML_MUSA
+		MK_CPPFLAGS += -I/usr/lib/llvm-10/include/openmp
+		MK_LDFLAGS  += -L/usr/lib/llvm-10/lib
+	endif # GGML_MUSA
 endif # GGML_NO_OPENMP
 
 ifdef GGML_OPENBLAS
@@ -588,15 +600,27 @@ else
 endif # GGML_CUDA_FA_ALL_QUANTS
 
 ifdef GGML_CUDA
-	ifneq ('', '$(wildcard /opt/cuda)')
-		CUDA_PATH ?= /opt/cuda
+	ifdef GGML_MUSA
+		ifneq ('', '$(wildcard /opt/musa)')
+			CUDA_PATH ?= /opt/musa
+		else
+			CUDA_PATH ?= /usr/local/musa
+		endif
+
+		MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include
+		MK_LDFLAGS   += -lmusa -lmublas -lmusart -lpthread -ldl -lrt -L$(CUDA_PATH)/lib -L/usr/lib64
+		MK_NVCCFLAGS += -x musa -mtgpu --cuda-gpu-arch=mp_22
 	else
-		CUDA_PATH ?= /usr/local/cuda
-	endif
+		ifneq ('', '$(wildcard /opt/cuda)')
+			CUDA_PATH ?= /opt/cuda
+		else
+			CUDA_PATH ?= /usr/local/cuda
+		endif
 
-	MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS
-	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
-	MK_NVCCFLAGS += -use_fast_math
+		MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS
+		MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
+		MK_NVCCFLAGS += -use_fast_math
+	endif # GGML_MUSA
 
 	OBJ_GGML += ggml/src/ggml-cuda.o
 	OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
@@ -606,9 +630,11 @@ ifdef LLAMA_FATAL_WARNINGS
 	MK_NVCCFLAGS += -Werror all-warnings
 endif # LLAMA_FATAL_WARNINGS
 
+ifndef GGML_MUSA
 ifndef JETSON_EOL_MODULE_DETECT
 	MK_NVCCFLAGS += --forward-unknown-to-host-compiler
 endif # JETSON_EOL_MODULE_DETECT
+endif # GGML_MUSA
 
 ifdef LLAMA_DEBUG
 	MK_NVCCFLAGS += -lineinfo
@@ -621,8 +647,12 @@ endif # GGML_CUDA_DEBUG
 ifdef GGML_CUDA_NVCC
 	NVCC = $(CCACHE) $(GGML_CUDA_NVCC)
 else
-	NVCC = $(CCACHE) nvcc
-endif #GGML_CUDA_NVCC
+	ifdef GGML_MUSA
+		NVCC = $(CCACHE) mcc
+	else
+		NVCC = $(CCACHE) nvcc
+	endif # GGML_MUSA
+endif # GGML_CUDA_NVCC
 
 ifdef CUDA_DOCKER_ARCH
 	MK_NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
@@ -693,9 +723,15 @@ define NVCC_COMPILE
 	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUDA -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endef # NVCC_COMPILE
 else
+	ifdef GGML_MUSA
+define NVCC_COMPILE
+	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -c $< -o $@
+endef # NVCC_COMPILE
+	else
 define NVCC_COMPILE
 	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endef # NVCC_COMPILE
+	endif # GGML_MUSA
 endif # JETSON_EOL_MODULE_DETECT
 
 ggml/src/ggml-cuda/%.o: \
@@ -859,15 +895,16 @@ ggml/src/ggml-metal-embed.o: \
 	ggml/src/ggml-common.h
 	@echo "Embedding Metal library"
 	@sed -e '/#include "ggml-common.h"/r ggml/src/ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml/src/ggml-metal.metal > ggml/src/ggml-metal-embed.metal
-	$(eval TEMP_ASSEMBLY=$(shell mktemp))
-	@echo ".section __DATA, __ggml_metallib"            >  $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_start"                 >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_start:"                       >> $(TEMP_ASSEMBLY)
-	@echo ".incbin \"ggml/src/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_end"                   >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_end:"                         >> $(TEMP_ASSEMBLY)
-	@$(AS) $(TEMP_ASSEMBLY) -o $@
-	@rm -f ${TEMP_ASSEMBLY}
+	$(eval TEMP_ASSEMBLY=$(shell mktemp -d))
+	@echo ".section __DATA, __ggml_metallib"            >  $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".globl _ggml_metallib_start"                 >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo "_ggml_metallib_start:"                       >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".incbin \"ggml/src/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".globl _ggml_metallib_end"                   >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo "_ggml_metallib_end:"                         >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	$(CC) $(CFLAGS) -c $(TEMP_ASSEMBLY)/ggml-metal-embed.s -o $@
+	@rm -f ${TEMP_ASSEMBLY}/ggml-metal-embed.s
+	@rmdir ${TEMP_ASSEMBLY}
 endif
 endif # GGML_METAL
 
@@ -950,6 +987,7 @@ $(info I CXX:       $(shell $(CXX)  --version | head -n 1))
 ifdef GGML_CUDA
 $(info I NVCC:      $(shell $(NVCC) --version | tail -n 1))
 CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])')
+ifndef GGML_MUSA
 ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
 
 ifndef CUDA_DOCKER_ARCH
@@ -959,6 +997,7 @@ endif # CUDA_POWER_ARCH
 endif # CUDA_DOCKER_ARCH
 
 endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
+endif # GGML_MUSA
 endif # GGML_CUDA
 $(info )
 
@@ -1029,7 +1068,6 @@ ggml/src/iqk/iqk_quantize.o: \
 	ggml/src/iqk/iqk_quantize.cpp \
 	ggml/src/iqk/iqk_quantize.h \
 	ggml/src/ggml-quants.h ggml/src/ggml-common.h ggml/include/ggml.h ggml/src/ggml-impl.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 ifndef GGML_NO_IQKMULMAT
 ggml/src/iqk/iqk_mul_mat.o: \
@@ -1189,6 +1227,7 @@ clean:
 	rm -rvf ggml/*.dll
 	rm -rvf ggml/*.so
 	rm -vrf ggml/src/*.o
+	rm -rvf ggml/src/llamafile/*.o
 	rm -rvf common/build-info.cpp
 	rm -vrf ggml/src/ggml-metal-embed.metal
 	rm -vrf ggml/src/ggml-cuda/*.o
@@ -1435,15 +1474,20 @@ libllava.a: examples/llava/llava.cpp \
 	$(CXX) $(CXXFLAGS) -static -fPIC -c $< -o $@ -Wno-cast-qual
 
 llama-llava-cli: examples/llava/llava-cli.cpp \
-	examples/llava/clip.h \
-	examples/llava/clip.cpp \
+	examples/llava/llava.cpp \
 	examples/llava/llava.h \
+	examples/llava/clip.cpp \
+	examples/llava/clip.h \
+	$(OBJ_ALL)
+	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
+
+llama-minicpmv-cli: examples/llava/minicpmv-cli.cpp \
 	examples/llava/llava.cpp \
+	examples/llava/llava.h \
+	examples/llava/clip.cpp \
+	examples/llava/clip.h \
 	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) -c examples/llava/clip.cpp  -o $(call GET_OBJ_FILE, examples/llava/clip.cpp) -Wno-cast-qual
-	$(CXX) $(CXXFLAGS) -c examples/llava/llava.cpp -o $(call GET_OBJ_FILE, examples/llava/llava.cpp)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $< examples/llava/clip.cpp examples/llava/llava.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) $(call GET_OBJ_FILE, examples/llava/llava.cpp) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
 
 ifeq ($(UNAME_S),Darwin)
 swift: examples/batched.swift
@@ -1589,42 +1633,41 @@ llama-q8dot: pocs/vdot/q8dot.cpp ggml/src/ggml.o \
 # Mark legacy binary targets as .PHONY so that they are always checked.
 .PHONY: main quantize perplexity embedding server
 
+# Define the object file target
+examples/deprecation-warning/deprecation-warning.o: examples/deprecation-warning/deprecation-warning.cpp
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
 # NOTE: We currently will always build the deprecation-warning `main` and `server` binaries to help users migrate.
 #  Eventually we will want to remove these target from building all the time.
-main: examples/deprecation-warning/deprecation-warning.cpp
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+main: examples/deprecation-warning/deprecation-warning.o
+	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
 	@echo "NOTICE: The 'main' binary is deprecated. Please use 'llama-cli' instead."
 
-server: examples/deprecation-warning/deprecation-warning.cpp
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+server: examples/deprecation-warning/deprecation-warning.o
+	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
 	@echo "NOTICE: The 'server' binary is deprecated. Please use 'llama-server' instead."
 
-quantize: examples/deprecation-warning/deprecation-warning.cpp
+quantize: examples/deprecation-warning/deprecation-warning.o
 ifneq (,$(wildcard quantize))
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
 	@echo "#########"
 	@echo "WARNING: The 'quantize' binary is deprecated. Please use 'llama-quantize' instead."
 	@echo "  Remove the 'quantize' binary to remove this warning."
 	@echo "#########"
 endif
 
-perplexity: examples/deprecation-warning/deprecation-warning.cpp
+perplexity: examples/deprecation-warning/deprecation-warning.o
 ifneq (,$(wildcard perplexity))
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
 	@echo "#########"
 	@echo "WARNING: The 'perplexity' binary is deprecated. Please use 'llama-perplexity' instead."
 	@echo "  Remove the 'perplexity' binary to remove this warning."
 	@echo "#########"
 endif
 
-embedding: examples/deprecation-warning/deprecation-warning.cpp
+embedding: examples/deprecation-warning/deprecation-warning.o
 ifneq (,$(wildcard embedding))
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
 	@echo "#########"
 	@echo "WARNING: The 'embedding' binary is deprecated. Please use 'llama-embedding' instead."
 	@echo "  Remove the 'embedding' binary to remove this warning."
diff --git a/common/common.cpp b/common/common.cpp
index 4d1d88c6..3b45d066 100644
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -684,14 +684,24 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
     }
     if (arg == "--lora") {
         CHECK_ARG
-        params.lora_adapter.emplace_back(argv[i], 1.0f);
+        params.lora_adapters.push_back({
+            std::string(argv[i]),
+            1.0,
+        });
         return true;
     }
     if (arg == "--lora-scaled") {
         CHECK_ARG
-        const char* lora_adapter = argv[i];
+        std::string lora_adapter = argv[i];
         CHECK_ARG
-        params.lora_adapter.emplace_back(lora_adapter, std::stof(argv[i]));
+        params.lora_adapters.push_back({
+            lora_adapter,
+            std::stof(argv[i]),
+        });
+        return true;
+    }
+    if (arg == "--lora-init-without-apply") {
+        params.lora_init_without_apply = true;
         return true;
     }
     if (arg == "--control-vector") {
@@ -1332,6 +1342,10 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         else { invalid_param = true; }
         return true;
     }
+    if (arg == "--no-warmup") {
+        params.warmup = false;
+        return true;
+    }
 #ifndef LOG_DISABLE_LOGS
     // Parse args for logging parameters
     if (log_param_single_parse(argv[i])) {
@@ -1454,6 +1468,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
     options.push_back({ "main infill", "       --in-prefix-bos",        "prefix BOS to user inputs, preceding the `--in-prefix` string" });
     options.push_back({ "main infill", "       --in-prefix STRING",     "string to prefix user inputs with (default: empty)" });
     options.push_back({ "main infill", "       --in-suffix STRING",     "string to suffix after user inputs with (default: empty)" });
+    options.push_back({ "main",        "       --no-warmup",            "skip warming up the model with an empty run" });
     options.push_back({ "server infill",
                                        "       --spm-infill",           "use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: %s)", params.spm_infill ? "enabled" : "disabled" });
 
@@ -1637,7 +1652,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
     options.push_back({ "server",      "       --host HOST",            "ip address to listen (default: %s)", params.hostname.c_str() });
     options.push_back({ "server",      "       --port PORT",            "port to listen (default: %d)", params.port });
     options.push_back({ "server",      "       --path PATH",            "path to serve static files from (default: %s)", params.public_path.c_str() });
-    options.push_back({ "server",      "       --embedding(s)",         "enable embedding endpoint (default: %s)", params.embedding ? "enabled" : "disabled" });
+    options.push_back({ "server",      "       --embedding(s)",         "restrict to only support embedding use case; use only with dedicated embedding models (default: %s)", params.embedding ? "enabled" : "disabled" });
     options.push_back({ "server",      "       --api-key KEY",          "API key to use for authentication (default: none)" });
     options.push_back({ "server",      "       --api-key-file FNAME",   "path to file containing API keys (default: none)" });
     options.push_back({ "server",      "       --ssl-key-file FNAME",   "path to file a PEM-encoded SSL private key" });
@@ -1657,6 +1672,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
                                                                         "https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template" });
     options.push_back({ "server",      "-sps,  --slot-prompt-similarity SIMILARITY",
                                                                         "how much the prompt of a request must match the prompt of a slot in order to use that slot (default: %.2f, 0.0 = disabled)\n", params.slot_prompt_similarity });
+    options.push_back({ "server",      "       --lora-init-without-apply",     "load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: %s)", params.lora_init_without_apply ? "enabled" : "disabled"});
 
 #ifndef LOG_DISABLE_LOGS
     options.push_back({ "logging" });
@@ -1769,6 +1785,17 @@ std::string string_get_sortable_timestamp() {
     return std::string(timestamp_no_ns) + "." + std::string(timestamp_ns);
 }
 
+void string_replace_all(std::string & s, const std::string & search, const std::string & replace) {
+    if (search.empty()) {
+        return; // Avoid infinite loop if 'search' is an empty string
+    }
+    size_t pos = 0;
+    while ((pos = s.find(search, pos)) != std::string::npos) {
+        s.replace(pos, search.length(), replace);
+        pos += replace.length();
+    }
+}
+
 void string_process_escapes(std::string & input) {
     std::size_t input_len = input.length();
     std::size_t output_idx = 0;
@@ -2042,8 +2069,8 @@ std::string fs_get_cache_file(const std::string & filename) {
 //
 // Model utils
 //
-
-std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params) {
+struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
+    llama_init_result iparams;
     auto mparams = llama_model_params_from_gpt_params(params);
 
     llama_model * model = nullptr;
@@ -2058,7 +2085,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
 
     if (model == NULL) {
         fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
-        return std::make_tuple(nullptr, nullptr);
+        return iparams;
     }
 
     auto cparams = llama_context_params_from_gpt_params(params);
@@ -2067,7 +2094,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
     if (lctx == NULL) {
         fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str());
         llama_free_model(model);
-        return std::make_tuple(nullptr, nullptr);
+        return iparams;
     }
 
     if (!params.control_vectors.empty()) {
@@ -2078,7 +2105,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
         if (cvec.n_embd == -1) {
             llama_free(lctx);
             llama_free_model(model);
-            return std::make_tuple(nullptr, nullptr);
+            return iparams;
         }
 
         int err = llama_control_vector_apply(lctx,
@@ -2090,21 +2117,26 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
         if (err) {
             llama_free(lctx);
             llama_free_model(model);
-            return std::make_tuple(nullptr, nullptr);
+            return iparams;
         }
     }
 
-    for (unsigned int i = 0; i < params.lora_adapter.size(); ++i) {
-        const std::string & lora_adapter = std::get<0>(params.lora_adapter[i]);
-        float lora_scale = std::get<1>(params.lora_adapter[i]);
-        auto adapter = llama_lora_adapter_init(model, lora_adapter.c_str());
-        if (adapter == nullptr) {
-            fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
+    // load and optionally apply lora adapters
+    for (auto & la : params.lora_adapters) {
+        llama_lora_adapter_container loaded_la;
+        loaded_la.path = la.path;
+        loaded_la.scale = la.scale;
+        loaded_la.adapter = llama_lora_adapter_init(model, la.path.c_str());
+        if (loaded_la.adapter == nullptr) {
+            fprintf(stderr, "%s: error: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
             llama_free(lctx);
             llama_free_model(model);
-            return std::make_tuple(nullptr, nullptr);
+            return iparams;
         }
-        llama_lora_adapter_set(lctx, adapter, lora_scale);
+        iparams.lora_adapters.push_back(loaded_la); // copy to list of loaded adapters
+    }
+    if (!params.lora_init_without_apply) {
+        llama_lora_adapters_apply(lctx, iparams.lora_adapters);
     }
 
     if (params.ignore_eos) {
@@ -2132,13 +2164,26 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
             tmp.clear();
             tmp.push_back(decoder_start_token_id);
         }
-        llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
+        if (llama_model_has_decoder(model)) {
+            llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
+        }
         llama_kv_cache_clear(lctx);
         llama_synchronize(lctx);
         llama_reset_timings(lctx);
     }
 
-    return std::make_tuple(model, lctx);
+    iparams.model   = model;
+    iparams.context = lctx;
+    return iparams;
+}
+
+void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lora_adapter_container> & lora_adapters) {
+    llama_lora_adapter_clear(ctx);
+    for (auto & la : lora_adapters) {
+        if (la.scale != 0.0f) {
+            llama_lora_adapter_set(ctx, la.adapter, la.scale);
+        }
+    }
 }
 
 struct llama_model_params llama_model_params_from_gpt_params(const gpt_params & params) {
@@ -3163,19 +3208,18 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l
     }
 
     fprintf(stream, "lora:\n");
-    for (std::tuple<std::string, float> la : params.lora_adapter) {
-        if (std::get<1>(la) != 1.0f) {
-            continue;
+    for (auto & la : params.lora_adapters) {
+        if (la.scale == 1.0f) {
+            fprintf(stream, "  - %s\n", la.path.c_str());
         }
-        fprintf(stream, "  - %s\n", std::get<0>(la).c_str());
     }
     fprintf(stream, "lora_scaled:\n");
-    for (std::tuple<std::string, float> la : params.lora_adapter) {
-        if (std::get<1>(la) == 1.0f) {
-            continue;
+    for (auto & la : params.lora_adapters) {
+        if (la.scale != 1.0f) {
+            fprintf(stream, "  - %s: %f\n", la.path.c_str(), la.scale);
         }
-        fprintf(stream, "  - %s: %f\n", std::get<0>(la).c_str(), std::get<1>(la));
     }
+    fprintf(stream, "lora_init_without_apply: %s # default: false\n", params.lora_init_without_apply ? "true" : "false");
     fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
     fprintf(stream, "min_keep: %d # default: 0 (disabled)\n", sparams.min_keep);
     fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat);
diff --git a/common/common.h b/common/common.h
index 979762e1..50035897 100644
--- a/common/common.h
+++ b/common/common.h
@@ -33,6 +33,15 @@
 
 #define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
 
+struct llama_lora_adapter_info {
+    std::string path;
+    float scale;
+};
+
+struct llama_lora_adapter_container : llama_lora_adapter_info {
+    struct llama_lora_adapter * adapter;
+};
+
 // build info
 extern int LLAMA_BUILD_NUMBER;
 extern char const * LLAMA_COMMIT;
@@ -126,8 +135,8 @@ struct gpt_params {
     std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
     std::vector<llama_model_kv_override> kv_overrides;
 
-    // TODO: avoid tuple, use struct
-    std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
+    bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_lora_adapter_apply)
+    std::vector<llama_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale
 
     std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale
 
@@ -278,6 +287,8 @@ std::vector<std::string> string_split(std::string input, char separator);
 std::string string_strip(const std::string & str);
 std::string string_get_sortable_timestamp();
 
+void string_replace_all(std::string & s, const std::string & search, const std::string & replace);
+
 template<class T>
 static std::vector<T> string_split(const std::string & str, char delim) {
     std::vector<T> values;
@@ -309,8 +320,13 @@ std::string fs_get_cache_file(const std::string & filename);
 // Model utils
 //
 
-// TODO: avoid tuplue, use struct
-std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params);
+struct llama_init_result {
+    struct llama_model   * model   = nullptr;
+    struct llama_context * context = nullptr;
+    std::vector<llama_lora_adapter_container> lora_adapters;
+};
+
+struct llama_init_result    llama_init_from_gpt_params(gpt_params & params);
 
 struct llama_model_params   llama_model_params_from_gpt_params  (const gpt_params & params);
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
@@ -318,6 +334,9 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
 struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const char * hf_token, const struct llama_model_params & params);
 struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const char * hf_token, const struct llama_model_params & params);
 
+// clear LoRA adapters from context, then apply new list of adapters
+void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lora_adapter_container> & lora_adapters);
+
 // Batch utils
 
 void llama_batch_clear(struct llama_batch & batch);
diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index 7a74cc20..b470a088 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -251,12 +251,7 @@ class Model:
 
         return [(self.map_tensor_name(name), data_torch)]
 
-    def extra_f32_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
-        del name, new_name, bid, n_dims  # unused
-
-        return False
-
-    def extra_f16_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
+    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
         del name, new_name, bid, n_dims  # unused
 
         return False
@@ -285,54 +280,46 @@ class Model:
             for new_name, data in ((n, d.squeeze().numpy()) for n, d in self.modify_tensors(data_torch, name, bid)):
                 data: np.ndarray  # type hint
                 n_dims = len(data.shape)
-                data_dtype = data.dtype
-                data_qtype: gguf.GGMLQuantizationType | None = None
-
-                # when both are True, f32 should win
-                extra_f32 = self.extra_f32_tensors(name, new_name, bid, n_dims)
-                extra_f16 = self.extra_f16_tensors(name, new_name, bid, n_dims)
+                data_qtype: gguf.GGMLQuantizationType | bool = self.tensor_force_quant(name, new_name, bid, n_dims)
 
                 # Most of the codebase that takes in 1D tensors or norms only handles F32 tensors
-                # Conditions should closely match those in llama_model_quantize_internal in llama.cpp
-                extra_f32 = any(cond for cond in (
-                    extra_f32,
-                    n_dims == 1,
-                    new_name.endswith("_norm.weight"),
-                ))
+                if n_dims <= 1 or new_name.endswith("_norm.weight"):
+                    data_qtype = gguf.GGMLQuantizationType.F32
 
+                # Conditions should closely match those in llama_model_quantize_internal in llama.cpp
                 # Some tensor types are always in float32
-                extra_f32 = extra_f32 or any(self.match_model_tensor_name(new_name, key, bid) for key in (
-                    gguf.MODEL_TENSOR.FFN_GATE_INP,
-                    gguf.MODEL_TENSOR.POS_EMBD,
-                    gguf.MODEL_TENSOR.TOKEN_TYPES,
-                ))
-
-                # if f16 desired, convert any float32 2-dim weight tensors to float16
-                extra_f16 = any(cond for cond in (
-                    extra_f16,
-                    (name.endswith(".weight") and n_dims >= 2),
-                ))
-
-                if self.ftype != gguf.LlamaFileType.ALL_F32 and extra_f16 and not extra_f32:
-                    if self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
-                        data = gguf.quantize_bf16(data)
-                        assert data.dtype == np.int16
-                        data_qtype = gguf.GGMLQuantizationType.BF16
-
-                    elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0 and gguf.can_quantize_to_q8_0(data):
-                        data = gguf.quantize_q8_0(data)
-                        assert data.dtype == np.uint8
-                        data_qtype = gguf.GGMLQuantizationType.Q8_0
+                if data_qtype is False and (
+                    any(
+                        self.match_model_tensor_name(new_name, key, bid)
+                        for key in (
+                            gguf.MODEL_TENSOR.FFN_GATE_INP,
+                            gguf.MODEL_TENSOR.POS_EMBD,
+                            gguf.MODEL_TENSOR.TOKEN_TYPES,
+                        )
+                    )
+                    or not name.endswith(".weight")
+                ):
+                    data_qtype = gguf.GGMLQuantizationType.F32
 
-                    else:  # default to float16 for quantized tensors
-                        if data_dtype != np.float16:
-                            data = data.astype(np.float16)
+                # No override (data_qtype is False), or wants to be quantized (data_qtype is True)
+                if isinstance(data_qtype, bool):
+                    if self.ftype == gguf.LlamaFileType.ALL_F32:
+                        data_qtype = gguf.GGMLQuantizationType.F32
+                    elif self.ftype == gguf.LlamaFileType.MOSTLY_F16:
                         data_qtype = gguf.GGMLQuantizationType.F16
+                    elif self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
+                        data_qtype = gguf.GGMLQuantizationType.BF16
+                    elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0:
+                        data_qtype = gguf.GGMLQuantizationType.Q8_0
+                    else:
+                        raise ValueError(f"Unknown file type: {self.ftype.name}")
 
-                if data_qtype is None:  # by default, convert to float32
-                    if data_dtype != np.float32:
-                        data = data.astype(np.float32)
-                    data_qtype = gguf.GGMLQuantizationType.F32
+                try:
+                    data = gguf.quants.quantize(data, data_qtype)
+                except gguf.QuantError as e:
+                    logger.warning("%s, %s", e, "falling back to F16")
+                    data_qtype = gguf.GGMLQuantizationType.F16
+                    data = gguf.quants.quantize(data, data_qtype)
 
                 shape = gguf.quant_shape_from_byte_shape(data.shape, data_qtype) if data.dtype == np.uint8 else data.shape
 
@@ -1570,6 +1557,34 @@ class LlamaModel(Model):
         return [(self.map_tensor_name(name), data_torch)]
 
     def prepare_tensors(self):
+        if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
+            if rope_scaling.get("rope_type", '').lower() == "llama3":
+                base = self.hparams.get("rope_theta", 10000.0)
+                dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+                freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
+
+                factor = rope_scaling.get("factor", 8.0)
+                low_freq_factor = rope_scaling.get("low_freq_factor", 1.0)
+                high_freq_factor = rope_scaling.get("high_freq_factor", 4.0)
+                old_context_len = self.hparams.get("original_max_position_embeddings", 8192)
+
+                low_freq_wavelen = old_context_len / low_freq_factor
+                high_freq_wavelen = old_context_len / high_freq_factor
+                assert low_freq_wavelen != high_freq_wavelen
+
+                rope_factors = []
+                for freq in freqs:
+                    wavelen = 2 * math.pi / freq
+                    if wavelen < high_freq_wavelen:
+                        rope_factors.append(1)
+                    elif wavelen > low_freq_wavelen:
+                        rope_factors.append(factor)
+                    else:
+                        smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
+                        rope_factors.append(1 / ((1 - smooth) / factor + smooth))
+
+                self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32))
+
         super().prepare_tensors()
 
         if self._experts is not None:
@@ -1754,7 +1769,7 @@ class DbrxModel(Model):
 
         return [(new_name, data_torch)]
 
-    def extra_f16_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
+    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
         del name, new_name, bid  # unused
 
         return n_dims > 1
@@ -2495,6 +2510,112 @@ class NomicBertModel(BertModel):
         self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
 
 
+@Model.register("XLMRobertaModel")
+class XLMRobertaModel(BertModel):
+    model_arch = gguf.MODEL_ARCH.BERT
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # we need the pad_token_id to know how to chop down position_embd matrix
+        if (pad_token_id := self.hparams.get("pad_token_id")) is not None:
+            self._position_offset = 1 + pad_token_id
+            if "max_position_embeddings" in self.hparams:
+                self.hparams["max_position_embeddings"] -= self._position_offset
+        else:
+            self._position_offset = None
+
+    def set_vocab(self):
+        # to avoid TypeError: Descriptors cannot be created directly
+        # exception when importing sentencepiece_model_pb2
+        os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
+        from sentencepiece import SentencePieceProcessor
+        from sentencepiece import sentencepiece_model_pb2 as model
+
+        tokenizer_path = self.dir_model / 'sentencepiece.bpe.model'
+        if not tokenizer_path.is_file():
+            raise FileNotFoundError(f"File not found: {tokenizer_path}")
+
+        sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue]
+        sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
+        assert sentencepiece_model.trainer_spec.model_type == 1  # UNIGRAM
+
+        add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix
+        remove_whitespaces = sentencepiece_model.normalizer_spec.remove_extra_whitespaces
+        precompiled_charsmap = sentencepiece_model.normalizer_spec.precompiled_charsmap
+
+        tokenizer = SentencePieceProcessor()
+        tokenizer.LoadFromFile(str(tokenizer_path))
+
+        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+
+        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
+        scores: list[float] = [-10000.0] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
+
+        for token_id in range(tokenizer.vocab_size()):
+            piece = tokenizer.IdToPiece(token_id)
+            text = piece.encode("utf-8")
+            score = tokenizer.GetScore(token_id)
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.IsUnknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.IsControl(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.IsUnused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.IsByte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+
+            tokens[token_id] = text
+            scores[token_id] = score
+            toktypes[token_id] = toktype
+
+        if vocab_size > len(tokens):
+            pad_count = vocab_size - len(tokens)
+            logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
+            for i in range(1, pad_count + 1):
+                tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
+                scores.append(-1000.0)
+                toktypes.append(SentencePieceTokenTypes.UNUSED)
+
+        # realign tokens (see HF tokenizer code)
+        tokens = [b'<s>', b'<pad>', b'</s>', b'<unk>'] + tokens[3:-1]
+        scores = [0.0, 0.0, 0.0, 0.0] + scores[3:-1]
+        toktypes = [
+            SentencePieceTokenTypes.CONTROL,
+            SentencePieceTokenTypes.CONTROL,
+            SentencePieceTokenTypes.CONTROL,
+            SentencePieceTokenTypes.UNKNOWN,
+        ] + toktypes[3:-1]
+
+        self.gguf_writer.add_tokenizer_model("t5")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+        self.gguf_writer.add_add_space_prefix(add_prefix)
+        self.gguf_writer.add_token_type_count(1)
+        self.gguf_writer.add_remove_extra_whitespaces(remove_whitespaces)
+        if precompiled_charsmap:
+            self.gguf_writer.add_precompiled_charsmap(precompiled_charsmap)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+        self.gguf_writer.add_add_bos_token(True)
+        self.gguf_writer.add_add_eos_token(True)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # position embeddings start at pad_token_id + 1, so just chop down the weight tensor
+        if name == "embeddings.position_embeddings.weight":
+            if self._position_offset is not None:
+                data_torch = data_torch[self._position_offset:,:]
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @Model.register("GemmaForCausalLM")
 class GemmaModel(Model):
     model_arch = gguf.MODEL_ARCH.GEMMA
@@ -2669,18 +2790,22 @@ class MambaModel(Model):
 
         return [(new_name, data_torch)]
 
-    def extra_f32_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
-        del n_dims  # unused
-
-        return bid is not None and new_name in (
-            self.format_tensor_name(n, bid, ".weight" if name.endswith(".weight") else "") for n in [
+    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
+        if bid is not None and new_name in (
+            self.format_tensor_name(
+                n, bid, ".weight" if name.endswith(".weight") else ""
+            )
+            for n in [
                 gguf.MODEL_TENSOR.SSM_CONV1D,
                 gguf.MODEL_TENSOR.SSM_X,
                 gguf.MODEL_TENSOR.SSM_DT,
                 gguf.MODEL_TENSOR.SSM_A,
                 gguf.MODEL_TENSOR.SSM_D,
             ]
-        )
+        ):
+            return gguf.GGMLQuantizationType.F32
+
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
 
 
 @Model.register("CohereForCausalLM")
@@ -3216,6 +3341,145 @@ class T5Model(Model):
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@Model.register("T5EncoderModel")
+class T5EncoderModel(Model):
+    model_arch = gguf.MODEL_ARCH.T5ENCODER
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.shared_token_embeddings_found = False
+
+    def set_vocab(self):
+        # to avoid TypeError: Descriptors cannot be created directly
+        # exception when importing sentencepiece_model_pb2
+        os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
+        from sentencepiece import SentencePieceProcessor
+        from sentencepiece import sentencepiece_model_pb2 as model
+
+        tokenizer_path = self.dir_model / 'tokenizer.model'
+
+        # many older models use spiece.model tokenizer model filename
+        if not tokenizer_path.is_file():
+            tokenizer_path = self.dir_model / 'spiece.model'
+
+        if not tokenizer_path.is_file():
+            raise FileNotFoundError(f"File not found: {tokenizer_path}")
+
+        sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue]
+        sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
+
+        # some models like Pile-T5 family use BPE tokenizer instead of Unigram
+        if sentencepiece_model.trainer_spec.model_type == 2:  # BPE
+            # assure the tokenizer model file name is correct
+            assert tokenizer_path.name == 'tokenizer.model'
+            return self._set_vocab_sentencepiece()
+        else:
+            assert sentencepiece_model.trainer_spec.model_type == 1  # UNIGRAM
+
+        add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix
+        remove_whitespaces = sentencepiece_model.normalizer_spec.remove_extra_whitespaces
+        precompiled_charsmap = sentencepiece_model.normalizer_spec.precompiled_charsmap
+
+        tokenizer = SentencePieceProcessor()
+        tokenizer.LoadFromFile(str(tokenizer_path))
+
+        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+
+        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
+        scores: list[float] = [-10000.0] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
+
+        for token_id in range(tokenizer.vocab_size()):
+            piece = tokenizer.IdToPiece(token_id)
+            text = piece.encode("utf-8")
+            score = tokenizer.GetScore(token_id)
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.IsUnknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.IsControl(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.IsUnused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.IsByte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+
+            tokens[token_id] = text
+            scores[token_id] = score
+            toktypes[token_id] = toktype
+
+        added_tokens_file = self.dir_model / 'added_tokens.json'
+        if added_tokens_file.is_file():
+            with open(added_tokens_file, "r", encoding="utf-8") as f:
+                added_tokens_json = json.load(f)
+                for key in added_tokens_json:
+                    token_id = added_tokens_json[key]
+                    if token_id >= vocab_size:
+                        logger.warning(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
+                        continue
+
+                    tokens[token_id] = key.encode("utf-8")
+                    scores[token_id] = -1000.0
+                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
+
+        if vocab_size > len(tokens):
+            pad_count = vocab_size - len(tokens)
+            logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
+            for i in range(1, pad_count + 1):
+                tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
+                scores.append(-1000.0)
+                toktypes.append(SentencePieceTokenTypes.UNUSED)
+
+        self.gguf_writer.add_tokenizer_model("t5")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+        self.gguf_writer.add_add_space_prefix(add_prefix)
+        self.gguf_writer.add_remove_extra_whitespaces(remove_whitespaces)
+        if precompiled_charsmap:
+            self.gguf_writer.add_precompiled_charsmap(precompiled_charsmap)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+        self.gguf_writer.add_add_bos_token(False)
+        self.gguf_writer.add_add_eos_token(True)
+
+    def set_gguf_parameters(self):
+        if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None:
+            logger.warning("Couldn't find context length in config.json, assuming default value of 512")
+            n_ctx = 512
+        self.gguf_writer.add_context_length(n_ctx)
+        self.gguf_writer.add_embedding_length(self.hparams["d_model"])
+        self.gguf_writer.add_feed_forward_length(self.hparams["d_ff"])
+        self.gguf_writer.add_block_count(self.hparams["num_layers"])
+        self.gguf_writer.add_head_count(self.hparams["num_heads"])
+        self.gguf_writer.add_key_length(self.hparams["d_kv"])
+        self.gguf_writer.add_value_length(self.hparams["d_kv"])
+        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_relative_attn_buckets_count(self.hparams["relative_attention_num_buckets"])
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        # T5 based models contain shared token embeddings tensors saved randomly as either "encoder.embed_tokens.weight",
+        # "decoder.embed_tokens.weight" or "shared.weight" tensor. In some models there are even multiple of them stored
+        # in the safetensors files. We use the first tensor from these three as the token embeddings for both encoder
+        # and decoder and ignore the remaining ones.
+        if name in ["decoder.embed_tokens.weight", "encoder.embed_tokens.weight", "shared.weight"]:
+            if not self.shared_token_embeddings_found:
+                name = "shared.weight"
+                self.shared_token_embeddings_found = True
+            else:
+                logger.debug(f"Skipping shared tensor {name!r} in safetensors so that convert can end normally.")
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
 @Model.register("JAISLMHeadModel")
 class JaisModel(Model):
     model_arch = gguf.MODEL_ARCH.JAIS
diff --git a/docs/backend/SYCL.md b/docs/backend/SYCL.md
index d36ac0a1..59a39fbb 100644
--- a/docs/backend/SYCL.md
+++ b/docs/backend/SYCL.md
@@ -80,7 +80,14 @@ The following release is verified with good quality:
 
 ### Intel GPU
 
-**Verified devices**
+SYCL backend supports Intel GPU Family:
+
+- Intel Data Center Max Series
+- Intel Flex Series, Arc Series
+- Intel Built-in Arc GPU
+- Intel iGPU in Core CPU (11th Generation Core CPU and newer, refer to [oneAPI supported GPU](https://www.intel.com/content/www/us/en/developer/articles/system-requirements/intel-oneapi-base-toolkit-system-requirements.html#inpage-nav-1-1)).
+
+#### Verified devices
 
 | Intel GPU                     | Status  | Verified Model                        |
 |-------------------------------|---------|---------------------------------------|
@@ -88,7 +95,7 @@ The following release is verified with good quality:
 | Intel Data Center Flex Series | Support | Flex 170                              |
 | Intel Arc Series              | Support | Arc 770, 730M, Arc A750               |
 | Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake       |
-| Intel iGPU                    | Support | iGPU in i5-1250P, i7-1260P, i7-1165G7 |
+| Intel iGPU                    | Support | iGPU in 13700k, i5-1250P, i7-1260P, i7-1165G7 |
 
 *Notes:*
 
@@ -237,6 +244,13 @@ Similarly, user targeting Nvidia GPUs should expect at least one SYCL-CUDA devic
 ### II. Build llama.cpp
 
 #### Intel GPU
+
+```
+./examples/sycl/build.sh
+```
+
+or
+
 ```sh
 # Export relevant ENV variables
 source /opt/intel/oneapi/setvars.sh
@@ -276,23 +290,26 @@ cmake --build build --config Release -j -v
 
 ### III. Run the inference
 
-1. Retrieve and prepare model
+#### Retrieve and prepare model
 
 You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
 
-2. Enable oneAPI running environment
+##### Check device
+
+1. Enable oneAPI running environment
 
 ```sh
 source /opt/intel/oneapi/setvars.sh
 ```
 
-3. List devices information
+2. List devices information
 
 Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
 
 ```sh
 ./build/bin/llama-ls-sycl-device
 ```
+
 This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following:
 ```
 found 2 SYCL devices:
@@ -304,12 +321,37 @@ found 2 SYCL devices:
 | 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
 ```
 
+#### Choose level-zero devices
+
+|Chosen Device ID|Setting|
+|-|-|
+|0|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action|
+|1|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
+|0 & 1|`export ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`|
+
+#### Execute
+
+Choose one of following methods to run.
+
+1. Script
+
+- Use device 0:
+
+```sh
+./examples/sycl/run_llama2.sh 0
+```
+- Use multiple devices:
+
+```sh
+./examples/sycl/run_llama2.sh
+```
 
-4. Launch inference
+2. Command line
+Launch inference
 
 There are two device selection modes:
 
-- Single device: Use one device target specified by the user.
+- Single device: Use one device assigned by user. Default device id is 0.
 - Multiple devices: Automatically choose the devices with the same backend.
 
 In two device selection modes, the default SYCL backend is level_zero, you can choose other backend supported by SYCL by setting environment variable ONEAPI_DEVICE_SELECTOR.
@@ -326,11 +368,6 @@ Examples:
 ```sh
 ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
 ```
-or run by script:
-
-```sh
-./examples/sycl/run_llama2.sh 0
-```
 
 - Use multiple devices:
 
@@ -338,12 +375,6 @@ or run by script:
 ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
 ```
 
-Otherwise, you can run the script:
-
-```sh
-./examples/sycl/run_llama2.sh
-```
-
 *Notes:*
 
 - Upon execution, verify the selected device(s) ID(s) in the output log, which can for instance be displayed as follow:
@@ -390,7 +421,7 @@ c. Verify installation
 In the oneAPI command line, run the following to print the available SYCL devices:
 
 ```
-sycl-ls
+sycl-ls.exe
 ```
 
 There should be one or more *level-zero* GPU devices displayed as **[ext_oneapi_level_zero:gpu]**. Below is example of such output detecting an *intel Iris Xe* GPU as a Level-zero SYCL device:
@@ -411,6 +442,18 @@ b. The new Visual Studio will install Ninja as default. (If not, please install
 
 ### II. Build llama.cpp
 
+You could download the release package for Windows directly, which including binary files and depended oneAPI dll files.
+
+Choose one of following methods to build from source code.
+
+1. Script
+
+```sh
+.\examples\sycl\win-build-sycl.bat
+```
+
+2. CMake
+
 On the oneAPI command line window, step into the llama.cpp main directory and run the following:
 
 ```
@@ -425,12 +468,8 @@ cmake -B build -G "Ninja" -DGGML_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPI
 cmake --build build --config Release -j
 ```
 
-Otherwise, run the `win-build-sycl.bat` wrapper which encapsulates the former instructions:
-```sh
-.\examples\sycl\win-build-sycl.bat
-```
-
 Or, use CMake presets to build:
+
 ```sh
 cmake --preset x64-windows-sycl-release
 cmake --build build-x64-windows-sycl-release -j --target llama-cli
@@ -442,7 +481,9 @@ cmake --preset x64-windows-sycl-debug
 cmake --build build-x64-windows-sycl-debug -j --target llama-cli
 ```
 
-Or, you can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project.
+3. Visual Studio
+
+You can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project.
 
 *Notes:*
 
@@ -450,23 +491,25 @@ Or, you can use Visual Studio to open llama.cpp folder as a CMake project. Choos
 
 ### III. Run the inference
 
-1. Retrieve and prepare model
+#### Retrieve and prepare model
 
-You can refer to the general [*Prepare and Quantize*](README#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
+You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
 
-2. Enable oneAPI running environment
+##### Check device
+
+1. Enable oneAPI running environment
 
 On the oneAPI command line window, run the following and step into the llama.cpp directory:
 ```
 "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64
 ```
 
-3. List devices information
+2. List devices information
 
 Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
 
 ```
-build\bin\ls-sycl-device.exe
+build\bin\llama-ls-sycl-device.exe
 ```
 
 This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following:
@@ -479,9 +522,27 @@ found 2 SYCL devices:
 | 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
 
 ```
+#### Choose level-zero devices
+
+|Chosen Device ID|Setting|
+|-|-|
+|0|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action|
+|1|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
+|0 & 1|`set ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`|
+
+#### Execute
+
+Choose one of following methods to run.
+
+1. Script
+
+```
+examples\sycl\win-run-llama2.bat
+```
 
+2. Command line
 
-4. Launch inference
+Launch inference
 
 There are two device selection modes:
 
@@ -508,11 +569,7 @@ build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website ca
 ```
 build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer
 ```
-Otherwise, run the following wrapper script:
 
-```
-.\examples\sycl\win-run-llama2.bat
-```
 
 Note:
 
@@ -526,17 +583,18 @@ Or
 use 1 SYCL GPUs: [0] with Max compute units:512
 ```
 
+
 ## Environment Variable
 
 #### Build
 
 | Name               | Value                             | Function                                    |
 |--------------------|-----------------------------------|---------------------------------------------|
-| GGML_SYCL          | ON (mandatory)                    | Enable build with SYCL code path.           |
+| GGML_SYCL          | ON (mandatory)                    | Enable build with SYCL code path.<br>FP32 path - recommended for better perforemance than FP16 on quantized model|
 | GGML_SYCL_TARGET   | INTEL *(default)* \| NVIDIA       | Set the SYCL target device type.            |
 | GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)* | Enable FP16 build with SYCL code path.      |
-| CMAKE_C_COMPILER   | icx                               | Set *icx* compiler for SYCL code path.      |
-| CMAKE_CXX_COMPILER | icpx *(Linux)*, icx *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
+| CMAKE_C_COMPILER   | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path.      |
+| CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
 
 #### Runtime
 
@@ -572,9 +630,18 @@ use 1 SYCL GPUs: [0] with Max compute units:512
   ```
   Otherwise, please double-check the GPU driver installation steps.
 
+- Can I report Ollama issue on Intel GPU to llama.cpp SYCL backend?
+
+  No. We can't support Ollama issue directly, because we aren't familiar with Ollama.
+
+  Sugguest reproducing on llama.cpp and report similar issue to llama.cpp. We will surpport it.
+
+  It's same for other projects including llama.cpp SYCL backend.
+
+
 ### **GitHub contribution**:
 Please add the **[SYCL]** prefix/tag in issues/PRs titles to help the SYCL-team check/address them without delay.
 
 ## TODO
 
-- Support row layer split for multiple card runs.
+- NA
diff --git a/docs/build.md b/docs/build.md
index d9d12c46..8b16d1a3 100644
--- a/docs/build.md
+++ b/docs/build.md
@@ -178,7 +178,11 @@ For Jetson user, if you have Jetson Orin, you can try this: [Offical Support](ht
   cmake --build build --config Release
   ```
 
-The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used. The following compilation options are also available to tweak performance:
+The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used.
+
+The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted. In Windows this setting is available in the NVIDIA control panel as `System Memory Fallback`.
+
+The following compilation options are also available to tweak performance:
 
 | Option                        | Legal values           | Default | Description                                                                                                                                                                                                                                                                             |
 |-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
@@ -192,6 +196,19 @@ The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/c
 | GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                         |
 | GGML_CUDA_FA_ALL_QUANTS       | Boolean                | false   | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer.                                                                                                  |
 
+### MUSA
+
+- Using `make`:
+  ```bash
+  make GGML_MUSA=1
+  ```
+- Using `CMake`:
+
+  ```bash
+  cmake -B build -DGGML_MUSA=ON
+  cmake --build build --config Release
+  ```
+
 ### hipBLAS
 
 This provides BLAS acceleration on HIP-supported AMD GPUs.
diff --git a/examples/baby-llama/baby-llama.cpp b/examples/baby-llama/baby-llama.cpp
index 4f6c3746..aca332e9 100644
--- a/examples/baby-llama/baby-llama.cpp
+++ b/examples/baby-llama/baby-llama.cpp
@@ -1,7 +1,6 @@
 #include "ggml.h"
 #include "train.h"
 
-#include <vector>
 #include <cassert>
 #include <cstdlib>
 #include <cstring>
diff --git a/examples/batched-bench/batched-bench.cpp b/examples/batched-bench/batched-bench.cpp
index 718f0a61..25e7c775 100644
--- a/examples/batched-bench/batched-bench.cpp
+++ b/examples/batched-bench/batched-bench.cpp
@@ -69,7 +69,7 @@ int main(int argc, char ** argv) {
     llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
 
     // ensure enough sequences are available
-    ctx_params.n_seq_max = *std::max_element(n_pl.begin(), n_pl.end());
+    ctx_params.n_seq_max = n_pl.empty() ? 1 : *std::max_element(n_pl.begin(), n_pl.end());
 
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 
diff --git a/examples/cvector-generator/cvector-generator.cpp b/examples/cvector-generator/cvector-generator.cpp
index d4e126ac..a12e90d8 100644
--- a/examples/cvector-generator/cvector-generator.cpp
+++ b/examples/cvector-generator/cvector-generator.cpp
@@ -414,9 +414,10 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     // load the model to get hparams
-    llama_model * model;
-    llama_context * ctx;
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // int n_ctx = llama_n_ctx(ctx);
     int n_layers = llama_n_layer(model);
diff --git a/examples/embedding/README.md b/examples/embedding/README.md
index e3705b45..12b372bf 100644
--- a/examples/embedding/README.md
+++ b/examples/embedding/README.md
@@ -9,13 +9,13 @@ To get started right away, run the following command, making sure to use the cor
 ### Unix-based systems (Linux, macOS, etc.):
 
 ```bash
-./llama-embedding -m ./path/to/model --log-disable -p "Hello World!" 2>/dev/null
+./llama-embedding -m ./path/to/model --pooling mean --log-disable -p "Hello World!" 2>/dev/null
 ```
 
 ### Windows:
 
 ```powershell
-llama-embedding.exe -m ./path/to/model --log-disable -p "Hello World!" 2>$null
+llama-embedding.exe -m ./path/to/model --pooling mean --log-disable -p "Hello World!" 2>$null
 ```
 
 The above command will output space-separated float values.
@@ -50,11 +50,11 @@ The above command will output space-separated float values.
 ### Unix-based systems (Linux, macOS, etc.):
 
 ```bash
-./embedding -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --embd-separator '<#sep#>' --embd-normalize 2  --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
+./llama-embedding -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --pooling mean --embd-separator '<#sep#>' --embd-normalize 2  --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
 ```
 
 ### Windows:
 
 ```powershell
-embedding.exe -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --embd-separator '<#sep#>' --embd-normalize 2  --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
+llama-embedding.exe -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --pooling mean --embd-separator '<#sep#>' --embd-normalize 2  --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
 ```
diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp
index 1466e5b2..b05aa006 100644
--- a/examples/embedding/embedding.cpp
+++ b/examples/embedding/embedding.cpp
@@ -31,13 +31,24 @@ static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & toke
 }
 
 static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd, int embd_norm) {
+    const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
+    const struct llama_model * model = llama_get_model(ctx);
+
     // clear previous kv_cache values (irrelevant for embeddings)
     llama_kv_cache_clear(ctx);
 
     // run model
     fprintf(stderr, "%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
-    if (llama_decode(ctx, batch) < 0) {
-        fprintf(stderr, "%s : failed to decode\n", __func__);
+    if (llama_model_has_encoder(model) && !llama_model_has_decoder(model)) {
+        // encoder-only model
+        if (llama_encode(ctx, batch) < 0) {
+            fprintf(stderr, "%s : failed to encode\n", __func__);
+        }
+    } else if (!llama_model_has_encoder(model) && llama_model_has_decoder(model)) {
+        // decoder-only model
+        if (llama_decode(ctx, batch) < 0) {
+            fprintf(stderr, "%s : failed to decode\n", __func__);
+        }
     }
 
     for (int i = 0; i < batch.n_tokens; i++) {
@@ -45,11 +56,22 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
             continue;
         }
 
-        // try to get sequence embeddings - supported only when pooling_type is not NONE
-        const float * embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
-        GGML_ASSERT(embd != NULL && "failed to get sequence embeddings");
+        const float * embd = nullptr;
+        int embd_pos = 0;
+
+        if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
+            // try to get token embeddings
+            embd = llama_get_embeddings_ith(ctx, i);
+            embd_pos = i;
+            GGML_ASSERT(embd != NULL && "failed to get token embeddings");
+        } else {
+            // try to get sequence embeddings - supported only when pooling_type is not NONE
+            embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
+            embd_pos = batch.seq_id[i][0];
+            GGML_ASSERT(embd != NULL && "failed to get sequence embeddings");
+        }
 
-        float * out = output + batch.seq_id[i][0] * n_embd;
+        float * out = output + embd_pos * n_embd;
         llama_embd_normalize(embd, out, n_embd, embd_norm);
     }
 }
@@ -79,11 +101,11 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model;
-    llama_context * ctx;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == NULL) {
         fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;
@@ -93,8 +115,9 @@ int main(int argc, char ** argv) {
     const int n_ctx = llama_n_ctx(ctx);
 
     const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
-    if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
-        fprintf(stderr, "%s: error: pooling type NONE not supported\n", __func__);
+
+    if (llama_model_has_encoder(model) && llama_model_has_decoder(model)) {
+        fprintf(stderr, "%s: error: computing embeddings in encoder-decoder models is not supported\n", __func__);
         return 1;
     }
 
@@ -153,13 +176,23 @@ int main(int argc, char ** argv) {
     const int n_prompts = prompts.size();
     struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
 
+    // count number of embeddings
+    int n_embd_count = 0;
+    if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
+        for (int k = 0; k < n_prompts; k++) {
+            n_embd_count += inputs[k].size();
+        }
+    } else {
+        n_embd_count = n_prompts;
+    }
+
     // allocate output
     const int n_embd = llama_n_embd(model);
-    std::vector<float> embeddings(n_prompts * n_embd, 0);
+    std::vector<float> embeddings(n_embd_count * n_embd, 0);
     float * emb = embeddings.data();
 
     // break into batches
-    int p = 0; // number of prompts processed already
+    int e = 0; // number of embeddings already stored
     int s = 0; // number of prompts in current batch
     for (int k = 0; k < n_prompts; k++) {
         // clamp to n_batch tokens
@@ -169,11 +202,11 @@ int main(int argc, char ** argv) {
 
         // encode if at capacity
         if (batch.n_tokens + n_toks > n_batch) {
-            float * out = emb + p * n_embd;
+            float * out = emb + e * n_embd;
             batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
-            llama_batch_clear(batch);
-            p += s;
+            e += pooling_type == LLAMA_POOLING_TYPE_NONE ? batch.n_tokens : s;
             s = 0;
+            llama_batch_clear(batch);
         }
 
         // add to batch
@@ -182,39 +215,62 @@ int main(int argc, char ** argv) {
     }
 
     // final batch
-    float * out = emb + p * n_embd;
+    float * out = emb + e * n_embd;
     batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
 
     if (params.embd_out.empty()) {
-        // print the first part of the embeddings or for a single prompt, the full embedding
         fprintf(stdout, "\n");
-        for (int j = 0; j < n_prompts; j++) {
-            fprintf(stdout, "embedding %d: ", j);
-            for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
-                if (params.embd_normalize == 0) {
-                    fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
-                } else {
-                    fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
+
+        if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
+            for (int j = 0; j < n_embd_count; j++) {
+                fprintf(stdout, "embedding %d: ", j);
+                for (int i = 0; i < std::min(3, n_embd); i++) {
+                    if (params.embd_normalize == 0) {
+                        fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
+                    } else {
+                        fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
+                    }
+                }
+                fprintf(stdout, " ... ");
+                for (int i = n_embd - 3; i < n_embd; i++) {
+                    if (params.embd_normalize == 0) {
+                        fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
+                    } else {
+                        fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
+                    }
                 }
+                fprintf(stdout, "\n");
             }
-            fprintf(stdout, "\n");
-        }
-
-        // print cosine similarity matrix
-        if (n_prompts > 1) {
-            fprintf(stdout, "\n");
-            printf("cosine similarity matrix:\n\n");
-            for (int i = 0; i < n_prompts; i++) {
-                fprintf(stdout, "%6.6s ", prompts[i].c_str());
+        } else {
+            // print the first part of the embeddings or for a single prompt, the full embedding
+            for (int j = 0; j < n_prompts; j++) {
+                fprintf(stdout, "embedding %d: ", j);
+                for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
+                    if (params.embd_normalize == 0) {
+                        fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
+                    } else {
+                        fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
+                    }
+                }
+                fprintf(stdout, "\n");
             }
-            fprintf(stdout, "\n");
-            for (int i = 0; i < n_prompts; i++) {
-                for (int j = 0; j < n_prompts; j++) {
-                    float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
-                    fprintf(stdout, "%6.2f ", sim);
+
+            // print cosine similarity matrix
+            if (n_prompts > 1) {
+                fprintf(stdout, "\n");
+                printf("cosine similarity matrix:\n\n");
+                for (int i = 0; i < n_prompts; i++) {
+                    fprintf(stdout, "%6.6s ", prompts[i].c_str());
                 }
-                fprintf(stdout, "%1.10s", prompts[i].c_str());
                 fprintf(stdout, "\n");
+                for (int i = 0; i < n_prompts; i++) {
+                    for (int j = 0; j < n_prompts; j++) {
+                        float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
+                        fprintf(stdout, "%6.2f ", sim);
+                    }
+                    fprintf(stdout, "%1.10s", prompts[i].c_str());
+                    fprintf(stdout, "\n");
+                }
             }
         }
     }
@@ -233,23 +289,23 @@ int main(int argc, char ** argv) {
             }
             fprintf(stdout, notArray ? "]\n    }" : "]");
             j++;
-            if (j < n_prompts) fprintf(stdout, notArray ? ",\n" : ","); else break;
+            if (j < n_embd_count) fprintf(stdout, notArray ? ",\n" : ","); else break;
         }
         fprintf(stdout, notArray ? "\n  ]" : "]\n");
 
         if (params.embd_out == "json+" && n_prompts > 1) {
             fprintf(stdout, ",\n  \"cosineSimilarity\": [\n");
-            for (int i = 0;;) { // at least two iteration (n_prompts > 1)
+            for (int i = 0;;) { // at least two iteration (n_embd_count > 1)
                 fprintf(stdout, "    [");
-                for (int j = 0;;) { // at least two iteration (n_prompts > 1)
+                for (int j = 0;;) { // at least two iteration (n_embd_count > 1)
                     float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
                     fprintf(stdout, "%6.2f", sim);
                     j++;
-                    if (j < n_prompts) fprintf(stdout, ", "); else break;
+                    if (j < n_embd_count) fprintf(stdout, ", "); else break;
                 }
                 fprintf(stdout, " ]");
                 i++;
-                if (i < n_prompts) fprintf(stdout, ",\n"); else break;
+                if (i < n_embd_count) fprintf(stdout, ",\n"); else break;
             }
             fprintf(stdout, "\n  ]");
         }
diff --git a/examples/eval-callback/eval-callback.cpp b/examples/eval-callback/eval-callback.cpp
index c8a3016a..ef35ba2c 100644
--- a/examples/eval-callback/eval-callback.cpp
+++ b/examples/eval-callback/eval-callback.cpp
@@ -62,7 +62,7 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
                     } else if (type == GGML_TYPE_I8) {
                         v = (float) *(int8_t *) &data[i];
                     } else {
-                        GGML_ASSERT(false);
+                        GGML_ABORT("fatal error");
                     }
                     printf("%12.4f", v);
                     sum += v;
@@ -163,9 +163,10 @@ int main(int argc, char ** argv) {
     params.warmup = false;
 
     // init
-    llama_model * model;
-    llama_context * ctx;
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == nullptr || ctx == nullptr) {
         fprintf(stderr, "%s : failed to init\n", __func__);
         return 1;
diff --git a/examples/export-lora/export-lora.cpp b/examples/export-lora/export-lora.cpp
index 150f7e8d..3176d6e2 100644
--- a/examples/export-lora/export-lora.cpp
+++ b/examples/export-lora/export-lora.cpp
@@ -50,20 +50,6 @@ static struct gguf_context * load_gguf(std::string & fname, struct ggml_context
     return ctx_gguf;
 }
 
-static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
-    std::string result;
-    for (size_t pos = 0; ; pos += search.length()) {
-        auto new_pos = s.find(search, pos);
-        if (new_pos == std::string::npos) {
-            result += s.substr(pos, s.size() - pos);
-            break;
-        }
-        result += s.substr(pos, new_pos - pos) + replace;
-        pos = new_pos;
-    }
-    s = std::move(result);
-}
-
 struct file_input {
     struct ggml_context * ctx_meta = nullptr;
     struct gguf_context * ctx_gguf = nullptr;
@@ -135,7 +121,7 @@ struct lora_merge_ctx {
 
     lora_merge_ctx(
             std::string & base_fname,
-            std::vector<std::tuple<std::string, float>> & lora_files,
+            std::vector<llama_lora_adapter_info> & lora_files,
             std::string & outfile,
             int n_threads) : base_model(base_fname, 0), n_threads(n_threads), fout(outfile, std::ios::binary) {
         fout.exceptions(std::ofstream::failbit); // fail fast on write errors
@@ -144,9 +130,9 @@ struct lora_merge_ctx {
             throw std::runtime_error("split model is not yet supported");
         }
 
-        for (auto lora_inp : lora_files) {
-            auto fname = std::get<0>(lora_inp);
-            auto scale = std::get<1>(lora_inp);
+        for (auto & lora_inp : lora_files) {
+            auto fname = lora_inp.path;
+            auto scale = lora_inp.scale;
             std::unique_ptr<file_input> adapter(new file_input(fname, scale));
             check_metadata_lora(adapter.get());
             adapters.push_back(std::move(adapter));
@@ -407,7 +393,7 @@ int main(int argc, char ** argv) {
 
     g_verbose = (params.verbosity == 1);
     try {
-        lora_merge_ctx ctx(params.model, params.lora_adapter, params.lora_outfile, params.n_threads);
+        lora_merge_ctx ctx(params.model, params.lora_adapters, params.lora_outfile, params.n_threads);
         ctx.run_merge();
     } catch (const std::exception & err) {
         fprintf(stderr, "%s\n", err.what());
diff --git a/examples/imatrix/imatrix.cpp b/examples/imatrix/imatrix.cpp
index 8afc6aa7..497c9d14 100644
--- a/examples/imatrix/imatrix.cpp
+++ b/examples/imatrix/imatrix.cpp
@@ -135,7 +135,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
         }
         else if (e.values.size() != (size_t)src1->ne[0]*n_as) {
             fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]*n_as);
-            exit(1); //GGML_ASSERT(false);
+            exit(1); //GGML_ABORT("fatal error");
         }
         if (m_params.verbosity > 1) {
             printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[2], (int)src1->type);
@@ -184,7 +184,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
         }
         else if (e.values.size() != (size_t)src1->ne[0]) {
             fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
-            exit(1); //GGML_ASSERT(false);
+            exit(1); //GGML_ABORT("fatal error");
         }
         ++e.ncall;
         if (m_params.verbosity > 1) {
@@ -619,10 +619,10 @@ int main(int argc, char ** argv) {
     params.warmup = false;
 
     // init
-    llama_model * model;
-    llama_context * ctx;
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
 
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == nullptr || ctx == nullptr) {
         fprintf(stderr, "%s : failed to init\n", __func__);
         return 1;
diff --git a/examples/infill/infill.cpp b/examples/infill/infill.cpp
index dc93d230..92d630b1 100644
--- a/examples/infill/infill.cpp
+++ b/examples/infill/infill.cpp
@@ -179,7 +179,10 @@ int main(int argc, char ** argv) {
 
     // load the model and apply lora adapter, if any
     LOG("%s: load the model and apply lora adapter, if any\n", __func__);
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    model = llama_init.model;
+    ctx = llama_init.context;
 
     if (model == NULL) {
         LOG_TEE("%s: error: unable to load model\n", __func__);
diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp
index a6497b6e..42918bfc 100644
--- a/examples/llama-bench/llama-bench.cpp
+++ b/examples/llama-bench/llama-bench.cpp
@@ -27,6 +27,14 @@
 #include "ggml-cann.h"
 #endif
 
+#ifdef _WIN32
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
+#include <windows.h>
+#endif
+
 // utils
 static uint64_t get_time_ns() {
     using clock = std::chrono::high_resolution_clock;
@@ -96,6 +104,27 @@ static std::string get_cpu_info() {
         }
         fclose(f);
     }
+#elif defined(_WIN32)
+    HKEY hKey;
+    if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
+                     TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
+                     0,
+                     KEY_READ,
+                     &hKey) != ERROR_SUCCESS) {
+        // fail to open registry key
+        return "";
+    }
+    char cpu_brand[256];
+    DWORD cpu_brand_size = sizeof(cpu_brand);
+    if (RegQueryValueExA(hKey,
+                        TEXT("ProcessorNameString"),
+                        NULL,
+                        NULL,
+                        (LPBYTE)cpu_brand,
+                        &cpu_brand_size) == ERROR_SUCCESS) {
+        id.assign(cpu_brand, cpu_brand_size);
+    }
+    RegCloseKey(hKey);
 #endif
     // TODO: other platforms
     return id;
@@ -150,7 +179,7 @@ static const char * output_format_str(output_formats format) {
         case JSON:     return "json";
         case MARKDOWN: return "md";
         case SQL:      return "sql";
-        default: GGML_ASSERT(!"invalid output format");
+        default: GGML_ABORT("invalid output format");
     }
 }
 
@@ -176,7 +205,7 @@ static const char * split_mode_str(llama_split_mode mode) {
         case LLAMA_SPLIT_MODE_NONE:  return "none";
         case LLAMA_SPLIT_MODE_LAYER: return "layer";
         case LLAMA_SPLIT_MODE_ROW:   return "row";
-        default: GGML_ASSERT(!"invalid split mode");
+        default: GGML_ABORT("invalid split mode");
     }
 }
 
@@ -1326,7 +1355,7 @@ static std::unique_ptr<printer> create_printer(output_formats format) {
         case SQL:
             return std::unique_ptr<printer>(new sql_printer());
     }
-    GGML_ASSERT(false);
+    GGML_ABORT("fatal error");
 }
 
 int main(int argc, char ** argv) {
diff --git a/examples/llava/CMakeLists.txt b/examples/llava/CMakeLists.txt
index e9fa73ac..bbf5fec5 100644
--- a/examples/llava/CMakeLists.txt
+++ b/examples/llava/CMakeLists.txt
@@ -36,3 +36,10 @@ set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-llava-cli)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
+
+set(TARGET llama-minicpmv-cli)
+add_executable(${TARGET} minicpmv-cli.cpp)
+set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-minicpmv-cli)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
diff --git a/examples/llava/README-minicpmv2.5.md b/examples/llava/README-minicpmv2.5.md
new file mode 100644
index 00000000..4affc1d0
--- /dev/null
+++ b/examples/llava/README-minicpmv2.5.md
@@ -0,0 +1,99 @@
+## MiniCPM-Llama3-V 2.5
+
+### Prepare models and code
+
+Download [MiniCPM-Llama3-V-2_5](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5) PyTorch model from huggingface to "MiniCPM-Llama3-V-2_5" folder.
+
+Clone llama.cpp:
+```bash
+git clone https://github.com/ggerganov/llama.cpp
+cd llama.cpp
+```
+
+### Usage
+
+Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5-gguf) by us)
+
+```bash
+python ./examples/minicpmv/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5
+python ./examples/minicpmv/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5
+python ./convert-hf-to-gguf.py ../MiniCPM-Llama3-V-2_5/model
+
+# quantize int4 version
+./llama-quantize ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf Q4_K_M
+```
+
+Build for Linux or Mac
+
+```bash
+make
+make llama-minicpmv-cli
+```
+
+Inference on Linux or Mac
+```
+# run f16 version
+./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+
+# run quantized int4 version
+./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
+
+# or run in interactive mode
+./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -i
+```
+
+### Android
+
+#### Build on Android device using Termux
+We found that build on Android device would bring better runtime performance, so we recommend to build on device.
+
+[Termux](https://github.com/termux/termux-app#installation) is a terminal app on Android device (no root required).
+
+Install tools in Termux:
+```
+apt update && apt upgrade -y
+apt install git make cmake
+```
+
+It's recommended to move your model inside the `~/` directory for best performance:
+```
+cd storage/downloads
+mv model.gguf ~/
+```
+
+#### Building the Project using Android NDK
+Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake.
+
+Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux:
+
+```bash
+mkdir build-android
+cd build-android
+export NDK=/your_ndk_path
+cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod ..
+make
+```
+
+Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice).
+
+Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
+
+(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
+```
+$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/
+$cd /data/data/com.termux/files/home/bin
+$chmod +x ./*
+```
+
+Download models and push them to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
+
+```
+$mv /sdcard/llama.cpp/ggml-model-Q4_K_M.gguf /data/data/com.termux/files/home/model/
+$mv /sdcard/llama.cpp/mmproj-model-f16.gguf /data/data/com.termux/files/home/model/
+```
+
+Now, you can start chatting:
+```
+$cd /data/data/com.termux/files/home/bin
+$./llama-minicpmv-cli -m ../model/ggml-model-Q4_K_M.gguf --mmproj ../model/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
+```
diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
index d23e282f..54aa822c 100644
--- a/examples/llava/clip.cpp
+++ b/examples/llava/clip.cpp
@@ -74,26 +74,27 @@ static std::string format(const char * fmt, ...) {
 // key constants
 //
 
-#define KEY_FTYPE          "general.file_type"
-#define KEY_NAME           "general.name"
-#define KEY_DESCRIPTION    "general.description"
-#define KEY_HAS_TEXT_ENC   "clip.has_text_encoder"
-#define KEY_HAS_VIS_ENC    "clip.has_vision_encoder"
-#define KEY_HAS_LLAVA_PROJ "clip.has_llava_projector"
-#define KEY_USE_GELU       "clip.use_gelu"
-#define KEY_N_EMBD         "clip.%s.embedding_length"
-#define KEY_N_FF           "clip.%s.feed_forward_length"
-#define KEY_N_BLOCK        "clip.%s.block_count"
-#define KEY_N_HEAD         "clip.%s.attention.head_count"
-#define KEY_LAYER_NORM_EPS "clip.%s.attention.layer_norm_epsilon"
-#define KEY_PROJ_DIM       "clip.%s.projection_dim"
-#define KEY_TOKENS         "tokenizer.ggml.tokens"
-#define KEY_N_POSITIONS    "clip.text.context_length"
-#define KEY_IMAGE_SIZE     "clip.vision.image_size"
-#define KEY_PATCH_SIZE     "clip.vision.patch_size"
-#define KEY_IMAGE_MEAN     "clip.vision.image_mean"
-#define KEY_IMAGE_STD      "clip.vision.image_std"
-#define KEY_PROJ_TYPE      "clip.projector_type"
+#define KEY_FTYPE               "general.file_type"
+#define KEY_NAME                "general.name"
+#define KEY_DESCRIPTION         "general.description"
+#define KEY_HAS_TEXT_ENC        "clip.has_text_encoder"
+#define KEY_HAS_VIS_ENC         "clip.has_vision_encoder"
+#define KEY_HAS_LLAVA_PROJ      "clip.has_llava_projector"
+#define KEY_HAS_MINICPMV_PROJ   "clip.has_minicpmv_projector"
+#define KEY_USE_GELU            "clip.use_gelu"
+#define KEY_N_EMBD              "clip.%s.embedding_length"
+#define KEY_N_FF                "clip.%s.feed_forward_length"
+#define KEY_N_BLOCK             "clip.%s.block_count"
+#define KEY_N_HEAD              "clip.%s.attention.head_count"
+#define KEY_LAYER_NORM_EPS      "clip.%s.attention.layer_norm_epsilon"
+#define KEY_PROJ_DIM            "clip.%s.projection_dim"
+#define KEY_TOKENS              "tokenizer.ggml.tokens"
+#define KEY_N_POSITIONS         "clip.text.context_length"
+#define KEY_IMAGE_SIZE          "clip.vision.image_size"
+#define KEY_PATCH_SIZE          "clip.vision.patch_size"
+#define KEY_IMAGE_MEAN          "clip.vision.image_mean"
+#define KEY_IMAGE_STD           "clip.vision.image_std"
+#define KEY_PROJ_TYPE           "clip.projector_type"
 
 #define KEY_MM_PATCH_MERGE_TYPE   "clip.vision.mm_patch_merge_type"
 #define KEY_IMAGE_GRID_PINPOINTS  "clip.vision.image_grid_pinpoints"
@@ -127,12 +128,20 @@ static std::string format(const char * fmt, ...) {
 #define TN_MVLM_PROJ_PEG   "mm.model.peg.%d.%s"
 #define TN_IMAGE_NEWLINE   "model.image_newline"
 
+#define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k"
+#define TN_MINICPMV_QUERY "resampler.query"
+#define TN_MINICPMV_PROJ "resampler.proj.weight"
+#define TN_MINICPMV_KV_PROJ "resampler.kv.weight"
+#define TN_MINICPMV_ATTN "resampler.attn.%s.%s"
+#define TN_MINICPMV_LN "resampler.ln_%s.%s"
+
 
 enum projector_type {
     PROJECTOR_TYPE_MLP,
     PROJECTOR_TYPE_MLP_NORM,
     PROJECTOR_TYPE_LDP,
     PROJECTOR_TYPE_LDPV2,
+    PROJECTOR_TYPE_RESAMPLER,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -140,6 +149,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_MLP, "mlp" },
     { PROJECTOR_TYPE_LDP, "ldp" },
     { PROJECTOR_TYPE_LDPV2, "ldpv2"},
+    { PROJECTOR_TYPE_RESAMPLER, "resampler"},
 };
 
 
@@ -200,17 +210,14 @@ static std::string gguf_data_to_str(enum gguf_type type, const void * data, int
 }
 
 static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
-    std::string result;
-    for (size_t pos = 0; ; pos += search.length()) {
-        auto new_pos = s.find(search, pos);
-        if (new_pos == std::string::npos) {
-            result += s.substr(pos, s.size() - pos);
-            break;
-        }
-        result += s.substr(pos, new_pos - pos) + replace;
-        pos = new_pos;
+    if (search.empty()) {
+        return; // Avoid infinite loop if 'search' is an empty string
+    }
+    size_t pos = 0;
+    while ((pos = s.find(search, pos)) != std::string::npos) {
+        s.replace(pos, search.length(), replace);
+        pos += replace.length();
     }
-    s = std::move(result);
 }
 
 static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
@@ -492,12 +499,33 @@ struct clip_vision_model {
     struct ggml_tensor * mm_model_mlp_2_b;
     struct ggml_tensor * mm_model_peg_0_w;
     struct ggml_tensor * mm_model_peg_0_b;
+
+    // MINICPMV projection
+    struct ggml_tensor * mm_model_pos_embed_k;
+    struct ggml_tensor * mm_model_query;
+    struct ggml_tensor * mm_model_proj;
+    struct ggml_tensor * mm_model_kv_proj;
+    struct ggml_tensor * mm_model_attn_q_w;
+    struct ggml_tensor * mm_model_attn_q_b;
+    struct ggml_tensor * mm_model_attn_k_w;
+    struct ggml_tensor * mm_model_attn_k_b;
+    struct ggml_tensor * mm_model_attn_v_w;
+    struct ggml_tensor * mm_model_attn_v_b;
+    struct ggml_tensor * mm_model_attn_o_w;
+    struct ggml_tensor * mm_model_attn_o_b;
+    struct ggml_tensor * mm_model_ln_q_w;
+    struct ggml_tensor * mm_model_ln_q_b;
+    struct ggml_tensor * mm_model_ln_kv_w;
+    struct ggml_tensor * mm_model_ln_kv_b;
+    struct ggml_tensor * mm_model_ln_post_w;
+    struct ggml_tensor * mm_model_ln_post_b;
 };
 
 struct clip_ctx {
     bool has_text_encoder    = false;
     bool has_vision_encoder  = false;
     bool has_llava_projector = false;
+    bool has_minicpmv_projector = false;
 
     struct clip_vision_model vision_model;
     projector_type proj_type = PROJECTOR_TYPE_MLP;
@@ -522,9 +550,11 @@ struct clip_ctx {
 
     ggml_backend_t backend       = NULL;
     ggml_gallocr_t compute_alloc = NULL;
+
+    struct clip_image_size * load_image_size;
 };
 
-static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs) {
+static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs, struct clip_image_size * load_image_size, bool is_inf = false) {
     if (!ctx->has_vision_encoder) {
         LOG_TEE("This gguf file seems to have no vision encoder\n");
         return nullptr;
@@ -533,20 +563,33 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     const auto & model = ctx->vision_model;
     const auto & hparams = model.hparams;
 
-    const int image_size           = hparams.image_size;
+    const int image_size = hparams.image_size;
+    int image_size_width  = image_size;
+    int image_size_height = image_size;
+    if (ctx->has_minicpmv_projector) {
+        if (load_image_size == nullptr) {
+            load_image_size = clip_image_size_init();
+        }
+        LOG_TEE("%s: %d %d\n", __func__, load_image_size->width, load_image_size->height);
+        image_size_width  = load_image_size->width;
+        image_size_height = load_image_size->height;
+        if (is_inf) {
+            image_size_width  = imgs->data->nx;
+            image_size_height = imgs->data->ny;
+        }
+    }
     const int patch_size           = hparams.patch_size;
-    const int num_patches          = ((image_size / patch_size) * (image_size / patch_size));
-    const int num_patches_per_side = image_size / patch_size; GGML_UNUSED(num_patches_per_side);
+    const int num_patches          = ((image_size_width / patch_size) * (image_size_height / patch_size));
     const int num_positions        = num_patches + (ctx->has_class_embedding ? 1 : 0);
     const int hidden_size          = hparams.hidden_size;
     const int n_head               = hparams.n_head;
     const int d_head               = hidden_size / n_head;
-    const int n_layer              = hparams.n_layer;
+    int n_layer                    = hparams.n_layer;
     const float eps                = hparams.eps;
 
     const int batch_size = imgs->size;
 
-    if (ctx->has_llava_projector) {
+    if (ctx->has_llava_projector || ctx->has_minicpmv_projector) {
         GGML_ASSERT(batch_size == 1);
     }
 
@@ -559,7 +602,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     struct ggml_context * ctx0 = ggml_init(params);
     struct ggml_cgraph * gf = ggml_new_graph(ctx0);
 
-    struct ggml_tensor * inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size, image_size, 3, batch_size);
+    struct ggml_tensor * inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size_width, image_size_height, 3, batch_size);
     ggml_set_name(inp_raw, "inp_raw");
     ggml_set_input(inp_raw);
 
@@ -572,19 +615,21 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         // inp = ggml_add(ctx0, inp, ggml_repeat(ctx0, model.patch_bias, inp));
         inp = ggml_add(ctx0, inp, model.patch_bias);
     }
-
-    // concat class_embeddings and patch_embeddings
     struct ggml_tensor * embeddings = inp;
-    if (ctx->has_class_embedding) {
-        embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size);
-        ggml_set_name(embeddings, "embeddings");
-        ggml_set_input(embeddings);
-        embeddings = ggml_acc(ctx0, embeddings, model.class_embedding,
-                embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0);
-        embeddings = ggml_acc(ctx0, embeddings, inp,
-                embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
-    }
+    struct ggml_tensor * pos_embed = nullptr;
 
+    if (ctx->has_llava_projector) {
+        // concat class_embeddings and patch_embeddings
+        if (ctx->has_class_embedding) {
+            embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size);
+            ggml_set_name(embeddings, "embeddings");
+            ggml_set_input(embeddings);
+            embeddings = ggml_acc(ctx0, embeddings, model.class_embedding,
+                    embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0);
+            embeddings = ggml_acc(ctx0, embeddings, inp,
+                    embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
+        }
+    }
 
     struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions);
     ggml_set_name(positions, "positions");
@@ -593,6 +638,14 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     embeddings =
         ggml_add(ctx0, embeddings, ggml_get_rows(ctx0, model.position_embeddings, positions));
 
+    if (ctx->has_minicpmv_projector) {
+        int pos_w = image_size_width/patch_size;
+        int pos_h = image_size_height/patch_size;
+        pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 4096, pos_w * pos_h, 1);
+        ggml_set_name(pos_embed, "pos_embed");
+        ggml_set_input(pos_embed);
+    }
+
     // pre-layernorm
     if (ctx->has_pre_norm) {
         embeddings = ggml_norm(ctx0, embeddings, eps);
@@ -602,6 +655,9 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     }
 
     // loop over layers
+    if (ctx->has_minicpmv_projector) {
+        n_layer += 1;
+    }
     for (int il = 0; il < n_layer - 1; il++) {
         struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states
 
@@ -691,7 +747,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     }
 
     // llava projector
-    {
+    if (ctx->has_llava_projector) {
         embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]);
 
         struct ggml_tensor * patches = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_patches);
@@ -869,6 +925,65 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             embeddings = peg_0;
         }
         else {
+            GGML_ABORT("fatal error");
+        }
+    }
+    // minicpmv projector
+    else if (ctx->has_minicpmv_projector)
+    {
+        if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
+            struct ggml_tensor * q = model.mm_model_query;
+            { // layernorm
+                q = ggml_norm(ctx0, q, eps);
+                q = ggml_add(ctx0, ggml_mul(ctx0, q, model.mm_model_ln_q_w), model.mm_model_ln_q_b);
+            }
+            struct ggml_tensor * v = ggml_mul_mat(ctx0, model.mm_model_kv_proj, embeddings);
+            { // layernorm
+                v = ggml_norm(ctx0, v, eps);
+                v = ggml_add(ctx0, ggml_mul(ctx0, v, model.mm_model_ln_kv_w), model.mm_model_ln_kv_b);
+            }
+            struct ggml_tensor * k;
+            { // position
+                // q = ggml_add(ctx0, q, model.mm_model_pos_embed);
+                k = ggml_add(ctx0, v, pos_embed);
+            }
+
+            { // attention
+                const int hidden_size = 4096;
+                const int d_head = 128;
+                const int n_head = hidden_size/d_head;
+                const int num_query = 96;
+
+                struct ggml_tensor * Q = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_q_w, q), model.mm_model_attn_q_b);
+                Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head));
+                struct ggml_tensor * K = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_k_w, k), model.mm_model_attn_k_b);
+                struct ggml_tensor * V = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_v_w, v), model.mm_model_attn_v_b);
+                // permute
+                Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, num_query, batch_size);
+                Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
+                Q = ggml_reshape_3d(ctx0, Q, d_head, num_query, n_head * batch_size);
+                K = ggml_reshape_4d(ctx0, K, d_head, n_head, num_positions, batch_size);
+                K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3));
+                K = ggml_reshape_3d(ctx0, K, d_head, num_positions, n_head * batch_size);
+                V = ggml_reshape_4d(ctx0, V, d_head, n_head, num_positions, batch_size);
+                V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3));
+                V = ggml_reshape_3d(ctx0, V, num_positions, d_head, n_head * batch_size);
+                struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+                KQ = ggml_soft_max_inplace(ctx0, KQ);
+                struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ);
+                KQV = ggml_reshape_4d(ctx0, KQV, d_head, num_query, n_head, batch_size);
+                KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+                KQV = ggml_cont_3d(ctx0, KQV, hidden_size, num_query, batch_size);
+
+                embeddings = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_o_w, KQV), model.mm_model_attn_o_b);
+            }
+            { // layernorm
+                embeddings = ggml_norm(ctx0, embeddings, eps);
+                embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_model_ln_post_w), model.mm_model_ln_post_b);
+            }
+            embeddings = ggml_mul_mat(ctx0, model.mm_model_proj, embeddings);
+        }
+        else {
             GGML_ASSERT(false);
         }
     }
@@ -1029,7 +1144,13 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             new_clip->has_llava_projector = gguf_get_val_bool(ctx, idx);
         }
 
-        GGML_ASSERT(new_clip->has_llava_projector); // see monatis/clip.cpp for image and/or text encoding for semantic search
+        idx = gguf_find_key(ctx, KEY_HAS_MINICPMV_PROJ);
+        if (idx != -1) {
+            new_clip->has_minicpmv_projector = gguf_get_val_bool(ctx, idx);
+        }
+
+        // GGML_ASSERT(new_clip->has_llava_projector); // see monatis/clip.cpp for image and/or text encoding for semantic search
+
         GGML_ASSERT(new_clip->has_vision_encoder);
         GGML_ASSERT(!new_clip->has_text_encoder);
 
@@ -1040,6 +1161,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             LOG_TEE("%s: text_encoder:   %d\n", __func__, new_clip->has_text_encoder);
             LOG_TEE("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
             LOG_TEE("%s: llava_projector:  %d\n", __func__, new_clip->has_llava_projector);
+            LOG_TEE("%s: minicpmv_projector:  %d\n", __func__, new_clip->has_minicpmv_projector);
             LOG_TEE("%s: model size:     %.2f MB\n", __func__, model_size / 1024.0 / 1024.0);
             LOG_TEE("%s: metadata size:  %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
         }
@@ -1281,6 +1403,27 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             vision_model.mm_model_peg_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "weight"));
             vision_model.mm_model_peg_0_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "bias"));
         }
+        else if (new_clip->proj_type == PROJECTOR_TYPE_RESAMPLER) {
+            // vision_model.mm_model_pos_embed = get_tensor(new_clip->ctx_data, TN_MINICPMV_POS_EMBD);
+            vision_model.mm_model_pos_embed_k = get_tensor(new_clip->ctx_data, TN_MINICPMV_POS_EMBD_K);
+            vision_model.mm_model_query = get_tensor(new_clip->ctx_data, TN_MINICPMV_QUERY);
+            vision_model.mm_model_proj = get_tensor(new_clip->ctx_data, TN_MINICPMV_PROJ);
+            vision_model.mm_model_kv_proj = get_tensor(new_clip->ctx_data, TN_MINICPMV_KV_PROJ);
+            vision_model.mm_model_attn_q_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "q", "weight"));
+            vision_model.mm_model_attn_k_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "k", "weight"));
+            vision_model.mm_model_attn_v_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "v", "weight"));
+            vision_model.mm_model_attn_q_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "q", "bias"));
+            vision_model.mm_model_attn_k_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "k", "bias"));
+            vision_model.mm_model_attn_v_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "v", "bias"));
+            vision_model.mm_model_attn_o_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "out", "weight"));
+            vision_model.mm_model_attn_o_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "out", "bias"));
+            vision_model.mm_model_ln_q_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "q", "weight"));
+            vision_model.mm_model_ln_q_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "q", "bias"));
+            vision_model.mm_model_ln_kv_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "kv", "weight"));
+            vision_model.mm_model_ln_kv_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "kv", "bias"));
+            vision_model.mm_model_ln_post_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "weight"));
+            vision_model.mm_model_ln_post_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "bias"));
+        }
         else {
             std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type];
             throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
@@ -1319,7 +1462,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         new_clip->compute_alloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(new_clip->backend));
         clip_image_f32_batch batch;
         batch.size = 1;
-        ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch);
+        ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch, nullptr, false);
         ggml_gallocr_reserve(new_clip->compute_alloc, gf);
         size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_clip->compute_alloc, 0);
         LOG_TEE("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0);
@@ -1328,6 +1471,17 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
     return new_clip;
 }
 
+void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size) {
+    ctx_clip->load_image_size = load_image_size;
+}
+
+struct clip_image_size * clip_image_size_init() {
+    struct clip_image_size * load_image_size = new struct clip_image_size();
+    load_image_size->width = 448;
+    load_image_size->height = 448;
+    return load_image_size;
+}
+
 struct clip_image_u8 * clip_image_u8_init() {
     return new clip_image_u8();
 }
@@ -1598,9 +1752,184 @@ static std::vector<clip_image_u8*> divide_to_patches_u8(const clip_image_u8 & im
     return patches;
 }
 
+static int ensure_divide(int length, int patch_size) {
+    return std::max(static_cast<int>(std::round(static_cast<float>(length) / patch_size) * patch_size), patch_size);
+}
+
+static std::pair<int, int> uhd_find_best_resize(std::pair<int, int> original_size, int scale_resolution, int patch_size, bool allow_upscale = false) {
+    int width = original_size.first;
+    int height = original_size.second;
+    if ((width * height > scale_resolution * scale_resolution) || allow_upscale) {
+        float r = static_cast<float>(width) / height;
+        height = static_cast<int>(scale_resolution / std::sqrt(r));
+        width = static_cast<int>(height * r);
+    }
+    int best_width = ensure_divide(width, patch_size);
+    int best_height = ensure_divide(height, patch_size);
+    return std::make_pair(best_width, best_height);
+}
+
+static std::pair<int, int> uhd_get_refine_size(std::pair<int, int> original_size, std::pair<int, int> grid, int scale_resolution, int patch_size, bool allow_upscale = false) {
+    int width, height;
+    std::tie(width, height) = original_size;
+    int grid_x, grid_y;
+    std::tie(grid_x, grid_y) = grid;
+
+    int refine_width = ensure_divide(width, grid_x);
+    int refine_height = ensure_divide(height, grid_y);
+
+    int grid_width = refine_width / grid_x;
+    int grid_height = refine_height / grid_y;
+
+   // auto best_grid_size = find_best_resize(std::make_tuple(grid_width, grid_height), scale_resolution, patch_size, allow_upscale); (old line)
+    auto best_grid_size = uhd_find_best_resize(std::make_pair(grid_width, grid_height), scale_resolution, patch_size, allow_upscale); // (new line) => fixes conversion for make_tuple to make_pair
+    int best_grid_width, best_grid_height;
+    std::tie(best_grid_width, best_grid_height) = best_grid_size;
+
+  //  std::pair<int, int> refine_size = std::make_tuple(best_grid_width * grid_x, best_grid_height * grid_y); (old line)
+    std::pair<int, int> refine_size = std::make_pair(best_grid_width * grid_x, best_grid_height * grid_y); // (new line)
+    return refine_size;
+}
+
+inline int clip(int x, int lower, int upper) {
+    return std::max(lower, std::min(x, upper));
+}
+
+static std::pair<int, int> uhd_best_grid(const int max_slice_nums, const int multiple, const float log_ratio) {
+    std::vector<int> candidate_split_grids_nums;
+    for (int i : {multiple - 1, multiple, multiple + 1}) {
+        if (i == 1 || i > max_slice_nums) {
+            continue;
+        }
+        candidate_split_grids_nums.push_back(i);
+    }
+
+    std::vector<std::pair<int, int>> candidate_grids;
+    for (int split_grids_nums : candidate_split_grids_nums) {
+        int m = 1;
+        while (m <= split_grids_nums) {
+            if (split_grids_nums % m == 0) {
+                candidate_grids.emplace_back(m, split_grids_nums / m);
+            }
+            ++m;
+        }
+    }
+
+    std::pair<int, int> best_grid{1, 1};
+    float min_error = std::numeric_limits<float>::infinity();
+    for (const auto& grid : candidate_grids) {
+        float error = std::abs(log_ratio - std::log(1.0 * grid.first / grid.second));
+        if (error < min_error) {
+            best_grid = grid;
+            min_error = error;
+        }
+    }
+    return best_grid;
+}
+
+// inspired from LLaVA-UHD:
+//    -> https://arxiv.org/pdf/2403.11703
+//    -> https://github.com/thunlp/LLaVA-UHD
+//    -> https://github.com/thunlp/LLaVA-UHD/blob/302301bc2175f7e717fb8548516188e89f649753/llava_uhd/train/llava-uhd/slice_logic.py#L118
+static std::vector<std::vector<clip_image_u8 *>> uhd_slice_image(const clip_image_u8 * img, const int max_slice_nums=9, const int scale_resolution=448, const int patch_size=14) {
+    const std::pair<int, int> original_size={img->nx,img->ny};
+    const int original_width = img->nx;
+    const int original_height = img->ny;
+    const float log_ratio = log(1.0*original_width/original_height);
+    const float ratio = 1.0 * original_width * original_height/ (scale_resolution * scale_resolution);
+    const int multiple = fmin(ceil(ratio), max_slice_nums);
+
+    std::vector<std::vector<clip_image_u8 *>> images;
+    LOG_TEE("%s: multiple %d\n", __func__, multiple);
+    images.push_back(std::vector<clip_image_u8 *>());
+
+    if (multiple <= 1) {
+        auto best_size = uhd_find_best_resize(original_size, scale_resolution, patch_size, true);
+        clip_image_u8 * source_image = clip_image_u8_init();
+        bicubic_resize(*img, *source_image, best_size.first, best_size.second);
+        // source_image = image.resize(best_size, Image.Resampling.BICUBIC)
+        images[images.size()-1].push_back(source_image);
+    }
+    else if (multiple > 1) {
+        auto best_size = uhd_find_best_resize(original_size, scale_resolution, patch_size);
+        clip_image_u8 * source_image = clip_image_u8_init();
+        bicubic_resize(*img, *source_image, best_size.first, best_size.second);
+        // source_image = image.copy().resize(best_resize, Image.Resampling.BICUBIC)
+        LOG_TEE("%s: image_size: %d %d; source_image size: %d %d\n", __func__, img->nx, img->ny, best_size.first, best_size.second);
+        images[images.size()-1].push_back(source_image);
+
+        std::pair<int, int> best_grid = uhd_best_grid(max_slice_nums, multiple, log_ratio);
+        LOG_TEE("%s: image_size: %d %d; best_grid: %d %d\n", __func__, img->nx, img->ny, best_grid.first, best_grid.second);
+
+        auto refine_size = uhd_get_refine_size(original_size, best_grid, scale_resolution, patch_size, true);
+        clip_image_u8 * refine_image = clip_image_u8_init();
+        bicubic_resize(*img, *refine_image, refine_size.first, refine_size.second);
+
+        LOG_TEE("%s: refine_image_size: %d %d; refine_size: %d %d\n", __func__, refine_image->nx, refine_image->ny, refine_size.first, refine_size.second);
+
+        // split_to_patches
+        int width = refine_image->nx;
+        int height = refine_image->ny;
+        int grid_x = int(width / best_grid.first);
+        int grid_y = int(height / best_grid.second);
+        for (int patches_i = 0, ic = 0; patches_i < height && ic < best_grid.second; patches_i += grid_y, ic += 1){
+            images.push_back(std::vector<clip_image_u8 *>());
+            for(int patches_j = 0, jc = 0; patches_j < width && jc < best_grid.first; patches_j += grid_x, jc += 1){
+                clip_image_u8 * patch = clip_image_u8_init();
+                patch->nx = grid_x;
+                patch->ny = grid_y;
+                patch->buf.resize(3 * patch->nx * patch->ny);
+                for (int y = patches_i; y < patches_i + grid_y; ++y) {
+                    for (int x = patches_j; x < patches_j + grid_x; ++x) {
+                        const int i = 3 * (y * refine_image->nx + x);
+                        const int j = 3 * ((y-patches_i) * patch->nx + (x-patches_j));
+                        patch->buf[j]   = refine_image->buf[i];
+                        patch->buf[j+1] = refine_image->buf[i+1];
+                        patch->buf[j+2] = refine_image->buf[i+2];
+                    }
+                }
+                images[images.size()-1].push_back(patch);
+            }
+        }
+    }
+    return images;
+}
+
+int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip) {
+    const int max_slice_nums=9;
+    const int scale_resolution=448;
+    const int original_width = ctx_clip->load_image_size->width;
+    const int original_height = ctx_clip->load_image_size->height;
+    const float log_ratio = log(1.0*original_width/original_height);
+    const float ratio = 1.0 * original_width * original_height/ (scale_resolution * scale_resolution);
+    const int multiple = fmin(ceil(ratio), max_slice_nums);
+    std::pair<int, int> best_grid = uhd_best_grid(max_slice_nums, multiple, log_ratio);
+    return best_grid.first;
+}
+
 // returns the normalized float tensor for llava-1.5, for spatial_unpad with anyres processing for llava-1.6 it returns the normalized image patch tensors as a vector
 // res_imgs memory is being allocated here, previous allocations will be freed if found
 bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch * res_imgs) {
+    if (clip_is_minicpmv(ctx)) {
+        std::vector<std::vector<clip_image_u8 *>> imgs = uhd_slice_image(img);
+        res_imgs->size = 0;
+        for (size_t i = 0; i < imgs.size(); ++i) {
+            res_imgs->size += imgs[i].size();
+        }
+        res_imgs->data = new clip_image_f32[res_imgs->size];
+        int idx = 0;
+        for (size_t i = 0; i < imgs.size(); ++i) {
+            for (size_t j = 0; j < imgs[i].size(); ++j) {
+                LOG_TEE("%s: %d %d\n", __func__,imgs[i][j]->nx,imgs[i][j]->ny);
+                clip_image_f32 * res = clip_image_f32_init();
+                normalize_image_u8_to_f32(imgs[i][j], res, ctx->image_mean, ctx->image_std);
+                res_imgs->data[idx++] = *res;
+                clip_image_f32_free(res);
+            }
+        }
+        return true;
+    }
+
     bool pad_to_square = true;
     if (!ctx->has_vision_encoder) {
         LOG_TEE("This gguf file seems to have no vision encoder\n");
@@ -1816,11 +2145,99 @@ int clip_n_patches(const struct clip_ctx * ctx) {
 
     if (ctx->proj_type == PROJECTOR_TYPE_LDP || ctx->proj_type == PROJECTOR_TYPE_LDPV2) {
         n_patches /= 4;
+    } else if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
+        n_patches = 96;
     }
 
     return n_patches;
 }
 
+static std::vector<std::vector<std::vector<float>>> get_1d_sincos_pos_embed_from_grid_new(int embed_dim, const std::vector<std::vector<float>> & pos) {
+    assert(embed_dim % 2 == 0);
+    int H = pos.size();
+    int W = pos[0].size();
+
+    std::vector<float> omega(embed_dim / 2);
+    for (int i = 0; i < embed_dim / 2; ++i) {
+        omega[i] = 1.0 / pow(10000.0, static_cast<float>(i) / (embed_dim / 2));
+    }
+
+    std::vector<std::vector<std::vector<float>>> emb(H, std::vector<std::vector<float>>(W, std::vector<float>(embed_dim)));
+    for (int h = 0; h < H; ++h) {
+        for (int w = 0; w < W; ++w) {
+            for (int d = 0; d < embed_dim / 2; ++d) {
+                float out_value = pos[h][w] * omega[d];
+                emb[h][w][d] = sin(out_value);
+                emb[h][w][d + embed_dim / 2] = cos(out_value);
+            }
+        }
+    }
+
+    return emb;
+}
+
+static std::vector<std::vector<std::vector<float>>> get_2d_sincos_pos_embed_from_grid(int embed_dim, const std::vector<std::vector<std::vector<float>>> & grid) {
+    assert(embed_dim % 2 == 0);
+    std::vector<std::vector<std::vector<float>>> emb_h = get_1d_sincos_pos_embed_from_grid_new(embed_dim / 2, grid[0]); // (H, W, D/2)
+    std::vector<std::vector<std::vector<float>>> emb_w = get_1d_sincos_pos_embed_from_grid_new(embed_dim / 2, grid[1]); // (H, W, D/2)
+
+    int H = emb_h.size();
+    int W = emb_h[0].size();
+    std::vector<std::vector<std::vector<float>>> emb(H, std::vector<std::vector<float>>(W, std::vector<float>(embed_dim)));
+
+    for (int h = 0; h < H; ++h) {
+        for (int w = 0; w < W; ++w) {
+            for (int d = 0; d < embed_dim / 2; ++d) {
+                emb[h][w][d] = emb_h[h][w][d];
+                emb[h][w][d + embed_dim / 2] = emb_w[h][w][d];
+            }
+        }
+    }
+    return emb;
+}
+
+static std::vector<std::vector<float>> get_2d_sincos_pos_embed(int embed_dim, const std::pair<int, int> image_size) {
+    int grid_h_size = image_size.first;
+    int grid_w_size = image_size.second;
+
+    std::vector<float> grid_h(grid_h_size);
+    std::vector<float> grid_w(grid_w_size);
+
+    for (int i = 0; i < grid_h_size; ++i) {
+        grid_h[i] = static_cast<float>(i);
+    }
+    for (int i = 0; i < grid_w_size; ++i) {
+        grid_w[i] = static_cast<float>(i);
+    }
+
+    std::vector<std::vector<float>> grid(grid_h_size, std::vector<float>(grid_w_size));
+    for (int h = 0; h < grid_h_size; ++h) {
+        for (int w = 0; w < grid_w_size; ++w) {
+            grid[h][w] = grid_w[w];
+        }
+    }
+    std::vector<std::vector<std::vector<float>>> grid_2d = {grid, grid};
+    for (int h = 0; h < grid_h_size; ++h) {
+        for (int w = 0; w < grid_w_size; ++w) {
+            grid_2d[0][h][w] = grid_h[h];
+            grid_2d[1][h][w] = grid_w[w];
+        }
+    }
+
+    std::vector<std::vector<std::vector<float>>> pos_embed_3d = get_2d_sincos_pos_embed_from_grid(embed_dim, grid_2d);
+
+    int H = image_size.first;
+    int W = image_size.second;
+    std::vector<std::vector<float>> pos_embed_2d(H * W, std::vector<float>(embed_dim));
+    for (int h = 0; h < H; ++h) {
+        for (int w = 0; w < W; ++w) {
+            pos_embed_2d[w * H + h] = pos_embed_3d[h][w];
+        }
+    }
+
+    return pos_embed_2d;
+}
+
 bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) {
     if (!ctx->has_vision_encoder) {
         LOG_TEE("This gguf file seems to have no vision encoder\n");
@@ -1843,18 +2260,27 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
     if (ctx->has_llava_projector) {
         GGML_ASSERT(batch_size == 1); // TODO: support multiple images
     }
+    if (ctx->has_minicpmv_projector) {
+        GGML_ASSERT(batch_size == 1);
+    }
 
     // build the inference graph
-    ggml_cgraph * gf = clip_image_build_graph(ctx, imgs);
+    ggml_cgraph * gf = clip_image_build_graph(ctx, imgs, ctx->load_image_size, true);
     ggml_gallocr_alloc_graph(ctx->compute_alloc, gf);
 
     // set inputs
     const auto & model = ctx->vision_model;
     const auto & hparams = model.hparams;
 
-    const int image_size    = hparams.image_size;
+    const int image_size = hparams.image_size;
+    int image_size_width  = image_size;
+    int image_size_height = image_size;
+    if (ctx->has_minicpmv_projector) {
+        image_size_width  = imgs->data[0].nx;
+        image_size_height = imgs->data[0].ny;
+    }
     const int patch_size    = hparams.patch_size;
-    const int num_patches   = ((image_size / patch_size) * (image_size / patch_size));
+    const int num_patches   = ((image_size_width / patch_size) * (image_size_height / patch_size));
     const int num_positions = num_patches + (ctx->has_class_embedding ? 1 : 0);
 
     {
@@ -1864,7 +2290,9 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         for (size_t i = 0; i < imgs->size; i++) {
             const int nx = imgs->data[i].nx;
             const int ny = imgs->data[i].ny;
-            GGML_ASSERT(nx == image_size && ny == image_size);
+            if (!ctx->has_minicpmv_projector) {
+                GGML_ASSERT(nx == image_size && ny == image_size);
+            }
 
             const int n = nx * ny;
 
@@ -1881,37 +2309,75 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         ggml_backend_tensor_set(inp_raw, data, 0, ggml_nbytes(inp_raw));
         free(data);
     }
+    if (ctx->has_minicpmv_projector) {
+        {
+            // inspired from siglip:
+            //    -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit
+            //    -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit/blob/d66538faeba44480d0bfaa42145eef26f9423199/modeling_siglip.py#L316
+            struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
+            int* positions_data = (int*)malloc(ggml_nbytes(positions));
+            for (int i = 0; i < num_positions; i++) {
+                positions_data[i] = std::floor(70.0*i/num_positions);
+            }
+            ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
+            free(positions_data);
+        }
 
-    {
-        if (ctx->has_class_embedding) {
-            struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings");
+        {
+            // inspired from resampler of Qwen-VL:
+            //    -> https://huggingface.co/Qwen/Qwen-VL/tree/main
+            //    -> https://huggingface.co/Qwen/Qwen-VL/blob/0547ed36a86561e2e42fecec8fd0c4f6953e33c4/visual.py#L23
+            struct ggml_tensor * pos_embed = ggml_graph_get_tensor(gf, "pos_embed");
+            if(ctx->load_image_size==nullptr){
+                ctx->load_image_size= clip_image_size_init();
+            }
+            int pos_w = ctx->load_image_size->width/patch_size;
+            int pos_h = ctx->load_image_size->height/patch_size;
+            int embed_dim = 4096;
+            auto pos_embed_t = get_2d_sincos_pos_embed(embed_dim, std::make_pair(pos_w, pos_h));
+
+            float * pos_embed_data = (float *)malloc(ggml_nbytes(pos_embed));
+            for(int i=0;i<pos_w * pos_h;++i){
+                for(int j=0;j<embed_dim;++j){
+                    pos_embed_data[i*embed_dim+j]=pos_embed_t[i][j];
+                }
+            }
 
-            void* zero_mem = malloc(ggml_nbytes(embeddings));
-            memset(zero_mem, 0, ggml_nbytes(embeddings));
-            ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings));
-            free(zero_mem);
+            ggml_backend_tensor_set(pos_embed, pos_embed_data, 0, ggml_nbytes(pos_embed));
+            free(pos_embed_data);
         }
-    }
+    } else {
+        {
+            if (ctx->has_class_embedding) {
+                struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings");
 
-    {
-        struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
+                void* zero_mem = malloc(ggml_nbytes(embeddings));
+                memset(zero_mem, 0, ggml_nbytes(embeddings));
+                ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings));
+                free(zero_mem);
+            }
+        }
+
+        {
+            struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
 
-        int* positions_data = (int*)malloc(ggml_nbytes(positions));
-        for (int i = 0; i < num_positions; i++) {
-            positions_data[i] = i;
+            int* positions_data = (int*)malloc(ggml_nbytes(positions));
+            for (int i = 0; i < num_positions; i++) {
+                positions_data[i] = i;
+            }
+            ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
+            free(positions_data);
         }
-        ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
-        free(positions_data);
-    }
 
-    {
-        struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
-        int* patches_data = (int*)malloc(ggml_nbytes(patches));
-        for (int i = 0; i < num_patches; i++) {
-            patches_data[i] = i + 1;
+        {
+            struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
+            int* patches_data = (int*)malloc(ggml_nbytes(patches));
+            for (int i = 0; i < num_patches; i++) {
+                patches_data[i] = i + 1;
+            }
+            ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
+            free(patches_data);
         }
-        ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
-        free(patches_data);
     }
 
     if (ggml_backend_is_cpu(ctx->backend)) {
@@ -2081,7 +2547,14 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
     if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
         return ctx->vision_model.mm_3_b->ne[0];
     }
+    if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
+        return 4096;
+    }
 
     std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type];
     throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
 }
+
+bool clip_is_minicpmv(const struct clip_ctx * ctx) {
+    return ctx->has_minicpmv_projector;
+}
diff --git a/examples/llava/clip.h b/examples/llava/clip.h
index ca363138..2ff4d399 100644
--- a/examples/llava/clip.h
+++ b/examples/llava/clip.h
@@ -18,14 +18,17 @@
 #    define CLIP_API
 #endif
 
-struct clip_ctx;
-
 #ifdef __cplusplus
 extern "C" {
 #endif
 
 struct clip_ctx;
 
+struct clip_image_size {
+    int width;
+    int height;
+};
+
 struct clip_image_u8_batch {
     struct clip_image_u8 * data;
     size_t size;
@@ -55,6 +58,10 @@ CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
 CLIP_API int clip_n_patches    (const struct clip_ctx * ctx);
 CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
 
+CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip);
+CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size);
+
+CLIP_API struct clip_image_size * clip_image_size_init();
 CLIP_API struct clip_image_u8  * clip_image_u8_init ();
 CLIP_API struct clip_image_f32 * clip_image_f32_init();
 
@@ -78,6 +85,8 @@ CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, cons
 
 CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype);
 
+CLIP_API bool clip_is_minicpmv(const struct clip_ctx * ctx);
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/examples/llava/llava.cpp b/examples/llava/llava.cpp
index 63878d17..916d9dc4 100644
--- a/examples/llava/llava.cpp
+++ b/examples/llava/llava.cpp
@@ -202,6 +202,33 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
     return true;
 }
 
+static clip_image_f32 * only_v2_5_reshape_by_patch(clip_image_f32 * image, int patch_size) {
+    int width = image->nx;
+    int height = image->ny;
+    int num_patches = (height / patch_size) * (width / patch_size);
+    clip_image_f32 * patch = clip_image_f32_init();
+    patch->nx = patch_size * num_patches;
+    patch->ny = patch_size;
+    patch->buf.resize(3 * patch->nx * patch->ny);
+
+    int patch_index = 0;
+
+    for (int i = 0; i < height; i += patch_size) {
+        for (int j = 0; j < width; j += patch_size) {
+            for (int pi = 0; pi < patch_size; ++pi) {
+                for (int pj = 0; pj < patch_size; ++pj) {
+                    int input_index = ((i + pi) * width + (j + pj)) * 3;
+                    int output_index = (pi * patch_size * num_patches + patch_index * patch_size + pj) * 3;
+                    patch->buf[output_index] = image->buf[input_index];
+                    patch->buf[output_index+1] = image->buf[input_index+1];
+                    patch->buf[output_index+2] = image->buf[input_index+2];
+                }
+            }
+            patch_index++;
+        }
+    }
+    return patch;
+}
 
 static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
     // std::vector<clip_image_f32*> img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336
@@ -218,7 +245,44 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
 
     const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip);
 
-    if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
+    if (clip_is_minicpmv(ctx_clip)) {
+        std::vector<float *> image_embd_v;
+        image_embd_v.resize(img_res_v.size);
+        struct clip_image_size * load_image_size = clip_image_size_init();
+        for (size_t i = 0; i < img_res_v.size; i++) {
+            const int64_t t_img_enc_step_start_us = ggml_time_us();
+            image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip));
+            int patch_size=14;
+            load_image_size->width = img_res_v.data[i].nx;
+            load_image_size->height = img_res_v.data[i].ny;
+            clip_add_load_image_size(ctx_clip, load_image_size);
+            const bool encoded = clip_image_encode(ctx_clip, n_threads, only_v2_5_reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]);
+            if (!encoded) {
+                LOG_TEE("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
+                return false;
+            }
+            const int64_t t_img_enc_steop_batch_us = ggml_time_us();
+            LOG_TEE("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)img_res_v.size, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0);
+        }
+        const int64_t t_img_enc_batch_us = ggml_time_us();
+        LOG_TEE("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
+
+        int n_img_pos_out = 0;
+        for (size_t i = 0; i < image_embd_v.size(); i++) {
+            std::memcpy(image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip), image_embd_v[i], clip_embd_nbytes(ctx_clip));
+            n_img_pos_out += clip_n_patches(ctx_clip);
+        }
+        *n_img_pos = n_img_pos_out;
+        for (size_t i = 0; i < image_embd_v.size(); i++) {
+            free(image_embd_v[i]);
+        }
+        image_embd_v.clear();
+        load_image_size->width = img->nx;
+        load_image_size->height = img->ny;
+        clip_add_load_image_size(ctx_clip, load_image_size);
+        LOG_TEE("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height);
+    }
+    else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
         // flat / default llava-1.5 type embedding
         *n_img_pos = clip_n_patches(ctx_clip);
         bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096
@@ -228,7 +292,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
 
             return false;
         }
-    } else {
+    }
+    else {
         // spatial_unpad llava-1.6 type embedding
         // TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working
         std::vector<float *> image_embd_v;
@@ -297,7 +362,11 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
 }
 
 bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
-    float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*6); // TODO: base on gridsize/llava model
+    int num_max_patches = 6;
+    if (clip_is_minicpmv(ctx_clip)) {
+        num_max_patches = 10;
+    }
+    float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model
     if (!image_embd) {
         LOG_TEE("Unable to allocate memory for image embeddings\n");
         return false;
diff --git a/examples/llava/llava.h b/examples/llava/llava.h
index 19212f6e..b6feb302 100644
--- a/examples/llava/llava.h
+++ b/examples/llava/llava.h
@@ -17,12 +17,11 @@
 #    define LLAVA_API
 #endif
 
-struct clip_ctx;
-
 #ifdef __cplusplus
 extern "C" {
 #endif
 
+struct clip_ctx;
 struct llava_image_embed {
     float * embed;
     int n_image_pos;
@@ -37,8 +36,8 @@ LLAVA_API bool llava_image_embed_make_with_clip_img(struct clip_ctx * ctx_clip,
 LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length);
 /** build an image embed from a path to an image filename */
 LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path);
-LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed);
 /** free an embedding made with llava_image_embed_make_* */
+LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed);
 
 /** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */
 LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past);
diff --git a/examples/llava/minicpmv-cli.cpp b/examples/llava/minicpmv-cli.cpp
new file mode 100644
index 00000000..f951b57b
--- /dev/null
+++ b/examples/llava/minicpmv-cli.cpp
@@ -0,0 +1,309 @@
+#include "ggml.h"
+#include "log.h"
+#include "common.h"
+#include "clip.h"
+#include "llava.h"
+#include "llama.h"
+
+#include <cstdio>
+#include <cstdlib>
+#include <vector>
+
+struct llava_context {
+    struct clip_ctx * ctx_clip = NULL;
+    struct llama_context * ctx_llama = NULL;
+    struct llama_model * model = NULL;
+};
+
+static void show_additional_info(int /*argc*/, char ** argv) {
+    LOG_TEE("\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
+    LOG_TEE("  note: a lower temperature value like 0.1 is recommended for better quality.\n");
+}
+
+static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
+    (void) level;
+    (void) user_data;
+    LOG_TEE("%s", text);
+}
+
+static struct llama_model * llava_init(gpt_params * params) {
+    llama_backend_init();
+    llama_numa_init(params->numa);
+
+    llama_model_params model_params = llama_model_params_from_gpt_params(*params);
+
+    llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
+    if (model == NULL) {
+        LOG_TEE("%s: error: unable to load model\n" , __func__);
+        return NULL;
+    }
+    return model;
+}
+
+static struct llava_context * llava_init_context(gpt_params * params, llama_model * model) {
+    auto prompt = params->prompt;
+    if (prompt.empty()) {
+        prompt = "describe the image in detail.";
+    }
+
+    llama_context_params ctx_params = llama_context_params_from_gpt_params(*params);
+    if (params->n_ctx < 2048) {
+        // warn user here, "Image processing requires at least 2048 context, setting context to 2048"
+        LOG_TEE("%s: warn: Image processing requires at least 2048 context, setting context to 2048\n" , __func__);
+        ctx_params.n_ctx = 2048;
+    } else {
+        ctx_params.n_ctx = params->n_ctx;
+    }
+
+    llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
+
+    if (ctx_llama == NULL) {
+        LOG_TEE("%s: error: failed to create the llama_context\n" , __func__);
+        return NULL;
+    }
+
+    auto ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
+
+    ctx_llava->ctx_llama = ctx_llama;
+    ctx_llava->model = model;
+    return ctx_llava;
+}
+
+static void llava_free(struct llava_context * ctx_llava) {
+    if (ctx_llava->ctx_clip) {
+        clip_free(ctx_llava->ctx_clip);
+        ctx_llava->ctx_clip = NULL;
+    }
+
+    llama_free(ctx_llava->ctx_llama);
+    llama_free_model(ctx_llava->model);
+    llama_backend_free();
+}
+
+static struct clip_ctx * clip_init_context(gpt_params * params) {
+    const char * clip_path = params->mmproj.c_str();
+
+    auto prompt = params->prompt;
+    if (prompt.empty()) {
+        prompt = "describe the image in detail.";
+    }
+    auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
+    return ctx_clip;
+}
+
+static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past) {
+    int N = (int) tokens.size();
+    for (int i = 0; i < N; i += n_batch) {
+        int n_eval = (int) tokens.size() - i;
+        if (n_eval > n_batch) {
+            n_eval = n_batch;
+        }
+        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
+            LOG_TEE("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
+            return false;
+        }
+        *n_past += n_eval;
+    }
+    return true;
+}
+
+static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
+    std::vector<llama_token> tokens;
+    tokens.push_back(id);
+    return eval_tokens(ctx_llama, tokens, 1, n_past);
+}
+
+static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
+    std::string              str2     = str;
+    std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos, true);
+    return eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
+}
+
+static void process_eval_image_embed(struct llava_context * ctx_llava, const struct llava_image_embed * embeds, int n_batch, int * n_past, int idx) {
+    float * image_embed = (float *)malloc(clip_embd_nbytes(ctx_llava->ctx_clip));
+    std::memcpy(image_embed, embeds->embed + idx * clip_n_patches(ctx_llava->ctx_clip) * clip_n_mmproj_embd(ctx_llava->ctx_clip), clip_embd_nbytes(ctx_llava->ctx_clip));
+
+    auto slice_embed = (llava_image_embed*)malloc(sizeof(llava_image_embed));
+    slice_embed->embed = image_embed;
+    slice_embed->n_image_pos = clip_n_patches(ctx_llava->ctx_clip);
+    llava_eval_image_embed(ctx_llava->ctx_llama, slice_embed, n_batch, n_past);
+    llava_image_embed_free(slice_embed);
+}
+
+static void process_image(struct llava_context * ctx_llava, struct llava_image_embed * embeds, gpt_params * params, int &n_past) {
+    std::string system_prompt;
+    int idx = 0;
+    int num_image_embeds = embeds->n_image_pos / clip_n_patches(ctx_llava->ctx_clip);
+    system_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n";
+    LOG_TEE("%s: image token past: %d\n", __func__, n_past);
+    eval_string(ctx_llava->ctx_llama, (system_prompt+"<image>").c_str(), params->n_batch, &n_past, false);
+    process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
+    eval_string(ctx_llava->ctx_llama, std::string("</image>").c_str(), params->n_batch, &n_past, false);
+    if (num_image_embeds > 1) {
+        size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip);
+        eval_string(ctx_llava->ctx_llama, std::string("<slice>").c_str(), params->n_batch, &n_past, false);
+        for (size_t i = 0; i < (num_image_embeds-1)/num_image_embeds_col; ++i) {
+            for (size_t j = 0; j < num_image_embeds_col; ++j) {
+                eval_string(ctx_llava->ctx_llama, std::string("<image>").c_str(), params->n_batch, &n_past, false);
+                process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
+                eval_string(ctx_llava->ctx_llama, std::string("</image>").c_str(), params->n_batch, &n_past, false);
+                if (j == num_image_embeds_col - 1) {
+                    eval_string(ctx_llava->ctx_llama, std::string("\n").c_str(), params->n_batch, &n_past, false);
+                }
+            }
+        }
+        eval_string(ctx_llava->ctx_llama, std::string("</slice>").c_str(), params->n_batch, &n_past, false);
+    }
+    LOG_TEE("%s: image token past: %d\n", __func__, n_past);
+}
+
+static const char * sample(struct llama_sampling_context * ctx_sampling,
+                           struct llama_context * ctx_llama,
+                           int * n_past) {
+    const llama_token id = llama_sampling_sample(ctx_sampling, ctx_llama, NULL);
+    llama_sampling_accept(ctx_sampling, ctx_llama, id, true);
+    static std::string ret;
+    if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
+        ret = "</s>";
+    } else {
+        ret = llama_token_to_piece(ctx_llama, id);
+    }
+    eval_id(ctx_llama, id, n_past);
+    return ret.c_str();
+}
+
+static struct llava_context * minicpmv_init(gpt_params * params, const std::string & fname, int &n_past){
+    auto ctx_clip = clip_init_context(params);
+    auto embeds = llava_image_embed_make_with_filename(ctx_clip, params->n_threads, fname.c_str());
+    if (!embeds) {
+        std::cerr << "error: failed to load image " << fname << ". Terminating\n\n";
+        return NULL;
+    }
+
+    // process the prompt
+    if (params->prompt.empty() && params->interactive == false) {
+        LOG_TEE("prompt should be given or interactive mode should be on");
+        return NULL;
+    }
+
+    auto model = llava_init(params);
+    if (model == NULL) {
+        fprintf(stderr, "%s: error: failed to init minicpmv model\n", __func__);
+        return NULL;
+    }
+    const int64_t t_llava_init_start_us = ggml_time_us();
+    auto ctx_llava = llava_init_context(params, model);
+    ctx_llava->ctx_clip = ctx_clip;
+    const int64_t t_llava_init_end_us = ggml_time_us();
+    float t_llava_init_ms = (t_llava_init_end_us - t_llava_init_start_us) / 1000.0;
+    LOG_TEE("\n%s: llava init in %8.2f ms.\n", __func__, t_llava_init_ms);
+
+    const int64_t t_process_image_start_us = ggml_time_us();
+    process_image(ctx_llava, embeds, params, n_past);
+    const int64_t t_process_image_end_us = ggml_time_us();
+    float t_process_image_ms = (t_process_image_end_us - t_process_image_start_us) / 1000.0;
+    LOG_TEE("\n%s: llama process image in %8.2f ms.\n", __func__, t_process_image_ms);
+
+    llava_image_embed_free(embeds);
+    return ctx_llava;
+}
+
+static struct llama_sampling_context * llama_init(struct llava_context * ctx_llava, gpt_params * params, std::string prompt, int &n_past, bool is_first = false){
+    std::string user_prompt = prompt;
+    if (!is_first) user_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n" + prompt;
+
+    eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false);
+    eval_string(ctx_llava->ctx_llama, "<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n", params->n_batch, &n_past, false);
+    // generate the response
+
+    LOG_TEE("\n");
+
+    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
+    return ctx_sampling;
+}
+
+static const char * llama_loop(struct llava_context * ctx_llava,struct llama_sampling_context * ctx_sampling, int &n_past){
+
+    const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
+    return tmp;
+}
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    gpt_params params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        show_additional_info(argc, argv);
+        return 1;
+    }
+
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("llava", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+    llama_log_set(llama_log_callback_logTee, nullptr);
+#endif // LOG_DISABLE_LOGS
+
+    if (params.mmproj.empty() || (params.image.empty())) {
+        gpt_params_print_usage(argc, argv, params);
+        show_additional_info(argc, argv);
+        return 1;
+    }
+
+    for (auto & image : params.image) {
+        int n_past = 0;
+        auto ctx_llava = minicpmv_init(&params, image, n_past);
+
+        if (!params.prompt.empty()) {
+            LOG_TEE("<user>%s\n", params.prompt.c_str());
+            LOG_TEE("<assistant>");
+            auto ctx_sampling = llama_init(ctx_llava, &params, params.prompt.c_str(), n_past, true);
+            const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
+            std::string response = "";
+            bool have_tmp = false;
+            for (int i = 0; i < max_tgt_len; i++) {
+                auto tmp = llama_loop(ctx_llava, ctx_sampling, n_past);
+                response += tmp;
+                if (strcmp(tmp, "</s>") == 0){
+                    if(!have_tmp)continue;
+                    else break;
+                }
+                if (strstr(tmp, "###")) break; // Yi-VL behavior
+                have_tmp = true;
+                printf("%s", tmp);
+                if (strstr(response.c_str(), "<user>")) break; // minicpm-v
+
+                fflush(stdout);
+            }
+            llama_sampling_free(ctx_sampling);
+        }else {
+            while (true) {
+                LOG_TEE("<user>");
+                std::string prompt;
+                std::getline(std::cin, prompt);
+                LOG_TEE("<assistant>");
+                auto ctx_sampling = llama_init(ctx_llava, &params, prompt, n_past, true);
+                const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
+                std::string response = "";
+                for (int i = 0; i < max_tgt_len; i++) {
+                    auto tmp = llama_loop(ctx_llava, ctx_sampling, n_past);
+                    response += tmp;
+                    if (strcmp(tmp, "</s>") == 0) break;
+                    if (strstr(tmp, "###")) break; // Yi-VL behavior
+                    printf("%s", tmp);// mistral llava-1.6
+                    if (strstr(response.c_str(), "<user>")) break; // minicpm-v
+                    fflush(stdout);
+                }
+                llama_sampling_free(ctx_sampling);
+            }
+        }
+        printf("\n");
+        llama_print_timings(ctx_llava->ctx_llama);
+
+        ctx_llava->model = NULL;
+        llava_free(ctx_llava);
+    }
+
+    return 0;
+}
diff --git a/examples/llava/minicpmv-convert-image-encoder-to-gguf.py b/examples/llava/minicpmv-convert-image-encoder-to-gguf.py
new file mode 100644
index 00000000..12cdd128
--- /dev/null
+++ b/examples/llava/minicpmv-convert-image-encoder-to-gguf.py
@@ -0,0 +1,382 @@
+import argparse
+import os
+import json
+import re
+
+import torch
+import numpy as np
+from gguf import *
+from transformers.models.idefics2.modeling_idefics2 import Idefics2VisionTransformer, Idefics2VisionConfig
+
+TEXT = "clip.text"
+VISION = "clip.vision"
+
+
+def add_key_str(raw_key: str, arch: str) -> str:
+    return raw_key.format(arch=arch)
+
+
+def should_skip_tensor(name: str, has_text: bool, has_vision: bool, has_minicpmv: bool) -> bool:
+    if name in (
+        "logit_scale",
+        "text_model.embeddings.position_ids",
+        "vision_model.embeddings.position_ids",
+    ):
+        return True
+
+    if has_minicpmv and name in ["visual_projection.weight"]:
+        return True
+
+    if name.startswith("v") and not has_vision:
+        return True
+
+    if name.startswith("t") and not has_text:
+        return True
+
+    return False
+
+
+def get_tensor_name(name: str) -> str:
+    if "projection" in name:
+        return name
+    if "mm_projector" in name:
+        name = name.replace("model.mm_projector", "mm")
+        name = re.sub(r'mm\.mlp\.mlp', 'mm.model.mlp', name, count=1)
+        name = re.sub(r'mm\.peg\.peg', 'mm.model.peg', name, count=1)
+        return name
+
+    return name.replace("text_model", "t").replace("vision_model", "v").replace("encoder.layers", "blk").replace("embeddings.", "").replace("_proj", "").replace("self_attn.", "attn_").replace("layer_norm", "ln").replace("layernorm", "ln").replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("embedding", "embd").replace("final", "post").replace("layrnorm", "ln")
+
+
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1))
+        + list(range(ord("¡"), ord("¬") + 1))
+        + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model-dir", help="Path to model directory cloned from HF Hub", required=True)
+ap.add_argument("--use-f32", action="store_true", default=False, help="Use f32 instead of f16")
+ap.add_argument("--text-only", action="store_true", required=False,
+                help="Save a text-only model. It can't be used to encode images")
+ap.add_argument("--vision-only", action="store_true", required=False,
+                help="Save a vision-only model. It can't be used to encode texts")
+ap.add_argument("--clip-model-is-vision", action="store_true", required=False,
+                help="The clip model is a pure vision model (ShareGPT4V vision extract for example)")
+ap.add_argument("--clip-model-is-openclip", action="store_true", required=False,
+                help="The clip model is from openclip (for ViT-SO400M type))")
+ap.add_argument("--minicpmv-projector", help="Path to minicpmv.projector file. If specified, save an image encoder for MiniCPM-V models.")
+ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp, ldpv2", choices=["mlp", "ldp", "ldpv2"], default="mlp")
+ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
+# Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711
+# Example --image_mean 0.5 0.5 0.5 --image_std 0.5 0.5 0.5
+default_image_mean = [0.48145466, 0.4578275, 0.40821073]
+default_image_std = [0.26862954, 0.26130258, 0.27577711]
+ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
+ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
+
+# with proper
+args = ap.parse_args()
+
+
+if args.text_only and args.vision_only:
+    print("--text-only and --image-only arguments cannot be specified at the same time.")
+    exit(1)
+
+if args.use_f32:
+    print("WARNING: Weights for the convolution op is always saved in f16, as the convolution op in GGML does not support 32-bit kernel weights yet.")
+
+# output in the same directory as the model if output_dir is None
+dir_model = args.model_dir
+
+if args.clip_model_is_vision or not os.path.exists(dir_model + "/vocab.json") or args.clip_model_is_openclip:
+    vocab = None
+    tokens = None
+else:
+    with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
+        vocab = json.load(f)
+        tokens = [key for key in vocab]
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if args.use_f32:
+    ftype = 0
+
+# if args.clip_model_is_vision or args.clip_model_is_openclip:
+#     model = CLIPVisionModel.from_pretrained(dir_model)
+#     processor = None
+# else:
+#     model = CLIPModel.from_pretrained(dir_model)
+#     processor = CLIPProcessor.from_pretrained(dir_model)
+
+default_vision_config = {
+        "hidden_size": 1152,
+        "image_size": 980,
+        "intermediate_size": 4304,
+        "model_type": "idefics2",
+        "num_attention_heads": 16,
+        "num_hidden_layers": 27,
+        "patch_size": 14,
+    }
+vision_config = Idefics2VisionConfig(**default_vision_config)
+model = Idefics2VisionTransformer(vision_config)
+
+processor = None
+# if model.attn_pool is not None:
+#     model.attn_pool = torch.nn.Identity()
+
+# model.blocks = model.blocks[:-1]
+model.load_state_dict(torch.load(os.path.join(dir_model, "minicpmv.clip")))
+
+fname_middle = None
+has_text_encoder = True
+has_vision_encoder = True
+has_minicpmv_projector = False
+if args.text_only:
+    fname_middle = "text-"
+    has_vision_encoder = False
+elif args.minicpmv_projector is not None:
+    fname_middle = "mmproj-"
+    has_text_encoder = False
+    has_minicpmv_projector = True
+elif args.vision_only:
+    fname_middle = "vision-"
+    has_text_encoder = False
+else:
+    fname_middle = ""
+
+output_dir = args.output_dir if args.output_dir is not None else dir_model
+os.makedirs(output_dir, exist_ok=True)
+output_prefix = os.path.basename(output_dir).replace("ggml_", "")
+fname_out = os.path.join(output_dir, f"{fname_middle}model-{ftype_str[ftype]}.gguf")
+fout = GGUFWriter(path=fname_out, arch="clip")
+
+fout.add_bool("clip.has_text_encoder", has_text_encoder)
+fout.add_bool("clip.has_vision_encoder", has_vision_encoder)
+fout.add_bool("clip.has_minicpmv_projector", has_minicpmv_projector)
+fout.add_file_type(ftype)
+if args.text_only:
+    fout.add_description("text-only CLIP model")
+elif args.vision_only and not has_minicpmv_projector:
+    fout.add_description("vision-only CLIP model")
+elif has_minicpmv_projector:
+    fout.add_description("image encoder for MiniCPM-V")
+    # add projector type
+    fout.add_string("clip.projector_type", "resampler")
+else:
+    fout.add_description("two-tower CLIP model")
+
+if has_vision_encoder:
+    # vision_model hparams
+    fout.add_uint32("clip.vision.image_size", 448)
+    fout.add_uint32("clip.vision.patch_size", 14)
+    fout.add_uint32(add_key_str(KEY_EMBEDDING_LENGTH, VISION), 1152)
+    fout.add_uint32(add_key_str(KEY_FEED_FORWARD_LENGTH, VISION), 4304)
+    fout.add_uint32("clip.vision.projection_dim", 0)
+    fout.add_uint32(add_key_str(KEY_ATTENTION_HEAD_COUNT, VISION), 16)
+    fout.add_float32(add_key_str(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
+    block_count = 26
+    fout.add_uint32(add_key_str(KEY_BLOCK_COUNT, VISION), block_count)
+
+    if processor is not None:
+        image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean
+        image_std = processor.image_processor.image_std if args.image_std is None or args.image_std == default_image_std else args.image_std
+    else:
+        image_mean = args.image_mean if args.image_mean is not None else default_image_mean
+        image_std = args.image_std if args.image_std is not None else default_image_std
+    fout.add_array("clip.vision.image_mean", image_mean)
+    fout.add_array("clip.vision.image_std", image_std)
+
+use_gelu = True
+fout.add_bool("clip.use_gelu", use_gelu)
+
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
+    """
+    embed_dim: output dimension for each position
+    pos: a list of positions to be encoded: size (M,)
+    out: (M, D)
+    """
+    assert embed_dim % 2 == 0
+    omega = np.arange(embed_dim // 2, dtype=np.float32)
+    omega /= embed_dim / 2.
+    omega = 1. / 10000 ** omega  # (D/2,)
+
+    pos = pos.reshape(-1)  # (M,)
+    out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product
+
+    emb_sin = np.sin(out)  # (M, D/2)
+    emb_cos = np.cos(out)  # (M, D/2)
+
+    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
+    return emb
+
+def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
+    assert embed_dim % 2 == 0
+
+    # use half of dimensions to encode grid_h
+    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
+    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)
+
+    emb = np.concatenate([emb_h, emb_w], axis=1)  # (H*W, D)
+    return emb
+
+
+# https://github.com/facebookresearch/mae/blob/efb2a8062c206524e35e47d04501ed4f544c0ae8/util/pos_embed.py#L20
+def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False):
+    """
+    grid_size: int of the grid height and width
+    return:
+    pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
+    """
+    if isinstance(grid_size, int):
+        grid_h_size, grid_w_size = grid_size, grid_size
+    else:
+        grid_h_size, grid_w_size = grid_size[0], grid_size[1]
+
+    grid_h = np.arange(grid_h_size, dtype=np.float32)
+    grid_w = np.arange(grid_w_size, dtype=np.float32)
+    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
+    grid = np.stack(grid, axis=0)
+
+    grid = grid.reshape([2, 1, grid_h_size, grid_w_size])
+    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
+    if cls_token:
+        pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
+    return pos_embed
+
+def _replace_name_resampler(s, v):
+    if re.match("resampler.pos_embed", s):
+        return {
+            s: v,
+            re.sub("pos_embed", "pos_embed_k", s): torch.from_numpy(get_2d_sincos_pos_embed(4096, (70, 70))),
+        }
+    if re.match("resampler.proj", s):
+        return {
+            re.sub("proj", "pos_embed_k", s): torch.from_numpy(get_2d_sincos_pos_embed(4096, (70, 70))),
+            re.sub("proj", "proj.weight", s): v.transpose(-1, -2).contiguous(),
+        }
+    if re.match("resampler.attn.in_proj_.*", s):
+        return {
+            re.sub("attn.in_proj_", "attn.q.", s): v.chunk(3, dim=0)[0],
+            re.sub("attn.in_proj_", "attn.k.", s): v.chunk(3, dim=0)[1],
+            re.sub("attn.in_proj_", "attn.v.", s): v.chunk(3, dim=0)[2],
+        }
+    return {s: v}
+
+if has_minicpmv_projector:
+    projector = torch.load(args.minicpmv_projector)
+    new_state_dict = {}
+    for k, v in projector.items():
+        kvs = _replace_name_resampler(k, v)
+        for nk, nv in kvs.items():
+            new_state_dict[nk] = nv
+    projector = new_state_dict
+    ftype_cur = 0
+    for name, data in projector.items():
+        name = get_tensor_name(name)
+        data = data.squeeze().numpy()
+
+        n_dims = len(data.shape)
+        if ftype == 1:
+            if name[-7:] == ".weight" and n_dims == 2:
+                print("  Converting to float16")
+                data = data.astype(np.float16)
+                ftype_cur = 1
+            else:
+                print("  Converting to float32")
+                data = data.astype(np.float32)
+                ftype_cur = 0
+        else:
+            if data.dtype != np.float32:
+                print("  Converting to float32")
+                data = data.astype(np.float32)
+                ftype_cur = 0
+
+        fout.add_tensor(name, data)
+        print(f"{name} - {ftype_str[ftype_cur]} - shape = {data.shape}")
+
+    print("Projector tensors added\n")
+
+def _replace_name(s, v):
+    s = "vision_model." + s
+    if re.match("vision_model.embeddings.position_embedding", s):
+        v = v.unsqueeze(0)
+        return {s: v}
+
+    return {s: v}
+
+state_dict = model.state_dict()
+new_state_dict = {}
+for k, v in state_dict.items():
+    kvs = _replace_name(k, v)
+    for nk, nv in kvs.items():
+        new_state_dict[nk] = nv
+state_dict = new_state_dict
+for name, data in state_dict.items():
+    if should_skip_tensor(name, has_text_encoder, has_vision_encoder, has_minicpmv_projector):
+        # we don't need this
+        print(f"skipping parameter: {name}")
+        continue
+
+    name = get_tensor_name(name)
+    data = data.squeeze().numpy()
+
+    n_dims = len(data.shape)
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype_cur = 0
+    if n_dims == 4:
+        print(f"tensor {name} is always saved in f16")
+        data = data.astype(np.float16)
+        ftype_cur = 1
+    elif ftype == 1:
+        if name[-7:] == ".weight" and n_dims == 2:
+            print("  Converting to float16")
+            data = data.astype(np.float16)
+            ftype_cur = 1
+        else:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+    else:
+        if data.dtype != np.float32:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+
+    print(f"{name} - {ftype_str[ftype_cur]} - shape = {data.shape}")
+    fout.add_tensor(name, data)
+
+
+fout.write_header_to_file()
+fout.write_kv_data_to_file()
+fout.write_tensors_to_file()
+fout.close()
+
+print("Done. Output file: " + fname_out)
diff --git a/examples/llava/minicpmv-surgery.py b/examples/llava/minicpmv-surgery.py
new file mode 100644
index 00000000..2b6bce7c
--- /dev/null
+++ b/examples/llava/minicpmv-surgery.py
@@ -0,0 +1,47 @@
+import argparse
+import os
+import torch
+from transformers import AutoModel, AutoTokenizer
+
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model", help="Path to MiniCPM-V-2.5 model")
+args = ap.parse_args()
+
+# find the model part that includes the the multimodal projector weights
+model = AutoModel.from_pretrained(args.model, trust_remote_code=True, local_files_only=True)
+checkpoint = model.state_dict()
+
+# get a list of mm tensor names
+mm_tensors = [k for k, v in checkpoint.items() if k.startswith("resampler")]
+
+# store these tensors in a new dictionary and torch.save them
+projector = {name: checkpoint[name].float() for name in mm_tensors}
+torch.save(projector, f"{args.model}/minicpmv.projector")
+
+clip_tensors = [k for k, v in checkpoint.items() if k.startswith("vpm")]
+if len(clip_tensors) > 0:
+    clip = {name.replace("vpm.", ""): checkpoint[name].float() for name in clip_tensors}
+    torch.save(clip, f"{args.model}/minicpmv.clip")
+
+    # added tokens should be removed to be able to convert Mistral models
+    if os.path.exists(f"{args.model}/added_tokens.json"):
+        with open(f"{args.model}/added_tokens.json", "w") as f:
+            f.write("{}\n")
+
+config = model.llm.config
+config._name_or_path = "openbmb/MiniCPM-Llama3-V-2.5"
+config.auto_map = {
+    "AutoConfig": "configuration_minicpm.MiniCPMConfig",
+    "AutoModel": "modeling_minicpm.MiniCPMModel",
+    "AutoModelForCausalLM": "modeling_minicpm.MiniCPMForCausalLM",
+    "AutoModelForSeq2SeqLM": "modeling_minicpm.MiniCPMForCausalLM",
+    "AutoModelForSequenceClassification": "modeling_minicpm.MiniCPMForSequenceClassification"
+}
+model.llm.save_pretrained(f"{args.model}/model")
+tok = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True)
+tok.save_pretrained(f"{args.model}/model")
+# os.system(f"cp {args.model}/modeling_minicpm.py {args.model}/MiniCPM_l3/modeling_minicpm.py")
+
+print("Done!")
+print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
+print(f"Also, use {args.model}/minicpmv.projector to prepare a minicpmv-encoder.gguf file.")
diff --git a/examples/llava/requirements.txt b/examples/llava/requirements.txt
index fbfd0cdd..dfe5fbe6 100644
--- a/examples/llava/requirements.txt
+++ b/examples/llava/requirements.txt
@@ -2,3 +2,4 @@
 --extra-index-url https://download.pytorch.org/whl/cpu
 pillow~=10.2.0
 torch~=2.2.1
+torchvision==0.17.1
diff --git a/examples/lookahead/lookahead.cpp b/examples/lookahead/lookahead.cpp
index fb20ad93..81cf1629 100644
--- a/examples/lookahead/lookahead.cpp
+++ b/examples/lookahead/lookahead.cpp
@@ -58,11 +58,11 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the target model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // Tokenize the prompt
     std::vector<llama_token> inp;
diff --git a/examples/lookup/lookup-create.cpp b/examples/lookup/lookup-create.cpp
index d713f6f2..5f04709f 100644
--- a/examples/lookup/lookup-create.cpp
+++ b/examples/lookup/lookup-create.cpp
@@ -22,11 +22,11 @@ int main(int argc, char ** argv){
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     GGML_ASSERT(model != nullptr);
 
     // tokenize the prompt
diff --git a/examples/lookup/lookup-stats.cpp b/examples/lookup/lookup-stats.cpp
index 2fe67100..400f3e0b 100644
--- a/examples/lookup/lookup-stats.cpp
+++ b/examples/lookup/lookup-stats.cpp
@@ -26,11 +26,11 @@ int main(int argc, char ** argv){
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // tokenize the prompt
     std::vector<llama_token> inp;
diff --git a/examples/lookup/lookup.cpp b/examples/lookup/lookup.cpp
index bb571bac..d53a9828 100644
--- a/examples/lookup/lookup.cpp
+++ b/examples/lookup/lookup.cpp
@@ -34,11 +34,11 @@ int main(int argc, char ** argv){
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // tokenize the prompt
     std::vector<llama_token> inp;
diff --git a/examples/main/main.cpp b/examples/main/main.cpp
index 61e960ea..6e0635a6 100644
--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@@ -207,7 +207,10 @@ int main(int argc, char ** argv) {
 
     // load the model and apply lora adapter, if any
     LOG("%s: load the model and apply lora adapter, if any\n", __func__);
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    model = llama_init.model;
+    ctx = llama_init.context;
     if (sparams.cfg_scale > 1.f) {
         struct llama_context_params lparams = llama_context_params_from_gpt_params(params);
         ctx_guidance = llama_new_context_with_model(model, lparams);
diff --git a/examples/parallel/parallel.cpp b/examples/parallel/parallel.cpp
index 7faeaec9..621a1c95 100644
--- a/examples/parallel/parallel.cpp
+++ b/examples/parallel/parallel.cpp
@@ -129,11 +129,11 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the target model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // load the prompts from an external file if there are any
     if (params.prompt.empty()) {
diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp
index dbe44539..372684f0 100644
--- a/examples/perplexity/perplexity.cpp
+++ b/examples/perplexity/perplexity.cpp
@@ -2018,11 +2018,11 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model;
-    llama_context * ctx;
-
     // load the model and apply lora adapter, if any
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == NULL) {
         fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;
diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp
index 10c0299b..9a08d625 100644
--- a/examples/quantize/quantize.cpp
+++ b/examples/quantize/quantize.cpp
@@ -99,7 +99,7 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
 }
 
 // usage:
-//  ./quantize [--allow-requantize] [--leave-output-tensor] [--pure] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
+//  ./llama-quantize [--allow-requantize] [--leave-output-tensor] [--pure] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
 //
 [[noreturn]]
 static void usage(const char * executable) {
diff --git a/examples/retrieval/retrieval.cpp b/examples/retrieval/retrieval.cpp
index eb89d16d..65b19ce7 100644
--- a/examples/retrieval/retrieval.cpp
+++ b/examples/retrieval/retrieval.cpp
@@ -148,11 +148,12 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model;
-    llama_context * ctx;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
+
     if (model == NULL) {
         fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;
diff --git a/examples/rpc/README.md b/examples/rpc/README.md
index e1da801f..adedc890 100644
--- a/examples/rpc/README.md
+++ b/examples/rpc/README.md
@@ -1,5 +1,9 @@
 ## Overview
 
+> [!IMPORTANT]
+> This example and the RPC backend are currently in a proof-of-concept development stage. As such, the functionality is fragile and
+> insecure. **Never run the RPC server on an open network or in a sensitive environment!**
+
 The `rpc-server` allows  running `ggml` backend on a remote host.
 The RPC backend communicates with one or several instances of `rpc-server` and offloads computations to them.
 This can be used for distributed LLM inference with `llama.cpp` in the following way:
diff --git a/examples/rpc/rpc-server.cpp b/examples/rpc/rpc-server.cpp
index 7c15d2aa..6342e648 100644
--- a/examples/rpc/rpc-server.cpp
+++ b/examples/rpc/rpc-server.cpp
@@ -16,7 +16,7 @@
 #include <stdio.h>
 
 struct rpc_server_params {
-    std::string host        = "0.0.0.0";
+    std::string host        = "127.0.0.1";
     int         port        = 50052;
     size_t      backend_mem = 0;
 };
@@ -114,6 +114,17 @@ int main(int argc, char * argv[]) {
         fprintf(stderr, "Invalid parameters\n");
         return 1;
     }
+
+    if (params.host != "127.0.0.1") {
+        fprintf(stderr, "\n");
+        fprintf(stderr, "!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n");
+        fprintf(stderr, "WARNING: Host ('%s') is != '127.0.0.1'\n", params.host.c_str());
+        fprintf(stderr, "         Never expose the RPC server to an open network!\n");
+        fprintf(stderr, "         This is an experimental feature and is not secure!\n");
+        fprintf(stderr, "!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n");
+        fprintf(stderr, "\n");
+    }
+
     ggml_backend_t backend = create_backend();
     if (!backend) {
         fprintf(stderr, "Failed to create backend\n");
diff --git a/examples/save-load-state/save-load-state.cpp b/examples/save-load-state/save-load-state.cpp
index 00c2277a..3ea7c790 100644
--- a/examples/save-load-state/save-load-state.cpp
+++ b/examples/save-load-state/save-load-state.cpp
@@ -28,10 +28,11 @@ int main(int argc, char ** argv) {
     std::string result2;
 
     // init
-    llama_model * model;
-    llama_context * ctx;
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
     if (model == nullptr || ctx == nullptr) {
         fprintf(stderr, "%s : failed to init\n", __func__);
         return 1;
@@ -47,7 +48,7 @@ int main(int argc, char ** argv) {
     // save state (rng, logits, embedding and kv_cache) to file
     {
         std::vector<uint8_t> state_mem(llama_state_get_size(ctx));
-        const size_t written = llama_state_get_data(ctx, state_mem.data());
+        const size_t written = llama_state_get_data(ctx, state_mem.data(), state_mem.size());
 
         FILE *fp_write = fopen("dump_state.bin", "wb");
         fwrite(state_mem.data(), 1, written, fp_write);
@@ -99,13 +100,16 @@ int main(int argc, char ** argv) {
 
     // load state (rng, logits, embedding and kv_cache) from file
     {
-        std::vector<uint8_t> state_mem(llama_state_get_size(ctx2));
+        std::vector<uint8_t> state_mem;
 
         FILE * fp_read = fopen("dump_state.bin", "rb");
+        fseek(fp_read, 0, SEEK_END);
+        state_mem.resize(ftell(fp_read));
+        fseek(fp_read, 0, SEEK_SET);
         const size_t read = fread(state_mem.data(), 1, state_mem.size(), fp_read);
         fclose(fp_read);
 
-        if (read != llama_state_set_data(ctx2, state_mem.data())) {
+        if (read != llama_state_set_data(ctx2, state_mem.data(), state_mem.size())) {
             fprintf(stderr, "\n%s : failed to read state\n", __func__);
             llama_free(ctx2);
             llama_free_model(model);
@@ -159,13 +163,16 @@ int main(int argc, char ** argv) {
 
     // load state (rng, logits, embedding and kv_cache) from file
     {
-        std::vector<uint8_t> state_mem(llama_state_get_size(ctx3));
+        std::vector<uint8_t> state_mem;
 
         FILE * fp_read = fopen("dump_state.bin", "rb");
+        fseek(fp_read, 0, SEEK_END);
+        state_mem.resize(ftell(fp_read));
+        fseek(fp_read, 0, SEEK_SET);
         const size_t read = fread(state_mem.data(), 1, state_mem.size(), fp_read);
         fclose(fp_read);
 
-        if (read != llama_state_set_data(ctx3, state_mem.data())) {
+        if (read != llama_state_set_data(ctx3, state_mem.data(), state_mem.size())) {
             fprintf(stderr, "\n%s : failed to read state\n", __func__);
             llama_free(ctx3);
             llama_free_model(model);
@@ -182,7 +189,7 @@ int main(int argc, char ** argv) {
     {
         // save kv of seq 0
         std::vector<uint8_t> seq_store(llama_state_seq_get_size(ctx3, 0));
-        const size_t ncopy = llama_state_seq_get_data(ctx3, seq_store.data(), 0);
+        const size_t ncopy = llama_state_seq_get_data(ctx3, seq_store.data(), seq_store.size(), 0);
         if (ncopy != seq_store.size()) {
             fprintf(stderr, "\n%s : seq copy data length %zd does not match expected length %zd\n", __func__, ncopy, seq_store.size());
             llama_free(ctx3);
@@ -196,7 +203,7 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s : kv cache cleared\n", __func__);
 
         // restore kv into seq 1
-        const size_t nset = llama_state_seq_set_data(ctx3, seq_store.data(), 1);
+        const size_t nset = llama_state_seq_set_data(ctx3, seq_store.data(), seq_store.size(), 1);
         if (nset != seq_store.size()) {
             fprintf(stderr, "\n%s : seq set data length %zd does not match expected length %zd\n", __func__, nset, seq_store.size());
             llama_free(ctx3);
diff --git a/examples/server/README.md b/examples/server/README.md
index 33a2b95c..e17595fe 100644
--- a/examples/server/README.md
+++ b/examples/server/README.md
@@ -207,47 +207,12 @@ model:
   -hff,  --hf-file FILE           Hugging Face model file (default: unused)
   -hft,  --hf-token TOKEN         Hugging Face access token (default: value from HF_TOKEN environment variable)
 
-retrieval:
-
-         --context-file FNAME     file to load context from (repeat to specify multiple files)
-         --chunk-size N           minimum length of embedded text chunks (default: 64)
-         --chunk-separator STRING
-                                  separator between chunks (default: '
-                                  ')
-
-passkey:
-
-         --junk N                 number of times to repeat the junk text (default: 250)
-         --pos N                  position of the passkey in the junk text (default: -1)
-
-imatrix:
-
-  -o,    --output FNAME           output file (default: 'imatrix.dat')
-         --output-frequency N     output the imatrix every N iterations (default: 10)
-         --save-frequency N       save an imatrix copy every N iterations (default: 0)
-         --process-output         collect data for the output tensor (default: false)
-         --no-ppl                 do not compute perplexity (default: true)
-         --chunk N                start processing the input from chunk N (default: 0)
-
-bench:
-
-  -pps                            is the prompt shared across parallel sequences (default: false)
-  -npp n0,n1,...                  number of prompt tokens
-  -ntg n0,n1,...                  number of text generation tokens
-  -npl n0,n1,...                  number of parallel prompts
-
-embedding:
-
-         --embd-normalize         normalisation for embendings (default: 2) (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
-         --embd-output-format     empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
-         --embd-separator         separator of embendings (default \n) for example "<#sep#>"
-
 server:
 
          --host HOST              ip address to listen (default: 127.0.0.1)
          --port PORT              port to listen (default: 8080)
          --path PATH              path to serve static files from (default: )
-         --embedding(s)           enable embedding endpoint (default: disabled)
+         --embedding(s)           restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)
          --api-key KEY            API key to use for authentication (default: none)
          --api-key-file FNAME     path to file containing API keys (default: none)
          --ssl-key-file FNAME     path to file a PEM-encoded SSL private key
@@ -267,7 +232,8 @@ server:
                                   https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
   -sps,  --slot-prompt-similarity SIMILARITY
                                   how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled)
-
+         --lora-init-without-apply
+                                  load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled)
 
 logging:
 
@@ -279,15 +245,6 @@ logging:
          --log-file FNAME         Specify a log filename (without extension)
          --log-new                Create a separate new log file on start. Each log file will have unique name: "<name>.<ID>.log"
          --log-append             Don't truncate the old log file.
-
-cvector:
-
-  -o,    --output FNAME           output file (default: 'control_vector.gguf')
-         --positive-file FNAME    positive prompts file, one prompt per line (default: 'examples/cvector-generator/positive.txt')
-         --negative-file FNAME    negative prompts file, one prompt per line (default: 'examples/cvector-generator/negative.txt')
-         --pca-batch N            batch size used for PCA. Larger batch runs faster, but uses more memory (default: 100)
-         --pca-iter N             number of iterations used for PCA (default: 1000)
-         --method {pca,mean}      dimensionality reduction method to be used (default: pca)
 ```
 
 
@@ -411,7 +368,8 @@ node index.js
 
 ## API Endpoints
 
-- **GET** `/health`: Returns the current state of the server:
+### GET `/health`: Returns the current state of the server
+
   - 503 -> `{"status": "loading model"}` if the model is still being loaded.
   - 500 -> `{"status": "error"}` if the model failed to load.
   - 200 -> `{"status": "ok", "slots_idle": 1, "slots_processing": 2 }` if the model is successfully loaded and the server is ready for further requests mentioned below.
@@ -420,7 +378,7 @@ node index.js
 
   If the query parameter `include_slots` is passed, `slots` field will contain internal slots data except if `--slots-endpoint-disable` is set.
 
-- **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
+### POST `/completion`: Given a `prompt`, it returns the predicted completion.
 
     *Options:*
 
@@ -498,7 +456,7 @@ node index.js
 
     `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["top_k", "tfs_z", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values.
 
-### Result JSON
+**Response format**
 
 - Note: When using streaming mode (`stream`), only `content` and `stop` will be returned until end of completion.
 
@@ -537,7 +495,7 @@ Notice that each `probs` is an array of length `n_probs`.
 - `tokens_evaluated`: Number of tokens evaluated in total from the prompt
 - `truncated`: Boolean indicating if the context size was exceeded during generation, i.e. the number of tokens provided in the prompt (`tokens_evaluated`) plus tokens generated (`tokens predicted`) exceeded the context size (`n_ctx`)
 
-- **POST** `/tokenize`: Tokenize a given text.
+### POST `/tokenize`: Tokenize a given text
 
     *Options:*
 
@@ -545,13 +503,15 @@ Notice that each `probs` is an array of length `n_probs`.
 
     `add_special`: Boolean indicating if special tokens, i.e. `BOS`, should be inserted.  Default: `false`
 
-- **POST** `/detokenize`: Convert tokens to text.
+### POST `/detokenize`: Convert tokens to text
 
     *Options:*
 
     `tokens`: Set the tokens to detokenize.
 
-- **POST** `/embedding`: Generate embedding of a given text just as [the embedding example](../embedding) does.
+### POST `/embedding`: Generate embedding of a given text
+
+The same as [the embedding example](../embedding) does.
 
     *Options:*
 
@@ -559,7 +519,9 @@ Notice that each `probs` is an array of length `n_probs`.
 
     `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `content`. You can determine the place of the image in the content as in the following: `Image: [img-21].\nCaption: This is a picture of a house`. In this case, `[img-21]` will be replaced by the embeddings of the image with id `21` in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 21}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
 
-- **POST** `/infill`: For code infilling. Takes a prefix and a suffix and returns the predicted completion as stream.
+### POST `/infill`: For code infilling.
+
+Takes a prefix and a suffix and returns the predicted completion as stream.
 
     *Options:*
 
@@ -571,7 +533,7 @@ Notice that each `probs` is an array of length `n_probs`.
 
 - **GET** `/props`: Return current server settings.
 
-### Result JSON
+**Response format**
 
 ```json
 {
@@ -589,7 +551,9 @@ Notice that each `probs` is an array of length `n_probs`.
 - `total_slots` - the total number of slots for process requests (defined by `--parallel` option)
 - `chat_template` - the model's original Jinja2 prompt template
 
-- **POST** `/v1/chat/completions`: OpenAI-compatible Chat Completions API. Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
+### POST `/v1/chat/completions`: OpenAI-compatible Chat Completions API
+
+Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
 
     *Options:*
 
@@ -641,7 +605,7 @@ Notice that each `probs` is an array of length `n_probs`.
     }'
     ```
 
-- **POST** `/v1/embeddings`: OpenAI-compatible embeddings API.
+### POST `/v1/embeddings`: OpenAI-compatible embeddings API
 
     *Options:*
 
@@ -675,9 +639,9 @@ Notice that each `probs` is an array of length `n_probs`.
     }'
     ```
 
-- **GET** `/slots`: Returns the current slots processing state. Can be disabled with `--slots-endpoint-disable`.
+### GET `/slots`: Returns the current slots processing state. Can be disabled with `--slots-endpoint-disable`.
 
-### Result JSON
+**Response format**
 
 ```json
 [
@@ -738,7 +702,7 @@ Notice that each `probs` is an array of length `n_probs`.
 ]
 ```
 
-- **GET** `/metrics`: [Prometheus](https://prometheus.io/) compatible metrics exporter endpoint if `--metrics` is enabled:
+### GET `/metrics`: Prometheus compatible metrics exporter endpoint if `--metrics` is enabled:
 
 Available metrics:
 - `llamacpp:prompt_tokens_total`: Number of prompt tokens processed.
@@ -750,13 +714,13 @@ Available metrics:
 - `llamacpp:requests_processing`: Number of requests processing.
 - `llamacpp:requests_deferred`: Number of requests deferred.
 
-- **POST** `/slots/{id_slot}?action=save`: Save the prompt cache of the specified slot to a file.
+### POST `/slots/{id_slot}?action=save`: Save the prompt cache of the specified slot to a file.
 
     *Options:*
 
     `filename`: Name of the file to save the slot's prompt cache. The file will be saved in the directory specified by the `--slot-save-path` server parameter.
 
-### Result JSON
+**Response format**
 
 ```json
 {
@@ -770,13 +734,13 @@ Available metrics:
 }
 ```
 
-- **POST** `/slots/{id_slot}?action=restore`: Restore the prompt cache of the specified slot from a file.
+### POST `/slots/{id_slot}?action=restore`: Restore the prompt cache of the specified slot from a file.
 
     *Options:*
 
     `filename`: Name of the file to restore the slot's prompt cache from. The file should be located in the directory specified by the `--slot-save-path` server parameter.
 
-### Result JSON
+**Response format**
 
 ```json
 {
@@ -790,9 +754,9 @@ Available metrics:
 }
 ```
 
-- **POST** `/slots/{id_slot}?action=erase`: Erase the prompt cache of the specified slot.
+### POST `/slots/{id_slot}?action=erase`: Erase the prompt cache of the specified slot.
 
-### Result JSON
+**Response format**
 
 ```json
 {
@@ -801,6 +765,42 @@ Available metrics:
 }
 ```
 
+### GET `/lora-adapters`: Get list of all LoRA adapters
+
+If an adapter is disabled, the scale will be set to 0.
+
+**Response format**
+
+```json
+[
+    {
+        "id": 0,
+        "path": "my_adapter_1.gguf",
+        "scale": 0.0
+    },
+    {
+        "id": 1,
+        "path": "my_adapter_2.gguf",
+        "scale": 0.0
+    }
+]
+```
+
+### POST `/lora-adapters`: Set list of LoRA adapters
+
+To disable an adapter, either remove it from the list below, or set scale to 0.
+
+**Request format**
+
+To know the `id` of the adapter, use GET `/lora-adapters`
+
+```json
+[
+  {"id": 0, "scale": 0.2},
+  {"id": 1, "scale": 0.8}
+]
+```
+
 ## More examples
 
 ### Change system prompt on runtime
diff --git a/examples/server/server.cpp b/examples/server/server.cpp
index 7813a295..360f571e 100644
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@@ -78,6 +78,7 @@ enum server_task_type {
     SERVER_TASK_TYPE_SLOT_SAVE,
     SERVER_TASK_TYPE_SLOT_RESTORE,
     SERVER_TASK_TYPE_SLOT_ERASE,
+    SERVER_TASK_TYPE_SET_LORA,
 };
 
 struct server_task {
@@ -622,6 +623,7 @@ struct server_response {
 struct server_context {
     llama_model * model = nullptr;
     llama_context * ctx = nullptr;
+    std::vector<llama_lora_adapter_container> lora_adapters;
 
     gpt_params params;
 
@@ -677,7 +679,11 @@ struct server_context {
         // dedicate one sequence to the system prompt
         params.n_parallel += 1;
 
-        std::tie(model, ctx) = llama_init_from_gpt_params(params);
+        llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+        model = llama_init.model;
+        ctx = llama_init.context;
+        lora_adapters = llama_init.lora_adapters;
         params.n_parallel -= 1; // but be sneaky about it
         if (model == nullptr) {
             LOG_ERROR("unable to load model", {{"model", params.model}});
@@ -900,7 +906,7 @@ struct server_context {
 
         slot.params.stream             = json_value(data, "stream",            false);
         slot.params.cache_prompt       = json_value(data, "cache_prompt",      false);
-        slot.params.n_predict          = json_value(data, "n_predict",         default_params.n_predict);
+        slot.params.n_predict          = json_value(data, "n_predict",         json_value(data, "max_tokens", default_params.n_predict));
         slot.sparams.top_k             = json_value(data, "top_k",             default_sparams.top_k);
         slot.sparams.top_p             = json_value(data, "top_p",             default_sparams.top_p);
         slot.sparams.min_p             = json_value(data, "min_p",             default_sparams.min_p);
@@ -969,6 +975,8 @@ struct server_context {
                 (prompt->is_array() &&  prompt->size() == 1 && prompt->at(0).is_string()) ||
                 (prompt->is_array() && !prompt->empty()     && prompt->at(0).is_number_integer())) {
                 slot.prompt = *prompt;
+            } else if (prompt->is_array() && prompt->size() == 1 && prompt->at(0).is_array()) {
+                slot.prompt = prompt->at(0);
             } else {
                 send_error(task, "\"prompt\" must be a string or an array of integers", ERROR_TYPE_INVALID_REQUEST);
                 return false;
@@ -1847,6 +1855,14 @@ struct server_context {
                     };
                     queue_results.send(result);
                 } break;
+            case SERVER_TASK_TYPE_SET_LORA:
+                {
+                    llama_lora_adapters_apply(ctx, lora_adapters);
+                    server_task_result result;
+                    result.id = task.id;
+                    result.data = json{{ "success", true }};
+                    queue_results.send(result);
+                } break;
         }
     }
 
@@ -3325,6 +3341,55 @@ int main(int argc, char ** argv) {
         return res.set_content(root.dump(), "application/json; charset=utf-8");
     };
 
+    const auto handle_lora_adapters_list = [&](const httplib::Request & req, httplib::Response & res) {
+        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
+        json result = json::array();
+        for (size_t i = 0; i < ctx_server.lora_adapters.size(); ++i) {
+            auto & la = ctx_server.lora_adapters[i];
+            result.push_back({
+                {"id", i},
+                {"path", la.path},
+                {"scale", la.scale},
+            });
+        }
+        res.set_content(result.dump(), "application/json");
+        res.status = 200; // HTTP OK
+    };
+
+    const auto handle_lora_adapters_apply = [&](const httplib::Request & req, httplib::Response & res) {
+        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
+
+        const std::vector<json> body = json::parse(req.body);
+        int max_idx = ctx_server.lora_adapters.size();
+
+        // clear existing value
+        for (auto & la : ctx_server.lora_adapters) {
+            la.scale = 0.0f;
+        }
+
+        // set value
+        for (auto entry : body) {
+            int id      = entry.at("id");
+            float scale = entry.at("scale");
+            if (0 <= id && id < max_idx) {
+                ctx_server.lora_adapters[id].scale = scale;
+            } else {
+                throw std::runtime_error("invalid adapter id");
+            }
+        }
+
+        server_task task;
+        task.type = SERVER_TASK_TYPE_SET_LORA;
+        const int id_task = ctx_server.queue_tasks.post(task);
+        ctx_server.queue_results.add_waiting_task_id(id_task);
+
+        server_task_result result = ctx_server.queue_results.recv(id_task);
+        ctx_server.queue_results.remove_waiting_task_id(id_task);
+
+        res.set_content(result.data.dump(), "application/json");
+        res.status = 200; // HTTP OK
+    };
+
     auto handle_static_file = [](unsigned char * content, size_t len, const char * mime_type) {
         return [content, len, mime_type](const httplib::Request &, httplib::Response & res) {
             res.set_content(reinterpret_cast<const char*>(content), len, mime_type);
@@ -3363,7 +3428,6 @@ int main(int argc, char ** argv) {
 
     // register API routes
     svr->Get ("/health",              handle_health);
-    svr->Get ("/slots",               handle_slots);
     svr->Get ("/metrics",             handle_metrics);
     svr->Get ("/props",               handle_props);
     svr->Get ("/v1/models",           handle_models);
@@ -3378,6 +3442,11 @@ int main(int argc, char ** argv) {
     svr->Post("/v1/embeddings",       handle_embeddings);
     svr->Post("/tokenize",            handle_tokenize);
     svr->Post("/detokenize",          handle_detokenize);
+    // LoRA adapters hotswap
+    svr->Get ("/lora-adapters",       handle_lora_adapters_list);
+    svr->Post("/lora-adapters",       handle_lora_adapters_apply);
+    // Save & load slots
+    svr->Get ("/slots",               handle_slots);
     if (!params.slot_save_path.empty()) {
         // only enable slot endpoints if slot_save_path is set
         svr->Post("/slots/:id_slot",  handle_slots_action);
diff --git a/examples/server/tests/features/lora.feature b/examples/server/tests/features/lora.feature
new file mode 100644
index 00000000..7b85988a
--- /dev/null
+++ b/examples/server/tests/features/lora.feature
@@ -0,0 +1,36 @@
+@llama.cpp
+@lora
+Feature: llama.cpp server
+
+  Background: Server startup
+    Given a server listening on localhost:8080
+    And   a model url https://huggingface.co/ggml-org/stories15M_MOE/resolve/main/stories15M_MOE-F16.gguf
+    And   a model file stories15M_MOE-F16.gguf
+    And   a model alias stories15M_MOE
+    And   a lora adapter file from https://huggingface.co/ggml-org/stories15M_MOE/resolve/main/moe_shakespeare15M.gguf
+    And   42 as server seed
+    And   1024 as batch size
+    And   1024 as ubatch size
+    And   2048 KV cache size
+    And   64 max tokens to predict
+    And   0.0 temperature
+    Then  the server is starting
+    Then  the server is healthy
+
+  Scenario: Completion LoRA disabled
+    Given switch off lora adapter 0
+    Given a prompt:
+    """
+    Look in thy glass
+    """
+    And   a completion request with no api error
+    Then  64 tokens are predicted matching little|girl|three|years|old
+
+  Scenario: Completion LoRA enabled
+    Given switch on lora adapter 0
+    Given a prompt:
+    """
+    Look in thy glass
+    """
+    And   a completion request with no api error
+    Then  64 tokens are predicted matching eye|love|glass|sun
diff --git a/examples/server/tests/features/steps/steps.py b/examples/server/tests/features/steps/steps.py
index df0814cc..6705a34f 100644
--- a/examples/server/tests/features/steps/steps.py
+++ b/examples/server/tests/features/steps/steps.py
@@ -7,6 +7,7 @@ import subprocess
 import sys
 import threading
 import time
+import requests
 from collections.abc import Sequence
 from contextlib import closing
 from re import RegexFlag
@@ -70,6 +71,7 @@ def step_server_config(context, server_fqdn: str, server_port: str):
     context.user_api_key = None
     context.response_format = None
     context.temperature = None
+    context.lora_file = None
 
     context.tasks_result = []
     context.concurrent_tasks = []
@@ -82,6 +84,12 @@ def step_download_hf_model(context, hf_file: str, hf_repo: str):
     context.model_hf_file = hf_file
     context.model_file = os.path.basename(hf_file)
 
+@step('a lora adapter file from {lora_file_url}')
+def step_download_lora_file(context, lora_file_url: str):
+    file_name = lora_file_url.split('/').pop()
+    context.lora_file = f'../../../{file_name}'
+    with open(context.lora_file, 'wb') as f:
+        f.write(requests.get(lora_file_url).content)
 
 @step('a model file {model_file}')
 def step_model_file(context, model_file: str):
@@ -849,6 +857,17 @@ async def step_erase_slot(context, slot_id):
             context.response = response
 
 
+@step('switch {on_or_off} lora adapter {lora_id:d}')
+@async_run_until_complete
+async def toggle_lora_adapter(context, on_or_off: str, lora_id: int):
+    async with aiohttp.ClientSession() as session:
+        async with session.post(f'{context.base_url}/lora-adapters',
+                                json=[{'id': lora_id, 'scale': 1 if on_or_off == 'on' else 0}],
+                                headers={"Content-Type": "application/json"}) as response:
+            context.response = response
+            print([{'id': lora_id, 'scale': 1 if on_or_off == 'on' else 0}])
+
+
 @step('the server responds with status code {status_code:d}')
 def step_server_responds_with_status_code(context, status_code):
     assert context.response.status == status_code
@@ -1326,6 +1345,8 @@ def start_server_background(context):
         server_args.extend(['--grp-attn-w', context.n_ga_w])
     if context.debug:
         server_args.append('--verbose')
+    if context.lora_file:
+        server_args.extend(['--lora', context.lora_file])
     if 'SERVER_LOG_FORMAT_JSON' not in os.environ:
         server_args.extend(['--log-format', "text"])
 
diff --git a/examples/server/tests/requirements.txt b/examples/server/tests/requirements.txt
index 2c741ea1..f2d7e5c5 100644
--- a/examples/server/tests/requirements.txt
+++ b/examples/server/tests/requirements.txt
@@ -4,3 +4,4 @@ huggingface_hub~=0.20.3
 numpy~=1.26.4
 openai~=1.30.3
 prometheus-client~=0.20.0
+requests~=2.32.3
diff --git a/examples/server/utils.hpp b/examples/server/utils.hpp
index db6b3b74..e6a1f069 100644
--- a/examples/server/utils.hpp
+++ b/examples/server/utils.hpp
@@ -355,24 +355,6 @@ static json oaicompat_completion_params_parse(
 
     llama_params["__oaicompat"] = true;
 
-    // Map OpenAI parameters to llama.cpp parameters
-    //
-    // For parameters that are defined by the OpenAI documentation (e.g.
-    // temperature), we explicitly specify OpenAI's intended default; we
-    // need to do that because sometimes OpenAI disagrees with llama.cpp
-    //
-    // https://platform.openai.com/docs/api-reference/chat/create
-    llama_sampling_params default_sparams;
-    llama_params["model"]             = json_value(body,   "model",             std::string("unknown"));
-    llama_params["frequency_penalty"] = json_value(body,   "frequency_penalty", 0.0);
-    llama_params["logit_bias"]        = json_value(body,   "logit_bias",        json::object());
-    llama_params["n_predict"]         = json_value(body,   "max_tokens",        -1);
-    llama_params["presence_penalty"]  = json_value(body,   "presence_penalty",  0.0);
-    llama_params["seed"]              = json_value(body,   "seed",              LLAMA_DEFAULT_SEED);
-    llama_params["stream"]            = json_value(body,   "stream",            false);
-    llama_params["temperature"]       = json_value(body,   "temperature",       1.0);
-    llama_params["top_p"]             = json_value(body,   "top_p",             1.0);
-
     // Apply chat template to the list of messages
     llama_params["prompt"] = format_chat(model, chat_template, body.at("messages"));
 
diff --git a/examples/simple/README.md b/examples/simple/README.md
index 49e24501..0ff34253 100644
--- a/examples/simple/README.md
+++ b/examples/simple/README.md
@@ -3,7 +3,7 @@
 The purpose of this example is to demonstrate a minimal usage of llama.cpp for generating text with a given prompt.
 
 ```bash
-./simple -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is"
+./llama-simple -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is"
 
 ...
 
diff --git a/examples/speculative/speculative.cpp b/examples/speculative/speculative.cpp
index 0939a1a6..b051a18f 100644
--- a/examples/speculative/speculative.cpp
+++ b/examples/speculative/speculative.cpp
@@ -66,7 +66,9 @@ int main(int argc, char ** argv) {
     llama_context * ctx_dft = NULL;
 
     // load the target model
-    std::tie(model_tgt, ctx_tgt) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init_tgt = llama_init_from_gpt_params(params);
+    model_tgt = llama_init_tgt.model;
+    ctx_tgt = llama_init_tgt.context;
 
     // load the draft model
     params.model = params.model_draft;
@@ -75,7 +77,9 @@ int main(int argc, char ** argv) {
         params.n_threads = params.n_threads_draft;
     }
     params.n_threads_batch = params.n_threads_batch_draft;
-    std::tie(model_dft, ctx_dft) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init_dft = llama_init_from_gpt_params(params);
+    model_dft = llama_init_dft.model;
+    ctx_dft = llama_init_dft.context;
 
     const bool vocab_type_tgt = llama_vocab_type(model_tgt);
     LOG("vocab_type tgt: %d\n", vocab_type_tgt);
diff --git a/examples/sycl/README.md b/examples/sycl/README.md
index 0e3acd35..8819d87f 100644
--- a/examples/sycl/README.md
+++ b/examples/sycl/README.md
@@ -12,9 +12,9 @@ This example program provides the tools for llama.cpp for SYCL on Intel GPU.
 
 List all SYCL devices with ID, compute capability, max work group size, ect.
 
-1. Build the llama.cpp for SYCL for all targets.
+1. Build the llama.cpp for SYCL for the specified target *(using GGML_SYCL_TARGET)*.
 
-2. Enable oneAPI running environment
+2. Enable oneAPI running environment *(if GGML_SYCL_TARGET is set to INTEL -default-)*
 
 ```
 source /opt/intel/oneapi/setvars.sh
@@ -29,19 +29,13 @@ source /opt/intel/oneapi/setvars.sh
 Check the ID in startup log, like:
 
 ```
-found 4 SYCL devices:
-  Device 0: Intel(R) Arc(TM) A770 Graphics,	compute capability 1.3,
-    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
-  Device 1: Intel(R) FPGA Emulation Device,	compute capability 1.2,
-    max compute_units 24,	max work group size 67108864,	max sub group size 64,	global mem size 67065057280
-  Device 2: 13th Gen Intel(R) Core(TM) i7-13700K,	compute capability 3.0,
-    max compute_units 24,	max work group size 8192,	max sub group size 64,	global mem size 67065057280
-  Device 3: Intel(R) Arc(TM) A770 Graphics,	compute capability 3.0,
-    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
+found 2 SYCL devices:
+|  |                   |                                       |       |Max    |        |Max  |Global |                     |
+|  |                   |                                       |       |compute|Max work|sub  |mem    |                     |
+|ID|        Device Type|                                   Name|Version|units  |group   |group|size   |       Driver version|
+|--|-------------------|---------------------------------------|-------|-------|--------|-----|-------|---------------------|
+| 0| [level_zero:gpu:0]|                Intel Arc A770 Graphics|    1.3|    512|    1024|   32| 16225M|            1.3.29138|
+| 1| [level_zero:gpu:1]|                 Intel UHD Graphics 750|    1.3|     32|     512|   32| 62631M|            1.3.29138|
 
 ```
 
-|Attribute|Note|
-|-|-|
-|compute capability 1.3|Level-zero running time, recommended |
-|compute capability 3.0|OpenCL running time, slower than level-zero in most cases|
diff --git a/examples/sycl/win-run-llama2.bat b/examples/sycl/win-run-llama2.bat
index f0385cdf..c2918d6d 100644
--- a/examples/sycl/win-run-llama2.bat
+++ b/examples/sycl/win-run-llama2.bat
@@ -6,4 +6,4 @@ set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
 @call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
 
 
-.\build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0
+.\build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0
diff --git a/examples/tokenize/tokenize.cpp b/examples/tokenize/tokenize.cpp
index 2afb6024..17f5e496 100644
--- a/examples/tokenize/tokenize.cpp
+++ b/examples/tokenize/tokenize.cpp
@@ -163,7 +163,7 @@ static void write_utf8_cstr_to_stdout(const char * str, bool & invalid_utf8) {
                 printf(">");
                 return;
             }
-            GGML_ASSERT(false && "MultiByteToWideChar() failed in an unexpected way.");
+            GGML_ABORT("MultiByteToWideChar() failed in an unexpected way.");
         }
 
         LPWSTR wstr = (LPWSTR) calloc(length_needed+1, sizeof(*wstr));
diff --git a/flake.lock b/flake.lock
index 940cda6a..f9e1548a 100644
--- a/flake.lock
+++ b/flake.lock
@@ -5,11 +5,11 @@
         "nixpkgs-lib": "nixpkgs-lib"
       },
       "locked": {
-        "lastModified": 1719994518,
-        "narHash": "sha256-pQMhCCHyQGRzdfAkdJ4cIWiw+JNuWsTX7f0ZYSyz0VY=",
+        "lastModified": 1722555600,
+        "narHash": "sha256-XOQkdLafnb/p9ij77byFQjDf5m5QYl9b2REiVClC+x4=",
         "owner": "hercules-ci",
         "repo": "flake-parts",
-        "rev": "9227223f6d922fee3c7b190b2cc238a99527bbb7",
+        "rev": "8471fe90ad337a8074e957b69ca4d0089218391d",
         "type": "github"
       },
       "original": {
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1721379653,
-        "narHash": "sha256-8MUgifkJ7lkZs3u99UDZMB4kbOxvMEXQZ31FO3SopZ0=",
+        "lastModified": 1723175592,
+        "narHash": "sha256-M0xJ3FbDUc4fRZ84dPGx5VvgFsOzds77KiBMW/mMTnI=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "1d9c2c9b3e71b9ee663d11c5d298727dace8d374",
+        "rev": "5e0ca22929f3342b19569b21b2f3462f053e497b",
         "type": "github"
       },
       "original": {
@@ -36,14 +36,14 @@
     },
     "nixpkgs-lib": {
       "locked": {
-        "lastModified": 1719876945,
-        "narHash": "sha256-Fm2rDDs86sHy0/1jxTOKB1118Q0O3Uc7EC0iXvXKpbI=",
+        "lastModified": 1722555339,
+        "narHash": "sha256-uFf2QeW7eAHlYXuDktm9c25OxOyCoUOQmh5SZ9amE5Q=",
         "type": "tarball",
-        "url": "https://github.com/NixOS/nixpkgs/archive/5daf0514482af3f97abaefc78a6606365c9108e2.tar.gz"
+        "url": "https://github.com/NixOS/nixpkgs/archive/a5d394176e64ab29c852d03346c1fc9b0b7d33eb.tar.gz"
       },
       "original": {
         "type": "tarball",
-        "url": "https://github.com/NixOS/nixpkgs/archive/5daf0514482af3f97abaefc78a6606365c9108e2.tar.gz"
+        "url": "https://github.com/NixOS/nixpkgs/archive/a5d394176e64ab29c852d03346c1fc9b0b7d33eb.tar.gz"
       }
     },
     "root": {
diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
index 66505753..1bb3d1ee 100644
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -50,9 +50,15 @@ else()
     set(GGML_BLAS_VENDOR_DEFAULT "Generic")
 endif()
 
+if (CMAKE_CROSSCOMPILING)
+    set(GGML_NATIVE_DEFAULT OFF)
+else()
+    set(GGML_NATIVE_DEFAULT ON)
+endif()
+
 # general
 option(GGML_STATIC "ggml: static link libraries"         OFF)
-option(GGML_NATIVE "ggml: enable -march=native flag"     ON)
+option(GGML_NATIVE "ggml: enable -march=native flag"     ${GGML_NATIVE_DEFAULT})
 option(GGML_LTO    "ggml: enable link time optimization" OFF)
 option(GGML_CCACHE "ggml: use ccache if available"       ON)
 
@@ -70,7 +76,7 @@ option(GGML_SANITIZE_ADDRESS   "ggml: enable address sanitizer"   OFF)
 option(GGML_SANITIZE_UNDEFINED "ggml: enable undefined sanitizer" OFF)
 
 # instruction set specific
-if (GGML_NATIVE)
+if (GGML_NATIVE OR NOT GGML_NATIVE_DEFAULT)
     set(INS_ENB OFF)
 else()
     set(INS_ENB ON)
@@ -108,6 +114,7 @@ option(GGML_LLAMAFILE                       "ggml: use LLAMAFILE"
 option(GGML_IQK_MUL_MAT                     "ggml: use optimized iqk matrix multiplications"  ON)
 
 option(GGML_CUDA                            "ggml: use CUDA"                                  OFF)
+option(GGML_MUSA                            "ggml: use MUSA"                                  OFF)
 option(GGML_CUDA_FORCE_DMMV                 "ggml: use dmmv instead of mmvq CUDA kernels"     OFF)
 option(GGML_CUDA_FORCE_MMQ                  "ggml: use mmq kernels instead of cuBLAS"         OFF)
 option(GGML_CUDA_FORCE_CUBLAS               "ggml: always use cuBLAS instead of mmq kernels"  OFF)
@@ -197,6 +204,7 @@ set(GGML_PUBLIC_HEADERS
     include/ggml-alloc.h
     include/ggml-backend.h
     include/ggml-blas.h
+    include/ggml-cann.h
     include/ggml-cuda.h
     include/ggml.h
     include/ggml-kompute.h
diff --git a/ggml/include/ggml-cuda.h b/ggml/include/ggml-cuda.h
index d7903c66..71bb6dcf 100644
--- a/ggml/include/ggml-cuda.h
+++ b/ggml/include/ggml-cuda.h
@@ -6,6 +6,9 @@
 #ifdef GGML_USE_HIPBLAS
 #define GGML_CUDA_NAME "ROCm"
 #define GGML_CUBLAS_NAME "hipBLAS"
+#elif defined(GGML_USE_MUSA)
+#define GGML_CUDA_NAME "MUSA"
+#define GGML_CUBLAS_NAME "muBLAS"
 #else
 #define GGML_CUDA_NAME "CUDA"
 #define GGML_CUBLAS_NAME "cuBLAS"
diff --git a/ggml/include/ggml-metal.h b/ggml/include/ggml-metal.h
index 6c3226c3..d483cf1a 100644
--- a/ggml/include/ggml-metal.h
+++ b/ggml/include/ggml-metal.h
@@ -50,6 +50,8 @@ GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void
 
 GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
 
+GGML_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
+
 GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
 
 // helper to check if the device supports a specific family
diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
index a0bcc67f..b9b0284b 100644
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -254,18 +254,8 @@
 
 #define GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
 
-#define GGML_ASSERT(x) \
-    do { \
-        if (!(x)) { \
-            fflush(stdout); \
-            fprintf(stderr, "GGML_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
-            ggml_print_backtrace(); \
-            abort(); \
-        } \
-    } while (0)
-
 #ifndef NDEBUG
-#define GGML_UNREACHABLE() GGML_ASSERT(!"statement should not be reached")
+#define GGML_UNREACHABLE() do { fprintf(stderr, "statement should be unreachable\n"); abort(); } while(0)
 #elif defined(__GNUC__)
 #define GGML_UNREACHABLE() __builtin_unreachable()
 #elif defined(_MSC_VER)
@@ -274,6 +264,17 @@
 #define GGML_UNREACHABLE() ((void) 0)
 #endif
 
+#ifdef __cplusplus
+#define GGML_NORETURN [[noreturn]]
+#elif defined(_MSC_VER)
+#define GGML_NORETURN __declspec(noreturn)
+#else
+#define GGML_NORETURN _Noreturn
+#endif
+
+#define GGML_ABORT(...) ggml_abort(__FILE__, __LINE__, __VA_ARGS__)
+#define GGML_ASSERT(x) if (!(x)) GGML_ABORT("GGML_ASSERT(%s) failed", #x)
+
 // used to copy the number of elements and stride in bytes of tensors into local variables.
 // main purpose is to reduce code duplication and improve readability.
 //
@@ -322,6 +323,9 @@
 extern "C" {
 #endif
 
+    GGML_NORETURN GGML_ATTRIBUTE_FORMAT(3, 4)
+    GGML_API void ggml_abort(const char * file, int line, const char * fmt, ...);
+
     enum ggml_status {
         GGML_STATUS_ALLOC_FAILED = -2,
         GGML_STATUS_FAILED = -1,
@@ -345,6 +349,7 @@ extern "C" {
     GGML_API ggml_bf16_t ggml_fp32_to_bf16(float);
     GGML_API float       ggml_bf16_to_fp32(ggml_bf16_t);  // consider just doing << 16
     GGML_API void        ggml_bf16_to_fp32_row(const ggml_bf16_t *, float *, int64_t);
+    GGML_API void        ggml_fp32_to_bf16_row_ref(const float *, ggml_bf16_t *, int64_t);
     GGML_API void        ggml_fp32_to_bf16_row(const float *, ggml_bf16_t *, int64_t);
 
     struct ggml_object;
@@ -653,8 +658,11 @@ extern "C" {
         GGML_CGRAPH_EVAL_ORDER_COUNT
     };
 
+    typedef uint32_t ggml_bitset_t;
+
     struct ggml_hash_set {
         size_t size;
+        ggml_bitset_t * used;
         struct ggml_tensor ** keys;
     };
 
@@ -668,7 +676,7 @@ extern "C" {
         struct ggml_tensor ** grads;
         struct ggml_tensor ** leafs;
 
-        struct ggml_hash_set visited_hash_table;
+        struct ggml_hash_set visited_hash_set;
 
         enum ggml_cgraph_eval_order order;
     };
@@ -715,8 +723,6 @@ extern "C" {
     GGML_API int64_t ggml_cycles(void);
     GGML_API int64_t ggml_cycles_per_ms(void);
 
-    GGML_API void    ggml_print_backtrace(void);
-
     // accepts a UTF-8 path, even on Windows
     GGML_API FILE *  ggml_fopen(const char * fname, const char * mode);
 
@@ -1151,16 +1157,17 @@ extern "C" {
 
     // group normalize along ne0*ne1*n_groups
     // used in stable-diffusion
-    // TODO: eps is hardcoded to 1e-6 for now
     GGML_API struct ggml_tensor * ggml_group_norm(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_groups);
+            int                   n_groups,
+            float                 eps);
 
     GGML_API struct ggml_tensor * ggml_group_norm_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_groups);
+            int                   n_groups,
+            float                 eps);
 
     // a - x
     // b - dy
@@ -1467,7 +1474,6 @@ extern "C" {
     // if mode & 2 == 1, GPT-NeoX style
     //
     // b is an int32 vector with size a->ne[2], it contains the positions
-    // c is freq factors (e.g. phi3-128k), (optional)
     GGML_API struct ggml_tensor * ggml_rope(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1484,6 +1490,7 @@ extern "C" {
             int                   mode);
 
     // custom RoPE
+    // c is freq factors (e.g. phi3-128k), (optional)
     GGML_API struct ggml_tensor * ggml_rope_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -2022,8 +2029,8 @@ extern "C" {
 
     // ggml_graph_plan() has to be called before ggml_graph_compute()
     // when plan.work_size > 0, caller must allocate memory for plan.work_data
-    GGML_API struct ggml_cplan ggml_graph_plan            (const struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/);
-    GGML_API enum ggml_status  ggml_graph_compute         (      struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
+    GGML_API struct ggml_cplan ggml_graph_plan   (const struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/);
+    GGML_API enum ggml_status  ggml_graph_compute(      struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
     // same as ggml_graph_compute() but the work data is allocated as a part of the context
     // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
     GGML_API enum ggml_status  ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
index 0b1ad48a..be6c2cc6 100644
--- a/ggml/src/CMakeLists.txt
+++ b/ggml/src/CMakeLists.txt
@@ -139,6 +139,17 @@ if (GGML_METAL)
         )
 endif()
 
+if (GGML_MUSA)
+    set(CMAKE_C_COMPILER clang)
+    set(CMAKE_C_EXTENSIONS OFF)
+    set(CMAKE_CXX_COMPILER clang++)
+    set(CMAKE_CXX_EXTENSIONS OFF)
+
+    set(GGML_CUDA ON)
+
+    list(APPEND GGML_CDEF_PUBLIC GGML_USE_MUSA)
+endif()
+
 if (GGML_OPENMP)
     find_package(OpenMP)
     if (OpenMP_FOUND)
@@ -147,6 +158,11 @@ if (GGML_OPENMP)
         add_compile_definitions(GGML_USE_OPENMP)
 
         set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
+
+        if (GGML_MUSA)
+            set(GGML_EXTRA_INCLUDES ${GGML_EXTRA_INCLUDES} "/usr/lib/llvm-10/include/openmp")
+            set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} "/usr/lib/llvm-10/lib/libomp.so")
+        endif()
     else()
         message(WARNING "OpenMP not found")
     endif()
@@ -257,11 +273,16 @@ endif()
 if (GGML_CUDA)
     cmake_minimum_required(VERSION 3.18)  # for CMAKE_CUDA_ARCHITECTURES
 
-    find_package(CUDAToolkit)
-
-    set(CMAKE_CUDA_USE_RESPONSE_FILE_FOR_INCLUDES 0)
-    set(CMAKE_CUDA_USE_RESPONSE_FILE_FOR_LIBRARIES 0)
-    set(CMAKE_CUDA_USE_RESPONSE_FILE_FOR_OBJECTS 0)
+    if (GGML_MUSA)
+        list(APPEND CMAKE_MODULE_PATH "/usr/local/musa/cmake/")
+        find_package(MUSAToolkit)
+        set(CUDAToolkit_FOUND ${MUSAToolkit_FOUND})
+    else()
+        find_package(CUDAToolkit)
+        set(CMAKE_CUDA_USE_RESPONSE_FILE_FOR_INCLUDES 0)
+        set(CMAKE_CUDA_USE_RESPONSE_FILE_FOR_LIBRARIES 0)
+        set(CMAKE_CUDA_USE_RESPONSE_FILE_FOR_OBJECTS 0)
+    endif()
 
     if (CUDAToolkit_FOUND)
         message(STATUS "CUDA found")
@@ -280,7 +301,11 @@ if (GGML_CUDA)
         endif()
         message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
 
-        enable_language(CUDA)
+        if (GGML_MUSA)
+            set(CMAKE_CUDA_COMPILER ${MUSAToolkit_MCC_EXECUTABLE})
+        else()
+            enable_language(CUDA)
+        endif()
 
         file(GLOB   GGML_HEADERS_CUDA "ggml-cuda/*.cuh")
         list(APPEND GGML_HEADERS_CUDA "../include/ggml-cuda.h")
@@ -344,21 +369,40 @@ if (GGML_CUDA)
             add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
         endif()
 
+        if (GGML_MUSA)
+            set_source_files_properties(${GGML_SOURCES_CUDA} PROPERTIES LANGUAGE CXX)
+            foreach(SOURCE ${GGML_SOURCES_CUDA})
+                set_property(SOURCE ${SOURCE} PROPERTY COMPILE_FLAGS "-x musa -mtgpu --cuda-gpu-arch=mp_22")
+            endforeach()
+        endif()
+
         if (GGML_STATIC)
             if (WIN32)
                 # As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
                 set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
             else ()
-                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+                if (GGML_MUSA)
+                    set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} MUSA::musart_static MUSA::mublas_static)
+                else()
+                    set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+                endif()
             endif()
         else()
-            set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
+            if (GGML_MUSA)
+                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} MUSA::musart MUSA::mublas)
+            else()
+                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
+            endif()
         endif()
 
         if (GGML_CUDA_NO_VMM)
             # No VMM requested, no need to link directly with the cuda driver lib (libcuda.so)
         else()
-            set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cuda_driver) # required by cuDeviceGetAttribute(), cuMemGetAllocationGranularity(...), ...
+            if (GGML_MUSA)
+                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} MUSA::musa_driver) # required by muDeviceGetAttribute(), muMemGetAllocationGranularity(...), ...
+            else()
+                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cuda_driver) # required by cuDeviceGetAttribute(), cuMemGetAllocationGranularity(...), ...
+            endif()
         endif()
     else()
         message(WARNING "CUDA not found")
@@ -816,11 +860,6 @@ if (GGML_CANN)
                 ${CANN_INSTALL_DIR}/acllib/include
             )
 
-            # TODO: find libs
-            link_directories(
-                ${CANN_INSTALL_DIR}/lib64
-            )
-
             add_subdirectory(ggml-cann/kernels)
             list(APPEND CANN_LIBRARIES
                 ascendcl
@@ -839,6 +878,7 @@ if (GGML_CANN)
 
             set(GGML_EXTRA_LIBS     ${GGML_EXTRA_LIBS}     ${CANN_LIBRARIES} )
             set(GGML_EXTRA_INCLUDES ${GGML_EXTRA_INCLUDES} ${CANN_INCLUDE_DIRS})
+            set(GGML_EXTRA_LIBDIRS  ${GGML_EXTRA_LIBDIRS}  ${CANN_INSTALL_DIR}/lib64)
             list(APPEND GGML_CDEF_PUBLIC GGML_USE_CANN)
         endif()
     else()
@@ -869,8 +909,10 @@ function(get_flags CCID CCVER)
         set(C_FLAGS   -Wdouble-promotion)
         set(CXX_FLAGS -Wno-array-bounds)
 
-        if (CCVER VERSION_GREATER_EQUAL 7.1.0)
-            list(APPEND CXX_FLAGS -Wno-format-truncation)
+        if (NOT GGML_MUSA)
+            if (CCVER VERSION_GREATER_EQUAL 7.1.0)
+                list(APPEND CXX_FLAGS -Wno-format-truncation)
+            endif()
         endif()
         if (CCVER VERSION_GREATER_EQUAL 8.1.0)
             list(APPEND CXX_FLAGS -Wextra-semi)
@@ -1278,6 +1320,7 @@ endif()
 target_compile_definitions(ggml PUBLIC  ${GGML_CDEF_PUBLIC})
 target_include_directories(ggml PUBLIC ../include)
 target_include_directories(ggml PRIVATE . ${GGML_EXTRA_INCLUDES})
+target_link_directories(ggml PRIVATE ${GGML_EXTRA_LIBDIRS})
 target_compile_features   (ggml PRIVATE c_std_11) # don't bump
 
 target_link_libraries(ggml PRIVATE Threads::Threads ${GGML_EXTRA_LIBS})
diff --git a/ggml/src/ggml-aarch64.c b/ggml/src/ggml-aarch64.c
index af53dea1..7adaadc9 100644
--- a/ggml/src/ggml-aarch64.c
+++ b/ggml/src/ggml-aarch64.c
@@ -16,6 +16,8 @@
 
 #if defined(__GNUC__)
 #pragma GCC diagnostic ignored "-Woverlength-strings"
+#elif defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
 #define UNUSED GGML_UNUSED
@@ -384,8 +386,8 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(blocklen);
 
 #if defined(__ARM_FEATURE_SVE)
-    if (svcntw() == 8) {
-        GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) &&
+    if (ggml_sve_cnt_b == QK8_0) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
                     "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
     }
 #endif
@@ -496,8 +498,8 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(blocklen);
 
 #if defined(__ARM_FEATURE_SVE)
-    if (svcntw() == 8) {
-        GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) &&
+    if (ggml_sve_cnt_b == QK8_0) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
                     "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
     }
 #endif
@@ -614,7 +616,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(blocklen);
 
 #if defined(__ARM_FEATURE_SVE) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
-    if (svcntw() == 8) {
+    if (ggml_sve_cnt_b == QK8_0) {
         const void * b_ptr = vx;
         const void * a_ptr = vy;
         float * res_ptr = s;
@@ -680,12 +682,12 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
         return;
     }
     else if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
-        GGML_ASSERT((ggml_cpu_has_sve() && (svcntw() == 8)) &&
+        GGML_ASSERT((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
                     "__ARM_FEATURE_SVE for vector size of 256-bits not defined, use the Q4_0_4_8 quantization format for optimal "
                     "performance");
     }
     else if (ggml_cpu_has_neon()) {
-        GGML_ASSERT(((ggml_cpu_has_sve() && (svcntw() == 8)) || ggml_cpu_has_matmul_int8()) &&
+        GGML_ASSERT(((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) || ggml_cpu_has_matmul_int8()) &&
                     "__ARM_FEATURE_SVE for vector size of 256-bits and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 "
                     "quantization format for optimal performance");
     }
@@ -745,8 +747,8 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(blocklen);
 
 #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
-    if (svcntw() == 8) {
-        GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) &&
+    if (ggml_sve_cnt_b == QK8_0) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
                     "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
     }
 #endif
@@ -1266,8 +1268,8 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(blocklen);
 
 #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
-    if (svcntw() == 8) {
-        GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) &&
+    if (ggml_sve_cnt_b == QK8_0) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
                     "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
     }
 #endif
@@ -1728,7 +1730,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(blocklen);
 
 #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
-    if (svcntw() == 8) {
+    if (ggml_sve_cnt_b == QK8_0) {
         const void * b_ptr = vx;
         const void * a_ptr = vy;
         float * res_ptr = s;
@@ -2139,12 +2141,12 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
         return;
     }
     else if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
-        GGML_ASSERT((ggml_cpu_has_sve() && (svcntw() == 8)) &&
+        GGML_ASSERT((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
                     "__ARM_FEATURE_SVE for vector size of 256-bits not defined, use the Q4_0_4_8 quantization format for optimal "
                     "performance");
     }
     else if (ggml_cpu_has_neon()) {
-        GGML_ASSERT(((ggml_cpu_has_sve() && (svcntw() == 8)) || ggml_cpu_has_matmul_int8()) &&
+        GGML_ASSERT(((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) || ggml_cpu_has_matmul_int8()) &&
                     "__ARM_FEATURE_SVE for vector size of 256-bits and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 "
                     "quantization format for optimal performance");
     }
diff --git a/ggml/src/ggml-alloc.c b/ggml/src/ggml-alloc.c
index e176b883..e485326a 100644
--- a/ggml/src/ggml-alloc.c
+++ b/ggml/src/ggml-alloc.c
@@ -91,8 +91,7 @@ void ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tenso
     if (talloc->offset + size > ggml_backend_buffer_get_size(talloc->buffer)) {
         fprintf(stderr, "%s: not enough space in the buffer to allocate %s (needed %zu, available %zu)\n",
                 __func__, tensor->name, size, ggml_backend_buffer_get_size(talloc->buffer) - talloc->offset);
-        GGML_ASSERT(!"not enough space in the buffer");
-        return;
+        GGML_ABORT("not enough space in the buffer");
     }
 
     void * addr = (char *)ggml_backend_buffer_get_base(talloc->buffer) + talloc->offset;
@@ -133,7 +132,7 @@ static void add_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offset,
             return;
         }
     }
-    GGML_ASSERT(!"out of allocated_tensors");
+    GGML_ABORT("out of allocated_tensors");
 }
 static void remove_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offset, const struct ggml_tensor * tensor) {
     for (int i = 0; i < 1024; i++) {
@@ -142,8 +141,7 @@ static void remove_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offs
             return;
         }
     }
-    fprintf(stderr, "tried to free tensor %s not found\n", tensor->name);
-    GGML_ASSERT(!"tensor not found");
+    GGML_ABORT("tried to free tensor %s not found\n", tensor->name);
 }
 #endif
 
@@ -176,8 +174,7 @@ static size_t ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * alloc, size_t siz
             // this should never happen
             fprintf(stderr, "%s: not enough space in the buffer to allocate %zu bytes, largest block available %zu bytes\n",
                     __func__, size, max_avail);
-            GGML_ASSERT(!"not enough space in the buffer");
-            GGML_UNREACHABLE();
+            GGML_ABORT("not enough space in the buffer");
         }
     }
 
@@ -443,7 +440,7 @@ void ggml_gallocr_free(ggml_gallocr_t galloc) {
         }
     }
 
-    free(galloc->hash_set.keys);
+    ggml_hash_set_free(&galloc->hash_set);
     free(galloc->hash_values);
     free(galloc->bufts);
     free(galloc->buffers);
@@ -456,7 +453,7 @@ void ggml_gallocr_free(ggml_gallocr_t galloc) {
 typedef struct ggml_gallocr * ggml_gallocr_t;
 
 static struct hash_node * ggml_gallocr_hash_get(ggml_gallocr_t galloc, struct ggml_tensor * t) {
-    size_t i = ggml_hash_find_or_insert(galloc->hash_set, t);
+    size_t i = ggml_hash_find_or_insert(&galloc->hash_set, t);
     return &galloc->hash_values[i];
 }
 
@@ -565,8 +562,8 @@ static int get_node_buffer_id(const int * node_buffer_ids, int i) {
 
 static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
     // clear hash tables
-    memset(galloc->hash_set.keys, 0, galloc->hash_set.size * sizeof(struct ggml_tensor *));
-    memset(galloc->hash_values,   0, galloc->hash_set.size * sizeof(struct hash_node));
+    ggml_hash_set_reset(&galloc->hash_set);
+    memset(galloc->hash_values, 0, sizeof(struct hash_node) * galloc->hash_set.size);
 
     // allocate leafs
     // these may be tensors that the application is not using in the graph, but may still want to allocate for other purposes
@@ -671,21 +668,19 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
 }
 
 bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
-    size_t hash_size = graph->visited_hash_table.size;
+    size_t min_hash_size = graph->n_nodes + graph->n_leafs;
+    // add 25% margin to avoid hash collisions
+    min_hash_size += min_hash_size / 4;
 
     // initialize hash table
-    if (galloc->hash_set.size < hash_size) {
-        free(galloc->hash_set.keys);
-        free(galloc->hash_values);
-        galloc->hash_set.size = hash_size;
-        galloc->hash_set.keys = calloc(hash_size, sizeof(struct ggml_tensor *));
-        galloc->hash_values   = calloc(hash_size, sizeof(struct hash_node));
+    if (galloc->hash_set.size < min_hash_size) {
+        ggml_hash_set_free(&galloc->hash_set);
+        galloc->hash_set = ggml_hash_set_new(min_hash_size);
         GGML_ASSERT(galloc->hash_set.keys != NULL);
+
+        free(galloc->hash_values);
+        galloc->hash_values = malloc(sizeof(struct hash_node) * galloc->hash_set.size);
         GGML_ASSERT(galloc->hash_values != NULL);
-    } else {
-        // reset hash table
-        memset(galloc->hash_set.keys, 0, sizeof(struct ggml_tensor *) * galloc->hash_set.size);
-        memset(galloc->hash_values,   0, sizeof(struct hash_node) * galloc->hash_set.size);
     }
 
     // reset allocators
@@ -817,8 +812,7 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
 }
 
 static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct tensor_alloc * talloc) {
-    ggml_backend_buffer_type_t buft = talloc->buffer_id != -1 ? galloc->bufts[talloc->buffer_id] : NULL;
-    size_t node_size = (node->data || node->view_src) ? 0 : ggml_backend_buft_get_alloc_size(buft, node);
+    size_t node_size = (node->data || node->view_src) ? 0 : ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
     return talloc->size_max >= node_size;
 }
 
diff --git a/ggml/src/ggml-backend.c b/ggml/src/ggml-backend.c
index d39cfed8..e1651cc6 100644
--- a/ggml/src/ggml-backend.c
+++ b/ggml/src/ggml-backend.c
@@ -351,15 +351,10 @@ void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t b
     }
 
     // an async copy would normally happen after all the queued operations on both backends are completed
-    // sync src, set_async dst
-    if (ggml_backend_buffer_is_host(src->buffer)) {
-        ggml_backend_synchronize(backend_src);
-        ggml_backend_tensor_set_async(backend_dst, dst, src->data, 0, ggml_nbytes(src));
-    } else {
-        ggml_backend_synchronize(backend_src);
-        ggml_backend_tensor_copy(src, dst);
-        ggml_backend_synchronize(backend_dst);
-    }
+    // to simulate the same behavior, we need to synchronize both backends first, and do a blocking copy
+    ggml_backend_synchronize(backend_src);
+    ggml_backend_synchronize(backend_dst);
+    ggml_backend_tensor_copy(src, dst);
 }
 
 // events
@@ -1055,11 +1050,10 @@ struct ggml_backend_sched {
     ggml_backend_buffer_type_t bufts[GGML_SCHED_MAX_BACKENDS];
     ggml_gallocr_t galloc;
 
-    // hash keys of the nodes in the graph
-    struct ggml_hash_set    hash_set;
-    // hash values
-    int * tensor_backend_id;
-    struct ggml_tensor * (* tensor_copies)[GGML_SCHED_MAX_BACKENDS][GGML_SCHED_MAX_COPIES];
+    // hash map of the nodes in the graph
+    struct ggml_hash_set  hash_set;
+    int                 * hv_tensor_backend_ids; // [hash_set.size]
+    struct ggml_tensor ** hv_tensor_copies;      // [hash_set.size][n_backends][n_copies]
 
     int * node_backend_ids; // [graph_size]
     int * leaf_backend_ids; // [graph_size]
@@ -1068,7 +1062,7 @@ struct ggml_backend_sched {
     int * prev_leaf_backend_ids; // [graph_size]
 
     // copy of the graph with modified inputs
-    struct ggml_cgraph * graph;
+    struct ggml_cgraph graph;
 
     // graph splits
     struct ggml_backend_sched_split * splits;
@@ -1087,19 +1081,16 @@ struct ggml_backend_sched {
     ggml_backend_sched_eval_callback callback_eval;
     void * callback_eval_user_data;
 
-    bool debug;
+    char * context_buffer;
+    size_t context_buffer_size;
 
-    // align context_buffer to GGML_MEM_ALIGN
-#ifdef _MSC_VER
-    __declspec(align(GGML_MEM_ALIGN))
-#else
-    __attribute__((aligned(GGML_MEM_ALIGN)))
-#endif
-    char context_buffer[GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + sizeof(struct ggml_cgraph)];
+    bool debug;
 };
 
-#define hash_id(tensor) ggml_hash_find_or_insert(sched->hash_set, tensor)
-#define tensor_backend_id(tensor) sched->tensor_backend_id[hash_id(tensor)]
+#define hash_id(tensor) ggml_hash_find_or_insert(&sched->hash_set, tensor)
+#define tensor_backend_id(tensor) sched->hv_tensor_backend_ids[hash_id(tensor)]
+#define tensor_id_copy(id, backend_id, copy_id) sched->hv_tensor_copies[(id) * sched->n_backends * sched->n_copies + (backend_id) * sched->n_copies + (copy_id)]
+#define tensor_copy(tensor, backend_id, copy_id) tensor_id_copy(hash_id(tensor), backend_id, copy_id)
 
 // returns the priority of the backend, lower id is higher priority
 static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backend_t backend) {
@@ -1169,7 +1160,6 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
         return cur_backend_id;
     }
 
-    // assign nodes that use weights to the backend of the weights
     // operations with weights are preferably run on the same backend as the weights
     for (int i = 0; i < GGML_MAX_SRC; i++) {
         const struct ggml_tensor * src = tensor->src[i];
@@ -1275,7 +1265,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
     sched->is_reset = false;
 
     struct ggml_init_params params = {
-        /* .mem_size =   */ sizeof(sched->context_buffer),
+        /* .mem_size =   */ sched->context_buffer_size,
         /* .mem_buffer = */ sched->context_buffer,
         /* .no_alloc =   */ true
     };
@@ -1284,39 +1274,43 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
 
     sched->ctx = ggml_init(params);
     if (sched->ctx == NULL) {
-        fprintf(stderr, "%s: failed to initialize context\n", __func__);
-        GGML_ASSERT(false);
+        GGML_ABORT("%s: failed to initialize context\n", __func__);
     }
 
     // pass 1: assign backends to ops with pre-allocated inputs
     for (int i = 0; i < graph->n_leafs; i++) {
         struct ggml_tensor * leaf = graph->leafs[i];
         int * leaf_backend_id = &tensor_backend_id(leaf);
-        if (*leaf_backend_id != -1) {
-            // do not overwrite user assignments
-            continue;
+        // do not overwrite user assignments
+        if (*leaf_backend_id == -1) {
+            *leaf_backend_id = ggml_backend_sched_backend_id_from_cur(sched, leaf);
         }
-        *leaf_backend_id = ggml_backend_sched_backend_id_from_cur(sched, leaf);
     }
 
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
         int * node_backend_id = &tensor_backend_id(node);
-        if (*node_backend_id != -1) {
-            // do not overwrite user assignments
-            continue;
-        }
-        *node_backend_id = ggml_backend_sched_backend_id_from_cur(sched, node);
-        // src
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (src == NULL) {
+        // do not overwrite user assignments
+        if (*node_backend_id == -1) {
+            *node_backend_id = ggml_backend_sched_backend_id_from_cur(sched, node);
+
+#if 0
+            // src
+            if (node->op == GGML_OP_NONE) {
                 continue;
             }
-            int * src_backend_id = &tensor_backend_id(src);
-            if (*src_backend_id == -1) {
-                *src_backend_id = ggml_backend_sched_backend_id_from_cur(sched, src);
+
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * src = node->src[j];
+                if (src == NULL) {
+                    continue;
+                }
+                int * src_backend_id = &tensor_backend_id(src);
+                if (*src_backend_id == -1) {
+                    *src_backend_id = ggml_backend_sched_backend_id_from_cur(sched, src);
+                }
             }
+#endif
         }
     }
 
@@ -1488,12 +1482,13 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         }
     }
 
-    // pass 4: split graph, find tensors that need to be copied
+    // pass 5: split graph, find tensors that need to be copied
     {
         int i_split = 0;
         struct ggml_backend_sched_split * split = &sched->splits[0];
         // find the backend of the first split, skipping view ops
-        for (int i = 0; i < graph->n_nodes; i++) {
+        int i = 0;
+        for (; i < graph->n_nodes; i++) {
             struct ggml_tensor * node = graph->nodes[i];
             if (!ggml_is_view_op(node->op)) {
                 split->backend_id = tensor_backend_id(node);
@@ -1502,9 +1497,8 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         }
         split->i_start = 0;
         split->n_inputs = 0;
-        memset(split->inputs, 0, sizeof(split->inputs)); //HACK
         int cur_backend_id = split->backend_id;
-        for (int i = 0; i < graph->n_nodes; i++) {
+        for (; i < graph->n_nodes; i++) {
             struct ggml_tensor * node = graph->nodes[i];
 
             if (ggml_is_view_op(node->op)) {
@@ -1513,7 +1507,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
 
             const int node_backend_id = tensor_backend_id(node);
 
-            GGML_ASSERT(node_backend_id != -1); // all nodes should be assigned by now
+            assert(node_backend_id != -1); // all nodes should be assigned by now
 
             // check if we should start a new split based on the sources of the current node
             bool need_new_split = false;
@@ -1527,7 +1521,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                     // by starting a new split, the memory of the previously offloaded weights can be reused
                     if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
                         int src_backend_id = tensor_backend_id(src);
-                        if (src_backend_id != -1 && src_backend_id != cur_backend_id) {
+                        if (src_backend_id != cur_backend_id) {
                             need_new_split = true;
                             break;
                         }
@@ -1536,9 +1530,9 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                     // FIXME: count the number of inputs instead of only checking when full
                     if (split->n_inputs == GGML_SCHED_MAX_SPLIT_INPUTS) {
                         const size_t id = hash_id(src);
-                        int src_backend_id = sched->tensor_backend_id[id];
+                        int src_backend_id = sched->hv_tensor_backend_ids[id];
                         bool supported = ggml_backend_sched_buffer_supported(sched, src, cur_backend_id);
-                        if (src_backend_id != cur_backend_id && sched->tensor_copies[hash_id(src)][cur_backend_id][0] == NULL && !supported) {
+                        if (src_backend_id != cur_backend_id && tensor_id_copy(id, cur_backend_id, 0) == NULL && !supported) {
                             //printf("starting new split because of too many inputs: node %s, input %s\n", node->name, src->name);
                             need_new_split = true;
                             break;
@@ -1570,12 +1564,12 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                     continue;
                 }
 
-                const int src_backend_id = tensor_backend_id(src);
+                size_t src_id = hash_id(src);
+                const int src_backend_id = sched->hv_tensor_backend_ids[src_id];
                 assert(src_backend_id != -1); // all inputs should be assigned by now
 
                 if (src->flags & GGML_TENSOR_FLAG_INPUT && sched->n_copies > 1) {
-                    size_t id = hash_id(src);
-                    if (sched->tensor_copies[id][src_backend_id][0] == NULL) {
+                    if (tensor_id_copy(src_id, src_backend_id, 0) == NULL) {
                         ggml_backend_t backend = sched->backends[src_backend_id];
                         for (int c = 0; c < sched->n_copies; c++) {
                             struct ggml_tensor * tensor_copy;
@@ -1589,7 +1583,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                                 ggml_set_input(tensor_copy);
                                 ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
                             }
-                            sched->tensor_copies[id][src_backend_id][c] = tensor_copy;
+                            tensor_id_copy(src_id, src_backend_id, c) = tensor_copy;
                             SET_CAUSE(tensor_copy, "4.cpy");
                         }
                         int n_graph_inputs = sched->n_graph_inputs++;
@@ -1598,11 +1592,9 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                     }
                 }
 
-                bool supported = ggml_backend_sched_buffer_supported(sched, src, cur_backend_id);
-                if (src_backend_id != cur_backend_id && !supported) {
+                if (src_backend_id != cur_backend_id && !ggml_backend_sched_buffer_supported(sched, src, cur_backend_id)) {
                     // create a copy of the input in the split's backend
-                    const size_t id = hash_id(src);
-                    if (sched->tensor_copies[id][cur_backend_id][0] == NULL) {
+                    if (tensor_id_copy(src_id, cur_backend_id, 0) == NULL) {
                         ggml_backend_t backend = sched->backends[cur_backend_id];
                         for (int c = 0; c < sched->n_copies; c++) {
                             struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
@@ -1611,14 +1603,14 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                                 ggml_set_input(tensor_copy);
                                 ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
                             }
-                            sched->tensor_copies[id][cur_backend_id][c] = tensor_copy;
+                            tensor_id_copy(src_id, cur_backend_id, c) = tensor_copy;
                             SET_CAUSE(tensor_copy, "4.cpy");
                         }
                         int n_inputs = split->n_inputs++;
                         GGML_ASSERT(n_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
                         split->inputs[n_inputs] = src;
                     }
-                    node->src[j] = sched->tensor_copies[id][cur_backend_id][sched->cur_copy];
+                    node->src[j] = tensor_id_copy(src_id, cur_backend_id, sched->cur_copy);
                 }
             }
         }
@@ -1630,7 +1622,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         ggml_backend_sched_print_assignments(sched, graph);
     }
 
-    // swap node_backend_ids and leaf_backend_ids and prevs
+    // swap node_backend_ids and leaf _backend_ids with prevs
     {
         int * tmp = sched->node_backend_ids;
         sched->node_backend_ids = sched->prev_node_backend_ids;
@@ -1641,9 +1633,19 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         sched->prev_leaf_backend_ids = tmp;
     }
 
-    // create copies of the graph for each split
-    // TODO: avoid this copy
-    struct ggml_cgraph * graph_copy = ggml_new_graph_custom(sched->ctx, graph->n_nodes + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2, false);
+    int graph_size = graph->n_nodes + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2;
+    if (sched->graph.size < graph_size) {
+        sched->graph.size = graph_size;
+        sched->graph.nodes = realloc(sched->graph.nodes, graph_size * sizeof(struct ggml_tensor *));
+        sched->graph.leafs = realloc(sched->graph.leafs, graph_size * sizeof(struct ggml_tensor *));
+        GGML_ASSERT(sched->graph.nodes != NULL);
+        GGML_ASSERT(sched->graph.leafs != NULL);
+    }
+    sched->graph.n_nodes = 0;
+    sched->graph.n_leafs = 0;
+
+    struct ggml_cgraph * graph_copy = &sched->graph;
+
     for (int i = 0; i < sched->n_splits; i++) {
         struct ggml_backend_sched_split * split = &sched->splits[i];
         split->graph = ggml_graph_view(graph, split->i_start, split->i_end);
@@ -1654,12 +1656,12 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
 
             struct ggml_tensor * input = split->inputs[j];
             const size_t input_id = hash_id(input);
-            struct ggml_tensor * input_cpy = sched->tensor_copies[input_id][split->backend_id][sched->cur_copy];
+            struct ggml_tensor * input_cpy = tensor_id_copy(input_id, split->backend_id, sched->cur_copy);
 
             // add a dependency to the input source so that it is not freed before the copy is done
             struct ggml_tensor * input_dep = ggml_view_tensor(sched->ctx, input);
             input_dep->src[0] = input;
-            sched->node_backend_ids[graph_copy->n_nodes] = sched->tensor_backend_id[input_id];
+            sched->node_backend_ids[graph_copy->n_nodes] = sched->hv_tensor_backend_ids[input_id];
             graph_copy->nodes[graph_copy->n_nodes++] = input_dep;
 
             // add a dependency to the input copy so that it is allocated at the start of the split
@@ -1681,7 +1683,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
             size_t id = hash_id(input);
             int backend_id = tensor_backend_id(input);
             for (int c = 0; c < sched->n_copies; c++) {
-                struct ggml_tensor * input_cpy = sched->tensor_copies[id][backend_id][c];
+                struct ggml_tensor * input_cpy = tensor_id_copy(id, backend_id, c);
                 sched->leaf_backend_ids[graph_copy->n_leafs] = backend_id;
                 graph_copy->leafs[graph_copy->n_leafs++] = input_cpy;
             }
@@ -1694,7 +1696,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                 struct ggml_tensor * input = split->inputs[j];
                 size_t id = hash_id(input);
                 for (int c = 0; c < sched->n_copies; c++) {
-                    struct ggml_tensor * input_cpy = sched->tensor_copies[id][backend_id][c];
+                    struct ggml_tensor * input_cpy = tensor_id_copy(id, backend_id, c);
                     sched->leaf_backend_ids[graph_copy->n_leafs] = backend_id;
                     graph_copy->leafs[graph_copy->n_leafs++] = input_cpy;
                 }
@@ -1708,13 +1710,11 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         sched->leaf_backend_ids[graph_copy->n_leafs] = tensor_backend_id(leaf);
         graph_copy->leafs[graph_copy->n_leafs++] = leaf;
     }
-
-    sched->graph = graph_copy;
 }
 
 static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
     bool backend_ids_changed = false;
-    for (int i = 0; i < sched->graph->n_nodes; i++) {
+    for (int i = 0; i < sched->graph.n_nodes; i++) {
         if (sched->node_backend_ids[i] != sched->prev_node_backend_ids[i] &&
             sched->bufts[sched->node_backend_ids[i]] != sched->bufts[sched->prev_node_backend_ids[i]]) {
             backend_ids_changed = true;
@@ -1722,7 +1722,7 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
         }
     }
     if (!backend_ids_changed) {
-        for (int i = 0; i < sched->graph->n_leafs; i++) {
+        for (int i = 0; i < sched->graph.n_leafs; i++) {
             if (sched->leaf_backend_ids[i] != sched->prev_leaf_backend_ids[i] &&
                 sched->bufts[sched->leaf_backend_ids[i]] != sched->bufts[sched->prev_leaf_backend_ids[i]]) {
                 backend_ids_changed = true;
@@ -1732,14 +1732,14 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
     }
 
     // allocate graph
-    if (backend_ids_changed || !ggml_gallocr_alloc_graph(sched->galloc, sched->graph)) {
+    if (backend_ids_changed || !ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
         // the re-allocation may cause the split inputs to be moved to a different address
         ggml_backend_sched_synchronize(sched);
 #ifndef NDEBUG
-        fprintf(stderr, "%s: failed to allocate graph, reserving\n", __func__);
+        fprintf(stderr, "%s: failed to allocate graph, reserving (backend_ids_changed = %d)\n", __func__, backend_ids_changed);
 #endif
-        ggml_gallocr_reserve_n(sched->galloc, sched->graph, sched->node_backend_ids, sched->leaf_backend_ids);
-        if (!ggml_gallocr_alloc_graph(sched->galloc, sched->graph)) {
+        ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids);
+        if (!ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
             fprintf(stderr, "%s: failed to allocate graph\n", __func__);
             return false;
         }
@@ -1760,7 +1760,7 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
         for (int j = 0; j < split->n_inputs; j++) {
             ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[j]);
             struct ggml_tensor * input = split->inputs[j];
-            struct ggml_tensor * input_cpy = sched->tensor_copies[hash_id(input)][split_backend_id][sched->cur_copy];
+            struct ggml_tensor * input_cpy = tensor_copy(input, split_backend_id, sched->cur_copy);
 
             if (input->flags & GGML_TENSOR_FLAG_INPUT) {
                 // inputs from the user must be copied immediately to prevent the user overwriting the data before the copy is done
@@ -1777,7 +1777,17 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                 } else {
                     ggml_backend_synchronize(split_backend);
                 }
-                ggml_backend_tensor_copy_async(input_backend, split_backend, input, input_cpy);
+                // try async copy, but if not possible, we can still use a sync copy without synchronizing the dst backend, since we handle the synchronization here with multiple copies and events
+                // TODO: add public function to facilitate this, since applications do not have direct access to the backend interface
+                if (!split_backend->iface.cpy_tensor_async || !split_backend->iface.cpy_tensor_async(input_backend, split_backend, input, input_cpy)) {
+                    ggml_backend_synchronize(input_backend);
+                    if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                        ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
+                    } else {
+                        ggml_backend_synchronize(split_backend);
+                    }
+                    ggml_backend_tensor_copy(input, input_cpy);
+                }
             }
         }
 
@@ -1846,21 +1856,23 @@ ggml_backend_sched_t ggml_backend_sched_new(
     struct ggml_backend_sched * sched = calloc(1, sizeof(struct ggml_backend_sched));
 
     sched->debug = getenv("GGML_SCHED_DEBUG") != NULL;
+    sched->n_backends = n_backends;
+    sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;
 
     // initialize hash table
-    sched->hash_set          = ggml_hash_set_new(graph_size);
-    sched->tensor_backend_id = calloc(sched->hash_set.size, sizeof(sched->tensor_backend_id[0]));
-    sched->tensor_copies     = calloc(sched->hash_set.size, sizeof(sched->tensor_copies[0]));
+    // FIXME: needs to be size*2 to account for leafs (do it in graph_split instead)
+    sched->hash_set    = ggml_hash_set_new(graph_size);
+    sched->hv_tensor_backend_ids = malloc(sched->hash_set.size * sizeof(sched->hv_tensor_backend_ids[0]));
+    sched->hv_tensor_copies      = malloc(sched->hash_set.size * sched->n_backends * sched->n_copies * sizeof(struct ggml_tensor *));
 
     const size_t nodes_size = graph_size + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2;
-    sched->node_backend_ids  = calloc(nodes_size, sizeof(sched->node_backend_ids[0]));
-    sched->leaf_backend_ids  = calloc(nodes_size, sizeof(sched->leaf_backend_ids[0]));
+    sched->node_backend_ids = calloc(nodes_size, sizeof(sched->node_backend_ids[0]));
+    sched->leaf_backend_ids = calloc(nodes_size, sizeof(sched->leaf_backend_ids[0]));
     sched->prev_node_backend_ids = calloc(nodes_size, sizeof(sched->prev_node_backend_ids[0]));
     sched->prev_leaf_backend_ids = calloc(nodes_size, sizeof(sched->prev_leaf_backend_ids[0]));
 
-    sched->n_backends = n_backends;
-
-    sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;
+    sched->context_buffer_size = GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + ggml_graph_overhead_custom(graph_size, false);
+    sched->context_buffer = malloc(sched->context_buffer_size);
 
     const int initial_splits_capacity = 16;
     sched->splits = calloc(initial_splits_capacity, sizeof(sched->splits[0]));
@@ -1895,37 +1907,37 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
     }
     ggml_gallocr_free(sched->galloc);
     ggml_free(sched->ctx);
+    ggml_hash_set_free(&sched->hash_set);
     free(sched->splits);
-    free(sched->hash_set.keys);
-    free(sched->tensor_backend_id);
-    free(sched->tensor_copies);
+    free(sched->hv_tensor_backend_ids);
+    free(sched->hv_tensor_copies);
     free(sched->node_backend_ids);
     free(sched->leaf_backend_ids);
     free(sched->prev_node_backend_ids);
     free(sched->prev_leaf_backend_ids);
+    free(sched->context_buffer);
+    free(sched->graph.nodes);
+    free(sched->graph.leafs);
     free(sched);
 }
 
 void ggml_backend_sched_reset(ggml_backend_sched_t sched) {
     // reset state for the next run
     if (!sched->is_reset) {
-        size_t hash_size = sched->hash_set.size;
-        memset(sched->hash_set.keys,      0, sizeof(sched->hash_set.keys[0])     * hash_size); // NOLINT
-        memset(sched->tensor_backend_id, -1, sizeof(sched->tensor_backend_id[0]) * hash_size);
-        memset(sched->tensor_copies,      0, sizeof(sched->tensor_copies[0])     * hash_size);
-
+        ggml_hash_set_reset(&sched->hash_set);
+        memset(sched->hv_tensor_backend_ids, -1, sched->hash_set.size * sizeof(sched->hv_tensor_backend_ids[0]));
+        memset(sched->hv_tensor_copies,       0, sched->hash_set.size * sched->n_backends * sched->n_copies * sizeof(struct ggml_tensor *));
         sched->is_reset = true;
     }
     sched->is_alloc = false;
 }
 
 bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) {
-    GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes);
+    GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes + measure_graph->n_leafs);
 
     ggml_backend_sched_split_graph(sched, measure_graph);
 
-    // TODO: extract this to a separate function
-    if (!ggml_gallocr_reserve_n(sched->galloc, sched->graph, sched->node_backend_ids, sched->leaf_backend_ids)) {
+    if (!ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids)) {
         return false;
     }
 
@@ -1936,10 +1948,11 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
 }
 
 bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
-    GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes);
+    GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes + graph->n_leafs);
 
     ggml_backend_sched_split_graph(sched, graph);
 
+
     if (!ggml_backend_sched_alloc_splits(sched)) {
         return false;
     }
@@ -2009,6 +2022,7 @@ void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct gg
     GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
     tensor_backend_id(node) = backend_index;
     SET_CAUSE(node, "usr");
+    sched->is_reset = false;
 }
 
 ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node) {
@@ -2051,9 +2065,9 @@ static struct ggml_tensor * graph_copy_dup_tensor(struct ggml_hash_set hash_set,
     GGML_ASSERT(src != NULL);
     GGML_ASSERT(src->data && "graph must be allocated");
 
-    size_t id = ggml_hash_insert(hash_set, src);
-    if (id == GGML_HASHTABLE_ALREADY_EXISTS) {
-        return node_copies[ggml_hash_find(hash_set, src)];
+    size_t id = ggml_hash_insert(&hash_set, src);
+    if (id == GGML_HASHSET_ALREADY_EXISTS) {
+        return node_copies[ggml_hash_find(&hash_set, src)];
     }
 
     struct ggml_tensor * dst = ggml_dup_tensor_layout(src->data && !src->view_src ? ctx_allocated : ctx_unallocated, src);
@@ -2078,7 +2092,7 @@ static struct ggml_tensor * graph_copy_dup_tensor(struct ggml_hash_set hash_set,
     return dst;
 }
 
-static void graph_copy_init_tensor(struct ggml_hash_set hash_set, struct ggml_tensor ** node_copies, bool * node_init, struct ggml_tensor * src) {
+static void graph_copy_init_tensor(struct ggml_hash_set * hash_set, struct ggml_tensor ** node_copies, bool * node_init, struct ggml_tensor * src) {
     size_t id = ggml_hash_find(hash_set, src);
     if (node_init[id]) {
         return;
@@ -2105,10 +2119,7 @@ static void graph_copy_init_tensor(struct ggml_hash_set hash_set, struct ggml_te
 }
 
 struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph) {
-    struct ggml_hash_set hash_set = {
-        /* .size = */ graph->visited_hash_table.size,
-        /* .keys = */ calloc(graph->visited_hash_table.size, sizeof(hash_set.keys[0])) // NOLINT
-    };
+    struct ggml_hash_set hash_set = ggml_hash_set_new(graph->visited_hash_set.size);
     struct ggml_tensor ** node_copies = calloc(hash_set.size, sizeof(node_copies[0])); // NOLINT
     bool * node_init = calloc(hash_set.size, sizeof(node_init[0]));
 
@@ -2123,7 +2134,7 @@ struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, s
 
     if (ctx_allocated == NULL || ctx_unallocated == NULL) {
         fprintf(stderr, "failed to allocate context for graph copy\n");
-        free(hash_set.keys);
+        ggml_hash_set_free(&hash_set);
         free(node_copies);
         free(node_init);
         ggml_free(ctx_allocated);
@@ -2146,7 +2157,7 @@ struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, s
     ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx_allocated, backend);
     if (buffer == NULL) {
         fprintf(stderr, "failed to allocate buffer for graph copy\n");
-        free(hash_set.keys);
+        ggml_hash_set_free(&hash_set);
         free(node_copies);
         free(node_init);
         ggml_free(ctx_allocated);
@@ -2164,19 +2175,19 @@ struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, s
     // copy data and init views
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
-        graph_copy_init_tensor(hash_set, node_copies, node_init, node);
+        graph_copy_init_tensor(&hash_set, node_copies, node_init, node);
     }
 
     // build graph copy
     struct ggml_cgraph * graph_copy = ggml_new_graph_custom(ctx_allocated, graph->size, false);
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
-        struct ggml_tensor * node_copy = node_copies[ggml_hash_find(hash_set, node)];
+        struct ggml_tensor * node_copy = node_copies[ggml_hash_find(&hash_set, node)];
         graph_copy->nodes[i] = node_copy;
     }
     graph_copy->n_nodes = graph->n_nodes;
 
-    free(hash_set.keys);
+    ggml_hash_set_free(&hash_set);
     free(node_copies);
     free(node_init);
 
diff --git a/ggml/src/ggml-blas.cpp b/ggml/src/ggml-blas.cpp
index a37aa407..71373173 100644
--- a/ggml/src/ggml-blas.cpp
+++ b/ggml/src/ggml-blas.cpp
@@ -275,8 +275,7 @@ GGML_CALL static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t
                 break;
 
             default:
-                fprintf(stderr, "%s: unsupported op %s\n", __func__, ggml_op_desc(node));
-                GGML_ASSERT(false);
+                GGML_ABORT("%s: unsupported op %s\n", __func__, ggml_op_desc(node));
         }
     }
 
diff --git a/ggml/src/ggml-cann.cpp b/ggml/src/ggml-cann.cpp
index 9bf7e332..06930ba2 100644
--- a/ggml/src/ggml-cann.cpp
+++ b/ggml/src/ggml-cann.cpp
@@ -120,7 +120,7 @@ static void ggml_cann_log(enum ggml_log_level level, const char* format, ...) {
             file, line);
     GGML_CANN_LOG_ERROR("  %s\n", stmt);
     // abort with GGML_ASSERT to get a stack trace
-    GGML_ASSERT(!"CANN error");
+    GGML_ABORT("CANN error");
 }
 
 /**
@@ -342,7 +342,7 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
         // memory should always buffered. these memory may still needed by
         // tasks in stream.
         // TODO, fix me.
-        GGML_ASSERT(!"Cann buffer pool full, increase MAX_CANN_BUFFERS\n");
+        GGML_ABORT("Cann buffer pool full, increase MAX_CANN_BUFFERS\n");
     }
 };
 
@@ -627,7 +627,6 @@ GGML_CALL static void* ggml_backend_cann_buffer_get_base(
 GGML_CALL static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor,
                                                        const void* src,
                                                        void* dst) {
-    GGML_ASSERT(tensor->op == GGML_OP_NONE);
 
     int64_t n_elems = ggml_nelements(tensor);
     int64_t groups = n_elems / QK4_0;
@@ -679,7 +678,6 @@ GGML_CALL static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor,
  */
 GGML_CALL static void ggml_backend_cann_transform_back_q4_0(
     const ggml_tensor* tensor, void* src, void* dst) {
-    GGML_ASSERT(tensor->op == GGML_OP_NONE);
 
     int64_t n_elems = ggml_nelements(tensor);
     int64_t groups = n_elems / QK4_0;
@@ -898,11 +896,10 @@ GGML_CALL static void ggml_backend_cann_buffer_init_tensor(
  * @param size Size of the data to be copied, in bytes.
  */
 GGML_CALL static void ggml_backend_cann_buffer_set_tensor(
-    ggml_backend_buffer_t buffer, ggml_tensor* tensor, const void* data,
+    ggml_backend_buffer_t buffer, ggml_tensor *tensor, const void *data,
     size_t offset, size_t size) {
-    // GGML_ASSERT(size == ggml_nbytes(tensor));
-    ggml_backend_cann_buffer_context* ctx =
-        (ggml_backend_cann_buffer_context*)buffer->context;
+    ggml_backend_cann_buffer_context *ctx =
+        (ggml_backend_cann_buffer_context *)buffer->context;
 
     ggml_cann_set_device(ctx->device);
     // TODO: refer to cann(#6017), it use thread's default stream.
@@ -910,22 +907,21 @@ GGML_CALL static void ggml_backend_cann_buffer_set_tensor(
     // Why aclrtSynchronizeDevice?
 
     if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpy(tensor->data, size, (const char*)data + offset,
-                              size, ACL_MEMCPY_HOST_TO_DEVICE));
+        ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size, data, size,
+                              ACL_MEMCPY_HOST_TO_DEVICE));
     } else {
-        void* transform_buffer = malloc(size);
-        ggml_backend_cann_transform(tensor, (const char*)data + offset,
-                                    transform_buffer);
+        void *transform_buffer = malloc(size);
+        ggml_backend_cann_transform(tensor, data, transform_buffer);
 
 #ifndef NDEBUG
-        void* check_buffer = malloc(size);
+        void *check_buffer = malloc(size);
         ggml_backend_cann_transform_back(tensor, transform_buffer,
                                          check_buffer);
-        GGML_ASSERT(memcmp((const char*)data + offset, check_buffer, size) ==
-                    0);
+        GGML_ASSERT(memcmp(data, check_buffer, size) == 0);
         free(check_buffer);
 #endif
-        ACL_CHECK(aclrtMemcpy(tensor->data, size, transform_buffer, size,
+        ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size,
+                              transform_buffer, size,
                               ACL_MEMCPY_HOST_TO_DEVICE));
         free(transform_buffer);
     }
@@ -947,21 +943,20 @@ GGML_CALL static void ggml_backend_cann_buffer_set_tensor(
 GGML_CALL static void ggml_backend_cann_buffer_get_tensor(
     ggml_backend_buffer_t buffer, const ggml_tensor* tensor, void* data,
     size_t offset, size_t size) {
-    GGML_ASSERT(size == ggml_nbytes(tensor));
     ggml_backend_cann_buffer_context* ctx =
         (ggml_backend_cann_buffer_context*)buffer->context;
 
     ggml_cann_set_device(ctx->device);
 
     if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpy((char*)data + offset, size, tensor->data, size,
+        ACL_CHECK(aclrtMemcpy(data, size, (char*)tensor->data + offset, size,
                               ACL_MEMCPY_DEVICE_TO_HOST));
     } else {
         void* transform_buffer = malloc(size);
-        ACL_CHECK(aclrtMemcpy(transform_buffer, size, tensor->data, size,
+        ACL_CHECK(aclrtMemcpy(transform_buffer, size,
+                              (char*)tensor->data + offset, size,
                               ACL_MEMCPY_DEVICE_TO_HOST));
-        ggml_backend_cann_transform_back(tensor, transform_buffer,
-                                         (char*)data + offset);
+        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
         free(transform_buffer);
     }
 }
@@ -1450,42 +1445,41 @@ ggml_backend_cann_get_default_buffer_type(ggml_backend_t backend) {
  * @param size Size of the data to copy in bytes.
  */
 GGML_CALL static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend,
-                                                         ggml_tensor* tensor,
-                                                         const void* data,
+                                                         ggml_tensor *tensor,
+                                                         const void *data,
                                                          size_t offset,
                                                          size_t size) {
-    ggml_backend_cann_context* cann_ctx =
-        (ggml_backend_cann_context*)backend->context;
+    ggml_backend_cann_context *cann_ctx =
+        (ggml_backend_cann_context *)backend->context;
 
     if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpyAsync(
-            tensor->data, size, (const char*)data + offset, size,
-            ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
+        ACL_CHECK(aclrtMemcpyAsync((char *)tensor->data + offset, size, data,
+                                   size, ACL_MEMCPY_HOST_TO_DEVICE,
+                                   cann_ctx->stream()));
     } else {
-        void* transform_buffer = malloc(size);
-        ggml_backend_cann_transform(tensor, (const char*)data + offset,
-                                    transform_buffer);
+        void *transform_buffer = malloc(size);
+        ggml_backend_cann_transform(tensor, data, transform_buffer);
 
 #ifndef NDEBUG
-        void* check_buffer = malloc(size);
+        void *check_buffer = malloc(size);
         ggml_backend_cann_transform_back(tensor, transform_buffer,
                                          check_buffer);
-        GGML_ASSERT(memcmp((const char*)data + offset, check_buffer, size));
+        GGML_ASSERT(memcmp(data, check_buffer, size));
         free(check_buffer);
 #endif
-        ACL_CHECK(aclrtMemcpyAsync(tensor->data, size, transform_buffer, size,
-                                   ACL_MEMCPY_HOST_TO_DEVICE,
-                                   cann_ctx->stream()));
+        ACL_CHECK(aclrtMemcpyAsync(
+            (char *)tensor->data + offset, size, transform_buffer, size,
+            ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
         ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
         free(transform_buffer);
     }
 }
 
 GGML_CALL static void ggml_backend_cann_get_tensor_async(
-    ggml_backend_t backend, const ggml_tensor* tensor, void* data,
+    ggml_backend_t backend, const ggml_tensor *tensor, void *data,
     size_t offset, size_t size) {
-    ggml_backend_cann_context* cann_ctx =
-        (ggml_backend_cann_context*)backend->context;
+    ggml_backend_cann_context *cann_ctx =
+        (ggml_backend_cann_context *)backend->context;
     ggml_backend_buffer_t buf =
         tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
@@ -1493,17 +1487,16 @@ GGML_CALL static void ggml_backend_cann_get_tensor_async(
                 "unsupported buffer type");
 
     if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpyAsync((char*)data + offset, size, tensor->data,
+        ACL_CHECK(aclrtMemcpyAsync(data, size, (char *)tensor->data + offset,
                                    size, ACL_MEMCPY_DEVICE_TO_HOST,
                                    cann_ctx->stream()));
     } else {
-        void* transform_buffer = malloc(size);
-        ACL_CHECK(aclrtMemcpyAsync(transform_buffer, size, tensor->data, size,
-                                   ACL_MEMCPY_DEVICE_TO_HOST,
-                                   cann_ctx->stream()));
+        void *transform_buffer = malloc(size);
+        ACL_CHECK(aclrtMemcpyAsync(
+            transform_buffer, size, (char *)tensor->data + offset, size,
+            ACL_MEMCPY_DEVICE_TO_HOST, cann_ctx->stream()));
         ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
-        ggml_backend_cann_transform_back(tensor, transform_buffer,
-                                         (char*)data + offset);
+        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
         free(transform_buffer);
     }
 }
@@ -1559,23 +1552,18 @@ GGML_CALL static bool ggml_backend_cann_cpy_tensor_async(
             return false;
         }
 
+        // need open both directions for memcpyasync between devices.
+        ggml_cann_set_device(cann_ctx_dst->device);
+        ACL_CHECK(aclrtDeviceEnablePeerAccess(cann_ctx_src->device, 0));
         ggml_cann_set_device(cann_ctx_src->device);
         ACL_CHECK(aclrtDeviceEnablePeerAccess(cann_ctx_dst->device, 0));
+
         ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size,
                                    ACL_MEMCPY_DEVICE_TO_DEVICE,
-                                   cann_ctx_dst->stream()));
-
-        // record event on src stream
-        if (!cann_ctx_src->copy_event) {
-            ACL_CHECK(aclrtCreateEvent(&cann_ctx_src->copy_event));
-        }
-
-        ACL_CHECK(
-            aclrtRecordEvent(cann_ctx_src->copy_event, cann_ctx_src->stream()));
+                                   cann_ctx_src->stream()));
 
-        // wait on dst stream for the copy to complete
-        ACL_CHECK(aclrtStreamWaitEvent(cann_ctx_dst->stream(),
-                                       cann_ctx_src->copy_event));
+        //TODO: workaround for Event didn`t work here.
+        aclrtSynchronizeStream(cann_ctx_src->stream());
     } else {
         // src and dst are on the same backend
         ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size,
@@ -1671,10 +1659,13 @@ GGML_CALL static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
             }
         case GGML_OP_MUL_MAT: {
             switch (op->src[0]->type) {
-                // case GGML_TYPE_Q4_0:
                 case GGML_TYPE_F16:
                 case GGML_TYPE_F32:
                 case GGML_TYPE_Q8_0:
+                    // TODO: fix me
+                    // Current groupsize should not be greater than k-1 in
+                    // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize().
+                case GGML_TYPE_Q4_0:
                     return true;
                 default:
                     return false;
@@ -1699,6 +1690,7 @@ GGML_CALL static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
                 case GGML_TYPE_F32:
                 case GGML_TYPE_F16:
                 case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q4_0:
                     return true;
                 default:
                     return false;
@@ -1763,8 +1755,8 @@ static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
  *
  * This function determines whether the CANN backend supports the given backend
  * buffer type by comparing the device context of the backend and buffer type.
- * It returns true if the device associated with the buffer type matches the
- * device associated with the backend.
+ * It returns true if the devices are same between the backend context and
+ * buffer type context.
  *
  * @param backend Pointer to the CANN backend.
  * @param buft Pointer to the backend buffer type to check.
@@ -1773,9 +1765,14 @@ static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
  */
 GGML_CALL static bool ggml_backend_cann_supports_buft(
     ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
-    return buft->iface.get_name == ggml_backend_cann_buffer_type_name;
-
-    GGML_UNUSED(backend);
+    if (ggml_backend_buft_is_cann(buft)) {
+        ggml_backend_cann_context * cann_ctx =
+                        (ggml_backend_cann_context *)backend->context;
+        ggml_backend_cann_buffer_type_context * buft_ctx =
+                        (ggml_backend_cann_buffer_type_context *)buft->context;
+        return buft_ctx->device == cann_ctx->device;
+    }
+    return false;
 }
 
 /**
@@ -1874,7 +1871,7 @@ static void ggml_backend_cann_event_wait(ggml_backend_t backend,
         ACL_CHECK(aclrtStreamWaitEvent(cann_ctx->stream(),
                                        (aclrtEvent)event->context));
     } else {
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 }
 
diff --git a/ggml/src/ggml-cann/acl_tensor.cpp b/ggml/src/ggml-cann/acl_tensor.cpp
index 960ce9a0..d120ce6a 100644
--- a/ggml/src/ggml-cann/acl_tensor.cpp
+++ b/ggml/src/ggml-cann/acl_tensor.cpp
@@ -37,6 +37,10 @@ aclDataType ggml_cann_type_mapping(ggml_type type) {
             return ACL_INT16;
         case GGML_TYPE_I32:
             return ACL_INT32;
+        case GGML_TYPE_Q4_0:
+            return ACL_INT4;
+        case GGML_TYPE_Q8_0:
+            return ACL_INT8;
         default:
             return ACL_DT_UNDEFINED;
     }
@@ -89,33 +93,6 @@ bool ggml_cann_need_bcast(const ggml_tensor* t0, const ggml_tensor* t1) {
     return false;
 }
 
-aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
-                                   size_t type_size, int64_t* ne, size_t* nb,
-                                   int64_t dims, aclFormat format,
-                                   size_t offset) {
-    int64_t tmp_ne[GGML_MAX_DIMS * 2];
-    int64_t tmp_stride[GGML_MAX_DIMS * 2];
-
-    memcpy(tmp_ne, ne, dims * sizeof(int64_t));
-    for (int i = 0; i < dims; i++) {
-        tmp_stride[i] = nb[i] / type_size;
-    }
-
-    std::reverse(tmp_ne, tmp_ne + dims);
-    std::reverse(tmp_stride, tmp_stride + dims);
-
-    int64_t acl_storage_len = 0;
-    for (int i = 0; i < dims; i++) {
-        acl_storage_len += (ne[i] - 1) * nb[i];
-    }
-
-    aclTensor* acl_tensor =
-        aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size,
-                        format, &acl_storage_len, 1, data_ptr);
-
-    return acl_tensor;
-}
-
 int64_t ggml_cann_get_bcast_shape(const ggml_tensor* src0,
                                   const ggml_tensor* src1,
                                   int64_t* bcast_src0_ne,
diff --git a/ggml/src/ggml-cann/acl_tensor.h b/ggml/src/ggml-cann/acl_tensor.h
index 7d0bf04e..4734a9cb 100644
--- a/ggml/src/ggml-cann/acl_tensor.h
+++ b/ggml/src/ggml-cann/acl_tensor.h
@@ -23,6 +23,9 @@
 #ifndef CANN_ACL_TENSOR_H
 #define CANN_ACL_TENSOR_H
 
+#include <algorithm>
+#include <cstring>
+
 #include <aclnn/aclnn_base.h>
 #include "common.h"
 
@@ -65,7 +68,8 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne = null
                              size_t offset = 0);
 
 /**
- * @brief   Creates an ACL tensor from provided parameters.
+ * @brief   Template for creating an ACL tensor from provided parameters. typename TYPE
+ *          should be size_t or float.
  *
  * @details This function creates an ACL tensor using the provided data pointer,
  *          data type, dimensions, strides, format, offset, and additional parameters.
@@ -83,10 +87,34 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne = null
  * @param   offset      Offset in bytes for the ACL tensor data. Defaults to 0.
  * @return  Pointer to the created ACL tensor.
  */
+template<typename TYPE>
 aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
-                             size_t type_size, int64_t* ne, size_t* nb,
-                             int64_t dims, aclFormat format = ACL_FORMAT_ND,
-                             size_t offset = 0);
+                                   TYPE type_size, int64_t* ne, TYPE* nb,
+                                   int64_t dims,
+                                   aclFormat format = ACL_FORMAT_ND,
+                                   size_t offset = 0) {
+    int64_t tmp_ne[GGML_MAX_DIMS * 2];
+    int64_t tmp_stride[GGML_MAX_DIMS * 2];
+
+    memcpy(tmp_ne, ne, dims * sizeof(int64_t));
+    for (int i = 0; i < dims; i++) {
+        tmp_stride[i] = nb[i] / type_size;
+    }
+
+    std::reverse(tmp_ne, tmp_ne + dims);
+    std::reverse(tmp_stride, tmp_stride + dims);
+
+    int64_t acl_storage_len = 0;
+    for (int i = 0; i < dims; i++) {
+        acl_storage_len += (ne[i] - 1) * nb[i];
+    }
+
+    aclTensor* acl_tensor =
+        aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size,
+                        format, &acl_storage_len, 1, data_ptr);
+
+    return acl_tensor;
+}
 
 /**
  * @brief   Checks if tensors require broadcasting based on their shapes.
diff --git a/ggml/src/ggml-cann/aclnn_ops.cpp b/ggml/src/ggml-cann/aclnn_ops.cpp
index a02efc82..8c4132f5 100644
--- a/ggml/src/ggml-cann/aclnn_ops.cpp
+++ b/ggml/src/ggml-cann/aclnn_ops.cpp
@@ -464,9 +464,11 @@ void ggml_cann_group_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     aclTensor* acl_src = ggml_cann_create_tensor(src);
     aclTensor* acl_dst = ggml_cann_create_tensor(dst);
 
-    const float eps = 1e-6f;  // TODO: make this a parameter
     int n_groups = dst->op_params[0];
 
+    float eps;
+    memcpy(&eps, dst->op_params + 1, sizeof(float));
+
     uint64_t workspaceSize = 0;
     aclOpExecutor* executor;
     void* workspaceAddr = nullptr;
@@ -844,7 +846,7 @@ void ggml_cann_pool2d(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
             ggml_cann_max_pool2d(ctx, dst);
             break;
         case GGML_OP_POOL_COUNT:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
             break;
     }
 }
@@ -910,6 +912,13 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 ((ggml_tensor*)dst->extra)->ne);
             return;
         }
+        if (dst->type == GGML_TYPE_Q4_0) {
+            aclrtlaunch_ascendc_quantize_f16_to_q4_0(
+                24, ctx.stream(), src->data, dst->data,
+                ((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb,
+                ((ggml_tensor*)dst->extra)->ne);
+            return;
+        }
         if (dst->type == GGML_TYPE_F16) {
             if (ggml_are_same_shape(src, dst)) {
                 cann_copy(ctx, acl_src, acl_dst);
@@ -931,9 +940,9 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                         ((ggml_tensor*)dst->extra)->nb);
                     return;
                 }
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
             }
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
         if (dst->type == GGML_TYPE_F32) {
             if (ggml_are_same_shape(src, dst)) {
@@ -955,12 +964,12 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                         ((ggml_tensor*)dst->extra)->nb);
                     return;
                 }
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
             }
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
         // TODO
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     } else if (src->type == GGML_TYPE_F32) {
         // TODO: if (src0->type == dst->type && ne00 == ne0 && nb00 == type_size
         //          && nb0 == type_size)
@@ -971,6 +980,13 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 ((ggml_tensor*)dst->extra)->ne);
             return;
         }
+        if (dst->type == GGML_TYPE_Q4_0) {
+            aclrtlaunch_ascendc_quantize_f32_to_q4_0(
+                24, ctx.stream(), src->data, dst->data,
+                ((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb,
+                ((ggml_tensor*)dst->extra)->ne);
+            return;
+        }
         if (dst->type == GGML_TYPE_F32) {
             if (ggml_are_same_shape(src, dst)) {
                 cann_copy(ctx, acl_src, acl_dst);
@@ -991,10 +1007,10 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                         ((ggml_tensor*)dst->extra)->nb);
                     return;
                 }
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
             } else {
                 // TODO: dst not contiguous
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
             }
         }
         if (dst->type == GGML_TYPE_F16) {
@@ -1017,11 +1033,11 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                         ((ggml_tensor*)dst->extra)->nb);
                     return;
                 }
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
             }
         }
         // TODO
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     } else {
         if (ggml_are_same_shape(src, dst)) {
             cann_copy(ctx, acl_src, acl_dst);
@@ -1029,7 +1045,7 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
             ACL_CHECK(aclDestroyTensor(acl_dst));
             return;
         }
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 }
 
@@ -1312,6 +1328,111 @@ aclnnStatus aclnnIm2col(void* workspace, uint64_t workspaceSize,
 #ifdef __cplusplus
 }
 #endif
+
+static void ggml_cann_im2col_2d_post_process(ggml_backend_cann_context& ctx,
+                                             ggml_tensor* dst,
+                                             ggml_tensor* src1,
+                                             aclTensor* tmp_cast_tensor,
+                                             aclTensor* tmp_im2col_tensor) {
+    // Permute: [N, IC * KH * KW, OW * OH] -> [N, OW * OH, IC * KH * KW]
+    int64_t dst_ne[] = {dst->ne[0], dst->ne[1] * dst->ne[2], dst->ne[3]};
+    size_t dst_nb[] = {dst->nb[0], dst->nb[1], dst->nb[3]};
+    aclTensor* acl_dst =
+        ggml_cann_create_tensor(dst, dst_ne, dst_nb, GGML_MAX_DIMS - 1);
+
+    int64_t permute_dim[] = {0, 2, 1};
+    if (src1->type != dst->type) {
+        aclnn_permute(ctx, tmp_cast_tensor, acl_dst, permute_dim, 3);
+    } else {
+        aclnn_permute(ctx, tmp_im2col_tensor, acl_dst, permute_dim, 3);
+    }
+
+    // release
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
+
+static void ggml_cann_im2col_1d_post_process(
+    ggml_backend_cann_context& ctx, ggml_tensor* dst, ggml_tensor* src1,
+    aclTensor* tmp_cast_tensor, aclTensor* tmp_im2col_tensor,
+    const std::vector<int64_t>& im2col_op_params) {
+    // get params
+    const int64_t KH = im2col_op_params[0];
+    const int64_t KW = im2col_op_params[1];
+    const int64_t IW = im2col_op_params[2];
+    const int64_t IC = im2col_op_params[3];
+    const int64_t N = im2col_op_params[4];
+    const int64_t OH = im2col_op_params[5];
+    const int64_t OW = im2col_op_params[6];
+    const int64_t s0 = im2col_op_params[7];
+    const int64_t p0 = im2col_op_params[8];
+    const int64_t d0 = im2col_op_params[9];
+    const int64_t n_bytes_factor = im2col_op_params[10];
+
+    // Permute: [N, IC * KH * KW, OW * OH] ->
+    // [N, OW * OH * n_bytes_factor, IC * KH * KW]
+    aclTensor* tmp_permute_tensor = nullptr;
+    ggml_cann_pool_alloc tmp_permute_allocator(ctx.pool());
+    tmp_permute_allocator.alloc(ggml_nbytes(dst) * n_bytes_factor);
+    void* tmp_permute_buffer = tmp_permute_allocator.get();
+
+    int64_t tmp_permute_ne[] = {IC * KH * KW, OW * OH * n_bytes_factor, N};
+    size_t tmp_permute_nb[GGML_MAX_DIMS - 1];
+    tmp_permute_nb[0] = ggml_type_size(dst->type);
+    for (int i = 1; i < GGML_MAX_DIMS - 1; i++) {
+        tmp_permute_nb[i] = tmp_permute_nb[i - 1] * tmp_permute_ne[i - 1];
+    }
+
+    tmp_permute_tensor = ggml_cann_create_tensor(
+        tmp_permute_buffer, ggml_cann_type_mapping(dst->type),
+        ggml_type_size(dst->type), tmp_permute_ne, tmp_permute_nb,
+        GGML_MAX_DIMS - 1, ACL_FORMAT_ND);
+
+    int64_t permute_dim[] = {0, 2, 1};
+    if (src1->type != dst->type) {
+        aclnn_permute(ctx, tmp_cast_tensor, tmp_permute_tensor, permute_dim, 3);
+    } else {
+        aclnn_permute(ctx, tmp_im2col_tensor, tmp_permute_tensor, permute_dim,
+                      3);
+    }
+
+    // number of times the kernel moves in W dimension
+    const int n_step_w = (IW + 2 * p0 - d0 * (KW - 1) - 1) / s0 + 1;
+    size_t offset;
+    void *cur_dst_buffer = dst->data, *cur_permute_buffer = tmp_permute_buffer;
+
+    // memory copy with offset to restore 1D im2col from 2d
+    if (IC > 1) {
+        offset = IC * KH * KW * n_step_w * ggml_type_size(dst->type);
+        size_t size_cpy = KH * KW * ggml_type_size(dst->type);
+
+        for (int c = 0; c < IC; c++) {
+            cur_permute_buffer = (char*)tmp_permute_buffer + offset +
+                                 KH * KW * c * ggml_type_size(dst->type);
+            cur_dst_buffer = (char*)dst->data +
+                             c * KH * KW * n_step_w * ggml_type_size(dst->type);
+
+            for (int i = 0; i < n_step_w; i++) {
+                ACL_CHECK(aclrtMemcpyAsync(
+                    cur_dst_buffer, size_cpy, cur_permute_buffer, size_cpy,
+                    ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
+                cur_dst_buffer =
+                    (char*)cur_dst_buffer + KH * KW * ggml_type_size(dst->type);
+                cur_permute_buffer = (char*)cur_permute_buffer +
+                                     KH * KW * IC * ggml_type_size(dst->type);
+            }
+        }
+    } else {
+        offset = KH * KW * n_step_w *
+                 ggml_type_size(dst->type);  // equal to ggml_nbytes(dst)
+        ACL_CHECK(aclrtMemcpyAsync(dst->data, offset,
+                                   (char*)tmp_permute_buffer + offset, offset,
+                                   ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
+    }
+
+    // release
+    ACL_CHECK(aclDestroyTensor(tmp_permute_tensor));
+}
+
 void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_tensor* src0 = dst->src[0];  // kernel
     ggml_tensor* src1 = dst->src[1];  // input
@@ -1320,21 +1441,23 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT(dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
 
+    GGML_TENSOR_BINARY_OP_LOCALS;
+
+    // aclnnIm2col only works on 2D. set s1, p1, d1 to 1 to perform 2D
+    // im2col and do post-processing to restore it to 1D.
+    const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
     const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
-    const int32_t s1 = ((const int32_t*)(dst->op_params))[1];
+    const int32_t s1 = is_2D ? ((const int32_t*)(dst->op_params))[1] : 1;
     const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
-    const int32_t p1 = ((const int32_t*)(dst->op_params))[3];
+    const int32_t p1 = is_2D ? ((const int32_t*)(dst->op_params))[3] : 1;
     const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
-    const int32_t d1 = ((const int32_t*)(dst->op_params))[5];
-    const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
+    const int32_t d1 = is_2D ? ((const int32_t*)(dst->op_params))[5] : 1;
 
-    GGML_TENSOR_BINARY_OP_LOCALS;
-
-    const int64_t N = is_2D ? ne13 : ne12;
-    const int64_t IC = is_2D ? ne12 : ne11;
-
-    const int64_t KH = is_2D ? ne01 : 1;
+    const int64_t N = ne13;
+    const int64_t IC = ne12;
+    const int64_t KH = ne01;
     const int64_t KW = ne00;
+    const int64_t IW = ne10;
 
     const int64_t OH = is_2D ? ne2 : 1;
     const int64_t OW = ne1;
@@ -1342,9 +1465,12 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
     GGML_ASSERT(nb10 == sizeof(float));
 
-    // im2col: [N,C,H,W] -> [N, IC * KH * KW, OW * OH]
+    // memory allocated increased to 3x when is_2D == false
+    const int64_t n_bytes_factor = is_2D ? 1 : 3;
+
+    // im2col: [N,C,H,W] -> [N, IC * KH * KW, OW * OH * n_bytes_factor]
     aclTensor* acl_src1 = ggml_cann_create_tensor(src1);
-    int64_t tmp_im2col_ne[] = {OW * OH, IC * KH * KW, N};
+    int64_t tmp_im2col_ne[] = {OW * OH * n_bytes_factor, IC * KH * KW, N};
     size_t tmp_im2col_nb[GGML_MAX_DIMS - 1];
 
     tmp_im2col_nb[0] = ggml_type_size(src1->type);
@@ -1356,8 +1482,10 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     // If dst is f16, tmp_buffer is f32, we need alloc src.typesize *
     // dst.elemcount.
     ggml_cann_pool_alloc im2col_allocator(
-        ctx.pool(), ggml_nelements(dst) * ggml_element_size(src1));
+        ctx.pool(),
+        ggml_nelements(dst) * ggml_element_size(src1) * n_bytes_factor);
     void* tmp_im2col_buffer = im2col_allocator.get();
+
     aclTensor* tmp_im2col_tensor = ggml_cann_create_tensor(
         tmp_im2col_buffer, ggml_cann_type_mapping(src1->type),
         ggml_type_size(src1->type), tmp_im2col_ne, tmp_im2col_nb,
@@ -1380,8 +1508,9 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                                           paddings, strides, tmp_im2col_tensor,
                                           &workspaceSize, &executor));
 
+    ggml_cann_pool_alloc workspace_allocator(ctx.pool());
     if (workspaceSize > 0) {
-        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspace_allocator.alloc(workspaceSize);
         workspaceAddr = workspace_allocator.get();
     }
 
@@ -1391,9 +1520,10 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     // Cast if dst is f16.
     aclTensor* tmp_cast_tensor = nullptr;
     ggml_cann_pool_alloc tmp_cast_allocator(ctx.pool());
+    void* tmp_cast_buffer = nullptr;
     if (src1->type != dst->type) {
-        tmp_cast_allocator.alloc(ggml_nbytes(dst));
-        void* tmp_cast_buffer = tmp_cast_allocator.get();
+        tmp_cast_allocator.alloc(ggml_nbytes(dst) * n_bytes_factor);
+        tmp_cast_buffer = tmp_cast_allocator.get();
         size_t temp_cast_nb[GGML_MAX_DIMS - 1];
         temp_cast_nb[0] = ggml_type_size(dst->type);
         for (int i = 1; i < GGML_MAX_DIMS - 1; i++) {
@@ -1408,24 +1538,21 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                    ggml_cann_type_mapping(dst->type));
     }
 
-    // Permute: [N, IC * KH * KW, OW * OH] -> [N, OW * OH, IC * KH * KW]
-    int64_t dst_ne[] = {dst->ne[0], dst->ne[1] * dst->ne[2], dst->ne[3]};
-    size_t dst_nb[] = {dst->nb[0], dst->nb[1], dst->nb[3]};
-    aclTensor* acl_dst =
-        ggml_cann_create_tensor(dst, dst_ne, dst_nb, GGML_MAX_DIMS - 1);
-
-    int64_t permute_dim[] = {0, 2, 1};
-    if (src1->type != dst->type) {
-        aclnn_permute(ctx, tmp_cast_tensor, acl_dst, permute_dim, 3);
+    // post-processing
+    if (is_2D) {
+        ggml_cann_im2col_2d_post_process(ctx, dst, src1, tmp_cast_tensor,
+                                         tmp_im2col_tensor);
     } else {
-        aclnn_permute(ctx, tmp_im2col_tensor, acl_dst, permute_dim, 3);
+        std::vector<int64_t> im2col_op_params = {
+            KH, KW, IW, IC, N, OH, OW, s0, p0, d0, n_bytes_factor};
+        ggml_cann_im2col_1d_post_process(ctx, dst, src1, tmp_cast_tensor,
+                                         tmp_im2col_tensor, im2col_op_params);
     }
 
     // release
     ACL_CHECK(aclDestroyTensor(acl_src1));
     ACL_CHECK(aclDestroyTensor(tmp_im2col_tensor));
     ACL_CHECK(aclDestroyTensor(tmp_cast_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_dst));
     ACL_CHECK(aclDestroyIntArray(kernel_size));
     ACL_CHECK(aclDestroyIntArray(dilations));
     ACL_CHECK(aclDestroyIntArray(paddings));
@@ -2219,7 +2346,7 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 ((ggml_tensor*)dst->extra)->nb);
             break;
         default:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
             break;
     }
 }
@@ -2352,21 +2479,33 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx,
  * @param dst The destination tensor where the result of the matrix
  * multiplication will be stored.
  */
-static void ggml_cann_mul_mat_q8_0(ggml_backend_cann_context& ctx,
-                                   ggml_tensor* dst) {
+static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
+                                   ggml_tensor* dst,
+                                   const enum ggml_type type) {
     ggml_tensor* src0 = dst->src[0];  // weight
     ggml_tensor* src1 = dst->src[1];  // input
 
     // The shape of the weight is NCHW. Matrix multiplication uses HW dims. HC
     // is regarded as batch. weight need transpose.
     int64_t weight_ne[] = {src0->ne[1], src0->ne[0]};
-    size_t weight_elem_size = sizeof(uint8_t);
-    size_t weight_nb[] = {weight_elem_size * src0->ne[0], weight_elem_size};
+    float weight_elem_size;
+    if (type == GGML_TYPE_Q4_0) {
+        weight_elem_size = float(sizeof(uint8_t)) / 2;
+    }
+    else if (type == GGML_TYPE_Q8_0) {
+        weight_elem_size = float(sizeof(uint8_t));
+    }
+    else {
+        GGML_ABORT("Only support Q4_0 and Q8_0 MUL_MAT");
+    }
+    float weight_nb[] = {weight_elem_size * src0->ne[0], weight_elem_size};
+
     // size of one matrix is element_size * height * width.
     size_t weight_stride = weight_elem_size * src0->ne[0] * src0->ne[1];
     size_t weight_size = weight_stride * src0->ne[2] * src0->ne[3];
 
     // scale stored at the end of weight. Also need transpose.
+    GGML_ASSERT(QK4_0 == QK8_0);
     int64_t scale_ne[] = {src0->ne[1], src0->ne[0] / QK8_0};
     size_t scale_elem_size = sizeof(uint16_t);
     size_t scale_nb[] = {src0->ne[0] / QK8_0 * scale_elem_size,
@@ -2381,10 +2520,10 @@ static void ggml_cann_mul_mat_q8_0(ggml_backend_cann_context& ctx,
     size_t input_nb[] = {input_elem_size, input_elem_size * src1->ne[0]};
     size_t input_stride = input_elem_size * src1->ne[0] * src1->ne[1];
 
+    ggml_cann_pool_alloc input_alloctor(ctx.pool());
     if (src1->type != GGML_TYPE_F16) {
         aclTensor* acl_src1_tensor = ggml_cann_create_tensor(src1);
-        ggml_cann_pool_alloc input_alloctor(
-            ctx.pool(), ggml_nelements(src1) * input_elem_size);
+        input_alloctor.alloc(ggml_nelements(src1) * input_elem_size);
         input_buffer = input_alloctor.get();
 
         int64_t* input_cast_ne = src1->ne;
@@ -2430,8 +2569,9 @@ static void ggml_cann_mul_mat_q8_0(ggml_backend_cann_context& ctx,
                 (char*)input_buffer + batch1 * input_stride, ACL_FLOAT16,
                 input_elem_size, input_ne, input_nb, 2);
             aclTensor* acl_weight_tensor = ggml_cann_create_tensor(
-                (char*)src0->data + batch0 * weight_stride, ACL_INT8,
-                weight_elem_size, weight_ne, weight_nb, 2);
+                (char*)src0->data + batch0 * weight_stride,
+                ggml_cann_type_mapping(type), weight_elem_size, weight_ne,
+                weight_nb, 2);
             aclTensor* acl_scale_tensor = ggml_cann_create_tensor(
                 scale_offset + batch0 * scale_stride, ACL_FLOAT16,
                 scale_elem_size, scale_ne, scale_nb, 2);
@@ -2485,14 +2625,12 @@ void ggml_cann_mul_mat(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         case GGML_TYPE_F16:
             ggml_cann_mat_mul_fp(ctx, dst);
             break;
-        // case GGML_TYPE_Q4_0:
-        //     ggml_cann_mul_mat_q4_0(ctx, dst);
-        //     break;
+        case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q8_0:
-            ggml_cann_mul_mat_q8_0(ctx, dst);
+            ggml_cann_mul_mat_quant(ctx, dst, type);
             break;
         default:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
             break;
     }
 }
diff --git a/ggml/src/ggml-cann/kernels/CMakeLists.txt b/ggml/src/ggml-cann/kernels/CMakeLists.txt
index f12a4d43..5b4fef91 100644
--- a/ggml/src/ggml-cann/kernels/CMakeLists.txt
+++ b/ggml/src/ggml-cann/kernels/CMakeLists.txt
@@ -9,6 +9,7 @@ file(GLOB SRC_FILES
     get_row_q8_0.cpp
     quantize_f32_q8_0.cpp
     quantize_f16_q8_0.cpp
+    quantize_float_to_q4_0.cpp
     dup.cpp
 )
 
@@ -29,4 +30,4 @@ ascendc_library(ascendc_kernels STATIC
     ${SRC_FILES}
 )
 
-#ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)
+# ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)
diff --git a/ggml/src/ggml-cann/kernels/ascendc_kernels.h b/ggml/src/ggml-cann/kernels/ascendc_kernels.h
index bf891475..7e153208 100644
--- a/ggml/src/ggml-cann/kernels/ascendc_kernels.h
+++ b/ggml/src/ggml-cann/kernels/ascendc_kernels.h
@@ -8,6 +8,8 @@
 
 #include "aclrtlaunch_ascendc_quantize_f32_q8_0.h"
 #include "aclrtlaunch_ascendc_quantize_f16_q8_0.h"
+#include "aclrtlaunch_ascendc_quantize_f16_to_q4_0.h"
+#include "aclrtlaunch_ascendc_quantize_f32_to_q4_0.h"
 
 #include "aclrtlaunch_ascendc_dup_by_rows_fp16.h"
 #include "aclrtlaunch_ascendc_dup_by_rows_fp32.h"
diff --git a/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp b/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp
new file mode 100644
index 00000000..9c8c86b6
--- /dev/null
+++ b/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp
@@ -0,0 +1,278 @@
+#include "kernel_operator.h"
+
+using namespace AscendC;
+
+#define BUFFER_NUM 2
+#define Group_Size 32
+
+template <typename SRC_T>
+class QUANTIZE_FLOAT_TO_Q4_0 {
+   public:
+    __aicore__ inline QUANTIZE_FLOAT_TO_Q4_0() {}
+    __aicore__ inline void init(GM_ADDR input, GM_ADDR output,
+                                int64_t *input_ne_ub, size_t *input_nb_ub,
+                                int64_t *output_ne_ub) {
+        // TODO: fix test_case CPY(type_src=f16,type_dst=q4_0,ne=[256,4,4,4],
+        //                         permute=[0,0,0,0]):
+        // [CPY] NMSE = 0.000008343 > 0.000001000 FAIL
+        int64_t op_block_num = GetBlockNum();
+        int64_t op_block_idx = GetBlockIdx();
+
+        // input stride of data elements
+        for (int i = 0; i < 4; i++) {
+            input_ne[i] = input_ne_ub[i];
+            input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
+            output_ne[i] = output_ne_ub[i];
+        }
+
+        // output stride of data elements
+        output_stride[0] = 1;
+        for (int i = 1; i < 4; i++) {
+            output_stride[i] = output_stride[i - 1] * output_ne[i - 1];
+        }
+
+        // scale saved one by one after data:. [group1_scale, group2_scale, ...]
+        scale_ne = input_ne;
+        scale_stride[0] = 1;
+        scale_stride[1] = input_ne[0] / Group_Size;
+        for (int i = 2; i < 4; i++) {
+            scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
+        }
+
+        // split input tensor by rows.
+        uint64_t nr = input_ne[1] * input_ne[2] * input_ne[3];
+        dr = nr / op_block_num;
+
+        uint64_t tails = nr % op_block_num;
+        if (op_block_idx < tails) {
+            dr += 1;
+            ir = dr * op_block_idx;
+        } else {
+            ir = dr * op_block_idx + tails;
+        }
+
+        group_size_in_row = scale_stride[1];
+        int64_t scale_offset = output_ne[0] * output_ne[1] * output_ne[2] *
+                              output_ne[3] * sizeof(uint8_t) / 2;
+
+        input_gm.SetGlobalBuffer((__gm__ SRC_T *)input);
+        output_gm.SetGlobalBuffer((__gm__ int8_t *)output);
+        scale_gm.SetGlobalBuffer((__gm__ half *)(output + scale_offset + ir *
+                                                 group_size_in_row *
+                                                 sizeof(half)));
+
+        pipe.InitBuffer(input_queue, BUFFER_NUM, Group_Size * sizeof(SRC_T));
+        pipe.InitBuffer(output_queue, BUFFER_NUM,
+                            Group_Size * sizeof(int8_t) / 2);
+        pipe.InitBuffer(cast_queue , 1, Group_Size * sizeof(float));
+        pipe.InitBuffer(work_queue, 1, Group_Size * sizeof(float));
+        pipe.InitBuffer(max_queue, 1, Group_Size * sizeof(float));
+        pipe.InitBuffer(min_queue, 1, Group_Size * sizeof(float));
+        pipe.InitBuffer(scale_queue, 1, Group_Size / 2 * sizeof(half));
+        pipe.InitBuffer(int8_queue, 1, Group_Size * sizeof(int8_t));
+        pipe.InitBuffer(half_queue, 1, Group_Size * sizeof(half));
+    }
+
+    __aicore__ inline void copy_in(uint32_t offset) {
+        LocalTensor<SRC_T> input_local = input_queue.AllocTensor<SRC_T>();
+        DataCopy(input_local, input_gm[offset], Group_Size);
+        input_queue.EnQue(input_local);
+    }
+
+    __aicore__ inline void copy_out(uint32_t offset) {
+        // reinterpretcast Group_Size(32) * int4b_t to Group_Size / 2 * int8_t,
+        // and using DataCopyPad to avoid 32 bits align.
+        LocalTensor<int4b_t> output_local = output_queue.DeQue<int4b_t>();
+        LocalTensor<int8_t> output_int8_local =
+                                    output_local.ReinterpretCast<int8_t>();
+
+        DataCopyExtParams dataCopyParams;
+        dataCopyParams.blockCount = 1;
+        dataCopyParams.blockLen = Group_Size / 2  * sizeof(int8_t);
+        DataCopyPad(output_gm[offset], output_int8_local, dataCopyParams);
+
+        output_queue.FreeTensor(output_local);
+    }
+
+    __aicore__ inline void input_to_cast(LocalTensor<float> cast_local,
+                                         LocalTensor<float> input_local) {
+        DataCopy(cast_local, input_local, Group_Size);
+    }
+
+    __aicore__ inline void input_to_cast(LocalTensor<float> cast_local,
+                                         LocalTensor<half> input_local) {
+        Cast(cast_local, input_local, RoundMode::CAST_NONE, Group_Size);
+    }
+
+    __aicore__ inline half calculate_group(int64_t row, int64_t group) {
+        const int64_t i3 = row / (input_ne[1] * input_ne[2]);
+        const int64_t i2 = (row - i3 * input_ne[1] * input_ne[2]) / input_ne[1];
+        const int64_t i1 =
+            row - i3 * input_ne[1] * input_ne[2] - i2 * input_ne[1];
+
+        const int64_t input_offset = i1 * input_stride[1] +
+                                     i2 * input_stride[2] +
+                                     i3 * input_stride[3] + Group_Size * group;
+
+        // output_offset is stride for output_gm which datatype is int8_t and
+        // divided by 2 is needed for int4b_t.
+        const int64_t output_offset = (i1 * output_stride[1] +
+                                       i2 * output_stride[2] +
+                                       i3 * output_stride[3] +
+                                       Group_Size * group) / 2;
+        copy_in(input_offset);
+
+        LocalTensor<SRC_T> input_local = input_queue.DeQue<SRC_T>();
+        LocalTensor<int4b_t> output_local = output_queue.AllocTensor<int4b_t>();
+        LocalTensor<float> cast_local = cast_queue.AllocTensor<float>();
+        LocalTensor<float> work_local = work_queue.AllocTensor<float>();
+        LocalTensor<float> max_local = max_queue.AllocTensor<float>();
+        LocalTensor<float> min_local = min_queue.AllocTensor<float>();
+        LocalTensor<int8_t> int8_local = int8_queue.AllocTensor<int8_t>();
+        LocalTensor<half> half_local = half_queue.AllocTensor<half>();
+
+        input_to_cast(cast_local, input_local);
+
+        ReduceMax(max_local, cast_local, work_local, Group_Size);
+        ReduceMin(min_local, cast_local, work_local, Group_Size);
+        const float max_value = max_local.GetValue(0);
+        const float min_value = min_local.GetValue(0);
+        float d = max_value;
+        if (min_value < 0 && (-1 * min_value) > max_value) {
+            d = min_value;
+        }
+
+        d = d / (-8);
+        if (d != 0) {
+            Muls(cast_local, cast_local, 1.0f / d, Group_Size);
+        }
+
+        // range: [-8,8] -> [0.5,16.5] -> [0,16] -> [0,15] -> [-8,7]
+        float scalar = 8.5f;
+        Adds(cast_local, cast_local, scalar, Group_Size);
+        Cast(cast_local, cast_local, RoundMode::CAST_FLOOR, Group_Size);
+        scalar = 15.0f;
+        Mins(cast_local, cast_local, scalar, Group_Size);
+        scalar = -8.0f;
+        Adds(cast_local, cast_local, scalar, Group_Size);
+
+        // float->half->int4b
+        Cast(half_local, cast_local, RoundMode::CAST_NONE, Group_Size);
+        Cast(output_local, half_local, RoundMode::CAST_NONE, Group_Size);
+
+        output_queue.EnQue(output_local);
+        copy_out(output_offset);
+
+        input_queue.FreeTensor(input_local);
+        work_queue.FreeTensor(work_local);
+        max_queue.FreeTensor(max_local);
+        min_queue.FreeTensor(min_local);
+        int8_queue.FreeTensor(int8_local);
+        half_queue.FreeTensor(half_local);
+        cast_queue.FreeTensor(cast_local);
+        return (half)d;
+    }
+
+    __aicore__ inline void calculate() {
+        LocalTensor<half> scale_local = scale_queue.AllocTensor<half>();
+        uint32_t scale_local_offset = 0;
+        uint32_t scale_global_offset = 0;
+        for (int64_t i = ir; i < ir + dr; i++) {
+            for (int64_t j = 0; j < group_size_in_row; j++) {
+                half scale = calculate_group(i, j);
+                scale_local.SetValue(scale_local_offset++, scale);
+                // Copy Group_Size/2 length data each time.
+                if (scale_local_offset == Group_Size / 2) {
+                    scale_local_offset = 0;
+                    // TODO: OPTIMIZE ME
+                    pipe_barrier(PIPE_ALL);
+                    DataCopy(scale_gm[scale_global_offset], scale_local,
+                                      Group_Size / 2);
+                    pipe_barrier(PIPE_ALL);
+                    scale_global_offset += Group_Size / 2;
+                }
+            }
+        }
+
+        if (scale_local_offset != 0) {
+            pipe_barrier(PIPE_ALL);
+            DataCopyExtParams dataCopyParams;
+            dataCopyParams.blockCount = 1;
+            dataCopyParams.blockLen = scale_local_offset * sizeof(half);
+            DataCopyPad(scale_gm[scale_global_offset], scale_local,
+                        dataCopyParams);
+            pipe_barrier(PIPE_ALL);
+        }
+        scale_queue.FreeTensor(scale_local);
+    }
+
+   private:
+    int64_t input_ne[4];
+    size_t input_stride[4];
+
+    int64_t *scale_ne;
+    size_t scale_stride[4];
+
+    int64_t output_ne[4];
+    size_t output_stride[4];
+
+    int64_t group_size_in_row;
+
+    int64_t ir;
+    int64_t dr;
+
+    TPipe pipe;
+    GlobalTensor<SRC_T> input_gm;
+    GlobalTensor<half> scale_gm;
+    GlobalTensor<int8_t> output_gm;
+    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
+    TQue<QuePosition::VECIN, BUFFER_NUM> work_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> max_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> min_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> scale_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> cast_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> int8_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> half_queue;
+};
+
+template <typename T>
+__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
+    auto gm_ptr = (__gm__ uint8_t *)gm;
+    auto ub_ptr = (uint8_t *)(ub);
+    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
+        *ub_ptr = *gm_ptr;
+    }
+}
+
+extern "C" __global__ __aicore__ void ascendc_quantize_f16_to_q4_0(
+    GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+    GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+    int64_t input_ne_ub[4];
+    size_t input_nb_ub[4];
+    int64_t output_ne_ub[4];
+
+    copy_to_ub(input_ne_gm, input_ne_ub, 32);
+    copy_to_ub(input_nb_gm, input_nb_ub, 32);
+    copy_to_ub(output_ne_gm, output_ne_ub, 32);
+
+    QUANTIZE_FLOAT_TO_Q4_0<half> op;
+    op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
+    op.calculate();
+}
+
+extern "C" __global__ __aicore__ void ascendc_quantize_f32_to_q4_0(
+    GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+    GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+    int64_t input_ne_ub[4];
+    size_t input_nb_ub[4];
+    int64_t output_ne_ub[4];
+
+    copy_to_ub(input_ne_gm, input_ne_ub, 32);
+    copy_to_ub(input_nb_gm, input_nb_ub, 32);
+    copy_to_ub(output_ne_gm, output_ne_ub, 32);
+
+    QUANTIZE_FLOAT_TO_Q4_0<float> op;
+    op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
+    op.calculate();
+}
diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h
index b58e3e4c..57fdeb82 100644
--- a/ggml/src/ggml-common.h
+++ b/ggml/src/ggml-common.h
@@ -26,7 +26,11 @@ typedef half2 ggml_half2;
 
 #define GGML_COMMON_DECL
 #elif defined(GGML_COMMON_DECL_CUDA)
+#if defined(GGML_COMMON_DECL_MUSA)
+#include <musa_fp16.h>
+#else
 #include <cuda_fp16.h>
+#endif
 #include <cstdint>
 
 typedef half  ggml_half;
@@ -527,7 +531,7 @@ static_assert(sizeof(block_iq6_k) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K
 #define GGML_TABLE_END() };
 
 #define GGML_COMMON_IMPL
-#elif defined(GGML_COMMON_IMPL_CUDA) || defined(GGML_COMMON_IMPL_HIP)
+#elif defined(GGML_COMMON_IMPL_CUDA) || defined(GGML_COMMON_IMPL_HIP) || defined(GGML_COMMON_IMPL_MUSA)
 #include <cstdint>
 
 #define GGML_TABLE_BEGIN(type, name, size) static const __device__ type name[size] = {
diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu
index 7641d5b5..f594cd26 100644
--- a/ggml/src/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda.cu
@@ -98,7 +98,7 @@ void ggml_cuda_error(const char * stmt, const char * func, const char * file, in
     GGML_CUDA_LOG_ERROR("  current device: %d, in function %s at %s:%d\n", id, func, file, line);
     GGML_CUDA_LOG_ERROR("  %s\n", stmt);
     // abort with GGML_ASSERT to get a stack trace
-    GGML_ASSERT(!"CUDA error");
+    GGML_ABORT("CUDA error");
 }
 
 // this is faster on Windows
@@ -130,7 +130,22 @@ static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device)
     }
     return res;
 #else
+
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
+    cudaError_t err;
+    if (getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr)
+    {
+        err = cudaMallocManaged(ptr, size);
+    }
+    else
+    {
+        err = cudaMalloc(ptr, size);
+    }
+    return err;
+#else
     return cudaMalloc(ptr, size);
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
+
 #endif
 }
 
@@ -167,7 +182,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
     for (int id = 0; id < info.device_count; ++id) {
         int device_vmm = 0;
 
-#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM) && !defined(GGML_USE_MUSA)
         CUdevice device;
         CU_CHECK(cuDeviceGet(&device, id));
         CU_CHECK(cuDeviceGetAttribute(&device_vmm, CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED, device));
@@ -179,7 +194,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
             alloc_prop.location.id = id;
             CU_CHECK(cuMemGetAllocationGranularity(&info.devices[id].vmm_granularity, &alloc_prop, CU_MEM_ALLOC_GRANULARITY_RECOMMENDED));
         }
-#endif // !defined(GGML_USE_HIPBLAS)
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM) && !defined(GGML_USE_MUSA)
         info.devices[id].vmm = !!device_vmm;
 
         cudaDeviceProp prop;
@@ -315,7 +330,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
 };
 
 // pool with virtual memory
-#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM) && !defined(GGML_USE_MUSA)
 struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
     static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB
 
@@ -409,14 +424,14 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
         GGML_ASSERT(ptr == (void *) (pool_addr + pool_used));
     }
 };
-#endif // !defined(GGML_USE_HIPBLAS)
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM) && !defined(GGML_USE_MUSA)
 
 std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int device) {
-#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM) && !defined(GGML_USE_MUSA)
     if (ggml_cuda_info().devices[device].vmm) {
         return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_vmm(device));
     }
-#endif
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM) && !defined(GGML_USE_MUSA)
     return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_leg(device));
 }
 
@@ -1341,7 +1356,7 @@ static void ggml_cuda_set_peer_access(const int n_tokens, int main_device) {
 static cudaError_t ggml_cuda_Memcpy2DPeerAsync(
     void * dst, int dstDevice, size_t dpitch, void * src, int srcDevice, size_t spitch, size_t width, size_t height, cudaStream_t stream) {
 
-#if !defined(GGML_USE_HIPBLAS)
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
     // cudaMemcpy2DAsync may fail with copies between vmm pools of different devices
     cudaMemcpy3DPeerParms p = {};
     p.dstDevice = dstDevice;
@@ -1355,7 +1370,7 @@ static cudaError_t ggml_cuda_Memcpy2DPeerAsync(
     GGML_UNUSED(dstDevice);
     GGML_UNUSED(srcDevice);
     return cudaMemcpy2DAsync(dst, dpitch, src, spitch, width, height, cudaMemcpyDeviceToDevice, stream);
-#endif // !defined(GGML_USE_HIPBLAS)
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
 }
 
 static void ggml_cuda_op_mul_mat(
@@ -1486,7 +1501,7 @@ static void ggml_cuda_op_mul_mat(
         }
 
         // If src0 is on a temporary compute buffers (partial offloading) there may be some padding that needs to be cleared:
-        if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
+        if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_is_quantized(src0->type) && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
             const int64_t nbytes_data    = ggml_row_size(src0->type, (dev[id].row_high - dev[id].row_low)*ne00);
             const int64_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
             CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data , 0, nbytes_padding, stream));
@@ -1596,7 +1611,7 @@ static void ggml_cuda_op_mul_mat(
                     CUDA_CHECK(ggml_cuda_cpy_tensor_2d(
                                 src1_ddf_i, src1, i03, i02, src1_col_0, src1_col_0+src1_ncols, stream));
                 } else {
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
                 }
 
                 if (quantize_src1 && !src1_is_contiguous) {
@@ -1828,6 +1843,9 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
         }
     }
 #else
+#ifdef GGML_USE_MUSA
+    GGML_ASSERT(false);
+#else // !GGML_USE_MUSA
     if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
         // there is no broadcast and src0, src1 are contiguous across dims 2, 3
         // use cublasGemmStridedBatchedEx
@@ -1870,6 +1888,7 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
                 cu_compute_type,
                 CUBLAS_GEMM_DEFAULT_TENSOR_OP));
     }
+#endif // GGML_USE_MUSA
 #endif
 
     if (dst->op_params[0] == GGML_PREC_DEFAULT) {
@@ -1881,10 +1900,9 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
 static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool split = ggml_backend_buffer_is_cuda_split(src0->buffer);
 
-    bool use_dequantize_mul_mat_vec = (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16)
+    bool use_dequantize_mul_mat_vec = ggml_cuda_dmmv_type_supported(src0->type)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
-        && src0->ne[0] % GGML_CUDA_DMMV_X == 0 && src0->ne[0] >= GGML_CUDA_DMMV_X*2
-        && src1->ne[1] == 1;
+        && src0->ne[0] % (GGML_CUDA_DMMV_X*2) == 0 && src1->ne[1] == 1;
     bool          use_mul_mat_vec_q =  ggml_is_quantized(src0->type)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
         && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
@@ -2340,33 +2358,35 @@ GGML_CALL static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend,
 }
 
 GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const ggml_tensor * src, ggml_tensor * dst) {
-    GGML_ASSERT(ggml_backend_is_cuda(backend_src) || ggml_backend_is_cuda(backend_dst));
-
     ggml_backend_buffer_t buf_src = src->view_src ? src->view_src->buffer : src->buffer;
     ggml_backend_buffer_t buf_dst = dst->view_src ? dst->view_src->buffer : dst->buffer;
 
-    if (!ggml_backend_buffer_is_cuda(src->buffer)) {
+    if (!ggml_backend_is_cuda(backend_src) || !ggml_backend_is_cuda(backend_dst)) {
         return false;
     }
 
-    if (!ggml_backend_buffer_is_cuda(dst->buffer)) {
+    if (!ggml_backend_buffer_is_cuda(src->buffer) || !ggml_backend_buffer_is_cuda(dst->buffer)) {
         return false;
     }
 
-    // device -> device
+    // device -> device copy
     ggml_backend_cuda_context * cuda_ctx_src = (ggml_backend_cuda_context *)backend_src->context;
     ggml_backend_cuda_context * cuda_ctx_dst = (ggml_backend_cuda_context *)backend_dst->context;
 
-    if (backend_src != backend_dst) {
-        ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *)buf_src->context;
-        ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *)buf_dst->context;
+    ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *)buf_src->context;
+    ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *)buf_dst->context;
 
-        GGML_ASSERT(cuda_ctx_src->device == buf_ctx_src->device);
-        GGML_ASSERT(cuda_ctx_dst->device == buf_ctx_dst->device);
+    if (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device) {
+#ifndef NDEBUG
+        GGML_CUDA_LOG_WARN("%s: backend and buffer devices do not match\n", __func__);
+#endif
+        return false;
+    }
 
+    if (backend_src != backend_dst) {
         // copy on src stream
         if (cuda_ctx_src->device == cuda_ctx_dst->device) {
-            CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_dst->stream()));
+            CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream()));
         } else {
 #ifdef GGML_CUDA_NO_PEER_COPY
             return false;
@@ -2375,7 +2395,7 @@ GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_
 #endif
         }
 
-        // record event on src stream
+        // record event on src stream after the copy
         if (!cuda_ctx_src->copy_event) {
             ggml_cuda_set_device(cuda_ctx_src->device);
             CUDA_CHECK(cudaEventCreateWithFlags(&cuda_ctx_src->copy_event, cudaEventDisableTiming));
@@ -2387,7 +2407,7 @@ GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_
         CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx_dst->stream(), cuda_ctx_src->copy_event, 0));
     } else {
         // src and dst are on the same backend
-        CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_dst->stream()));
+        CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream()));
     }
     return true;
 }
@@ -2724,11 +2744,12 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_MUL_MAT_ID:
             {
                 struct ggml_tensor * a = op->src[0];
-                if (op->op == GGML_OP_MUL_MAT) {
-                    struct ggml_tensor * b = op->src[1];
-                    if (a->ne[3] != b->ne[3]) {
-                        return false;
-                    }
+                struct ggml_tensor * b = op->src[1];
+                if (b->type == GGML_TYPE_F16 && a->type != GGML_TYPE_F16) {
+                    return false;
+                }
+                if (op->op == GGML_OP_MUL_MAT && a->ne[3] != b->ne[3]) {
+                    return false;
                 }
                 switch (a->type) {
                     case GGML_TYPE_F32:
@@ -2867,7 +2888,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
             return true;
         case GGML_OP_FLASH_ATTN_EXT:
 #if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-            return op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128;
+            return (op->src[0]->ne[0] == 64 && op->src[1]->type == GGML_TYPE_F16) || op->src[0]->ne[0] == 128;
 #else
             if (op->src[0]->ne[0] == 128) {
                 return true;
@@ -2953,7 +2974,7 @@ static void ggml_backend_cuda_event_wait(ggml_backend_t backend, ggml_backend_ev
 
         CUDA_CHECK(cudaLaunchHostFunc(cuda_ctx->stream(), wait_fn, event));
 #endif
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 }
 
@@ -3035,7 +3056,7 @@ GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size
         return false;
     }
 
-#if CUDART_VERSION >= 11100
+#if CUDART_VERSION >= 11100 || defined(GGML_USE_MUSA)
     cudaError_t err = cudaHostRegister(buffer, size, cudaHostRegisterPortable | cudaHostRegisterReadOnly);
     if (err != cudaSuccess) {
         // clear the error
diff --git a/ggml/src/ggml-cuda/argsort.cu b/ggml/src/ggml-cuda/argsort.cu
index 15757ca1..607ded85 100644
--- a/ggml/src/ggml-cuda/argsort.cu
+++ b/ggml/src/ggml-cuda/argsort.cu
@@ -81,7 +81,7 @@ static void argsort_f32_i32_cuda(const float * x, int * dst, const int ncols, co
     } else if (order == GGML_SORT_ORDER_DESC) {
         k_argsort_f32_i32<GGML_SORT_ORDER_DESC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
     } else {
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 }
 
diff --git a/ggml/src/ggml-cuda/binbcast.cu b/ggml/src/ggml-cuda/binbcast.cu
index 76cc01b2..62d115f1 100644
--- a/ggml/src/ggml-cuda/binbcast.cu
+++ b/ggml/src/ggml-cuda/binbcast.cu
@@ -259,7 +259,7 @@ static void ggml_cuda_op_bin_bcast(
     } else {
         fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__,
             ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type));
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 }
 
diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
index 07a53bcd..9aff6c13 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -12,6 +12,10 @@
 #else
 #define GGML_COMMON_DECL_CUDA
 #define GGML_COMMON_IMPL_CUDA
+#if defined(GGML_USE_MUSA)
+#define GGML_COMMON_DECL_MUSA
+#define GGML_COMMON_IMPL_MUSA
+#endif
 #endif
 #include "ggml-common.h"
 
@@ -23,111 +27,11 @@
 #include <vector>
 
 #if defined(GGML_USE_HIPBLAS)
-#include <hip/hip_runtime.h>
-#include <hipblas/hipblas.h>
-#include <hip/hip_fp16.h>
-#ifdef __HIP_PLATFORM_AMD__
-// for rocblas_initialize()
-#include "rocblas/rocblas.h"
-#endif // __HIP_PLATFORM_AMD__
-#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F
-#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
-#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
-#define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
-#define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
-#define CUBLAS_OP_N HIPBLAS_OP_N
-#define CUBLAS_OP_T HIPBLAS_OP_T
-#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
-#define CUBLAS_TF32_TENSOR_OP_MATH 0
-#define CUDA_R_16F  HIPBLAS_R_16F
-#define CUDA_R_32F  HIPBLAS_R_32F
-#define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
-#define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6
-#define cublasCreate hipblasCreate
-#define cublasDestroy hipblasDestroy
-#define cublasGemmEx hipblasGemmEx
-#define cublasGemmBatchedEx hipblasGemmBatchedEx
-#define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx
-#define cublasHandle_t hipblasHandle_t
-#define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS
-#define cublasSetStream hipblasSetStream
-#define cublasSgemm hipblasSgemm
-#define cublasStatus_t hipblasStatus_t
-#define cudaDataType_t hipblasDatatype_t //deprecated, new hipblasDatatype not in 5.6
-#define cudaDeviceCanAccessPeer hipDeviceCanAccessPeer
-#define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
-#define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
-#define cudaDeviceProp hipDeviceProp_t
-#define cudaDeviceSynchronize hipDeviceSynchronize
-#define cudaError_t hipError_t
-#define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled
-#define cudaErrorPeerAccessNotEnabled hipErrorPeerAccessNotEnabled
-#define cudaEventCreateWithFlags hipEventCreateWithFlags
-#define cudaEventDisableTiming hipEventDisableTiming
-#define cudaEventRecord hipEventRecord
-#define cudaEventSynchronize hipEventSynchronize
-#define cudaEvent_t hipEvent_t
-#define cudaEventDestroy hipEventDestroy
-#define cudaFree hipFree
-#define cudaFreeHost hipHostFree
-#define cudaGetDevice hipGetDevice
-#define cudaGetDeviceCount hipGetDeviceCount
-#define cudaGetDeviceProperties hipGetDeviceProperties
-#define cudaGetErrorString hipGetErrorString
-#define cudaGetLastError hipGetLastError
-#define cudaHostRegister hipHostRegister
-#define cudaHostRegisterPortable hipHostRegisterPortable
-#define cudaHostRegisterReadOnly hipHostRegisterReadOnly
-#define cudaHostUnregister hipHostUnregister
-#define cudaLaunchHostFunc hipLaunchHostFunc
-#define cudaMalloc hipMalloc
-#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
-#define cudaMemcpy hipMemcpy
-#define cudaMemcpyAsync hipMemcpyAsync
-#define cudaMemcpyPeerAsync hipMemcpyPeerAsync
-#define cudaMemcpy2DAsync hipMemcpy2DAsync
-#define cudaMemcpyDeviceToDevice hipMemcpyDeviceToDevice
-#define cudaMemcpyDeviceToHost hipMemcpyDeviceToHost
-#define cudaMemcpyHostToDevice hipMemcpyHostToDevice
-#define cudaMemcpyKind hipMemcpyKind
-#define cudaMemset hipMemset
-#define cudaMemsetAsync hipMemsetAsync
-#define cudaMemGetInfo hipMemGetInfo
-#define cudaOccupancyMaxPotentialBlockSize hipOccupancyMaxPotentialBlockSize
-#define cudaSetDevice hipSetDevice
-#define cudaStreamCreateWithFlags hipStreamCreateWithFlags
-#define cudaStreamDestroy hipStreamDestroy
-#define cudaStreamFireAndForget hipStreamFireAndForget
-#define cudaStreamNonBlocking hipStreamNonBlocking
-#define cudaStreamPerThread hipStreamPerThread
-#define cudaStreamSynchronize hipStreamSynchronize
-#define cudaStreamWaitEvent(stream, event, flags) hipStreamWaitEvent(stream, event, flags)
-#define cudaStream_t hipStream_t
-#define cudaSuccess hipSuccess
-#define __trap() do { abort(); __builtin_unreachable(); } while(0)
-#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
-#define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED
-#define CUBLAS_STATUS_ALLOC_FAILED HIPBLAS_STATUS_ALLOC_FAILED
-#define CUBLAS_STATUS_INVALID_VALUE HIPBLAS_STATUS_INVALID_VALUE
-#define CUBLAS_STATUS_ARCH_MISMATCH HIPBLAS_STATUS_ARCH_MISMATCH
-#define CUBLAS_STATUS_MAPPING_ERROR HIPBLAS_STATUS_MAPPING_ERROR
-#define CUBLAS_STATUS_EXECUTION_FAILED HIPBLAS_STATUS_EXECUTION_FAILED
-#define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR
-#define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED
+#include "vendors/hip.h"
+#elif defined(GGML_USE_MUSA)
+#include "vendors/musa.h"
 #else
-#include <cuda_runtime.h>
-#include <cuda.h>
-#include <cublas_v2.h>
-#include <cuda_fp16.h>
-
-#if CUDART_VERSION < 11020
-#define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED
-#define CUBLAS_TF32_TENSOR_OP_MATH CUBLAS_TENSOR_OP_MATH
-#define CUBLAS_COMPUTE_16F CUDA_R_16F
-#define CUBLAS_COMPUTE_32F CUDA_R_32F
-#define cublasComputeType_t cudaDataType_t
-#endif // CUDART_VERSION < 11020
-
+#include "vendors/cuda.h"
 #endif // defined(GGML_USE_HIPBLAS)
 
 #define STRINGIZE_IMPL(...) #__VA_ARGS__
@@ -168,7 +72,7 @@ void ggml_cuda_error(const char * stmt, const char * func, const char * file, in
 
 #define CUDA_CHECK(err) CUDA_CHECK_GEN(err, cudaSuccess, cudaGetErrorString)
 
-#if CUDART_VERSION >= 12000
+#if CUDART_VERSION >= 12000 || defined(GGML_USE_MUSA)
     static const char * cublas_get_error_str(const cublasStatus_t err) {
         return cublasGetStatusString(err);
     }
@@ -200,7 +104,7 @@ static const char * cu_get_error_str(CUresult err) {
 #define CU_CHECK(err) CUDA_CHECK_GEN(err, CUDA_SUCCESS, cu_get_error_str)
 #endif
 
-#if CUDART_VERSION >= 11100
+#if CUDART_VERSION >= 11100 || defined(GGML_USE_MUSA)
 #define GGML_CUDA_ASSUME(x) __builtin_assume(x)
 #else
 #define GGML_CUDA_ASSUME(x)
@@ -212,93 +116,7 @@ typedef half2 dfloat2;
 #else
 typedef float dfloat; // dequantize float
 typedef float2 dfloat2;
-#endif //GGML_CUDA_F16
-
-#if defined(GGML_USE_HIPBLAS)
-#define __CUDA_ARCH__ 1300
-
-#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
-    defined(__gfx1150__) || defined(__gfx1151__)
-#define RDNA3
-#endif
-
-#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || defined(__gfx1033__) || \
-    defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || defined(__gfx1037__)
-#define RDNA2
-#endif
-
-#if defined(__gfx1010__) || defined(__gfx1012__)
-#define RDNA1
-#endif
-
-#ifndef __has_builtin
-    #define __has_builtin(x) 0
-#endif
-
-typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
-typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
-static __device__ __forceinline__ int __vsubss4(const int a, const int b) {
-    const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
-    const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
-#if __has_builtin(__builtin_elementwise_sub_sat)
-    const int8x4_t c = __builtin_elementwise_sub_sat(va, vb);
-    return reinterpret_cast<const int &>(c);
-#else
-    int8x4_t c;
-    int16_t tmp;
-#pragma unroll
-    for (int i = 0; i < 4; i++) {
-        tmp = va[i] - vb[i];
-        if(tmp > std::numeric_limits<int8_t>::max()) tmp = std::numeric_limits<int8_t>::max();
-        if(tmp < std::numeric_limits<int8_t>::min()) tmp = std::numeric_limits<int8_t>::min();
-        c[i] = tmp;
-    }
-    return reinterpret_cast<int &>(c);
-#endif // __has_builtin(__builtin_elementwise_sub_sat)
-}
-
-static __device__ __forceinline__ int __vsub4(const int a, const int b) {
-    return __vsubss4(a, b);
-}
-
-static __device__ __forceinline__ unsigned int __vcmpeq4(unsigned int a, unsigned int b) {
-    const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
-    const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
-    unsigned int c;
-    uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
-#pragma unroll
-    for (int i = 0; i < 4; ++i) {
-        vc[i] = va[i] == vb[i] ? 0xff : 0x00;
-    }
-    return c;
-}
-
-static __device__ __forceinline__ unsigned int __vcmpne4(unsigned int a, unsigned int b) {
-    const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
-    const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
-    unsigned int c;
-    uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
-#pragma unroll
-    for (int i = 0; i < 4; ++i) {
-        vc[i] = va[i] == vb[i] ? 0x00 : 0xff;
-    }
-    return c;
-}
-
-#if defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
-// __shfl_xor() for half2 was added in ROCm 5.6
-static __device__ __forceinline__ half2 __shfl_xor(half2 var, int laneMask, int width) {
-    typedef union half2_b32 {
-        half2 val;
-        int   b32;
-    } half2_b32_t;
-    half2_b32_t tmp;
-    tmp.val = var;
-    tmp.b32 = __shfl_xor(tmp.b32, laneMask, width);
-    return tmp.val;
-}
-#endif // defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
-#endif // defined(GGML_USE_HIPBLAS)
+#endif // GGML_CUDA_F16
 
 #if (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
 #define FP16_AVAILABLE
@@ -348,7 +166,7 @@ static __device__ void no_device_code(
 #ifdef __CUDA_ARCH__
 #define NO_DEVICE_CODE no_device_code(__FILE__, __LINE__, __FUNCTION__, __CUDA_ARCH__, STRINGIZE(__CUDA_ARCH_LIST__))
 #else
-#define NO_DEVICE_CODE //GGML_ASSERT(false && "NO_DEVICE_CODE not valid in host code.")
+#define NO_DEVICE_CODE //GGML_ABORT("NO_DEVICE_CODE not valid in host code.")
 #endif // __CUDA_ARCH__
 
 static __device__ __forceinline__ float warp_reduce_sum(float x) {
@@ -455,7 +273,7 @@ static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half
     const uint32_t mask_high = 0xFFFF0000 * (float(__high2half(a)) > float(__high2half(b)));
     return mask_low | mask_high;
 }
-#endif // CUDART_VERSION < 12000
+#endif // CUDART_VERSION < CUDART_HMASK
 
 static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, int c) {
 #if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
diff --git a/ggml/src/ggml-cuda/convert.cu b/ggml/src/ggml-cuda/convert.cu
index f76c80dc..70305404 100644
--- a/ggml/src/ggml-cuda/convert.cu
+++ b/ggml/src/ggml-cuda/convert.cu
@@ -162,7 +162,6 @@ static __global__ void dequantize_block_iq2_tn(const void * __restrict__ vx, dst
     const int64_t tid = threadIdx.x;
     const int64_t n   = tid/32;
     const int64_t l   = tid - 32*n;
-    const int64_t is  = 8*n + l/16;
 
     const uint8_t q = x[i].qs[32*n + l];
     dst_t * y = yy + i*QK_K + 128*n;
diff --git a/ggml/src/ggml-cuda/cpy.cu b/ggml/src/ggml-cuda/cpy.cu
index 3db57034..aad34bfe 100644
--- a/ggml/src/ggml-cuda/cpy.cu
+++ b/ggml/src/ggml-cuda/cpy.cu
@@ -451,7 +451,7 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
     } else {
         fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__,
                 ggml_type_name(src0->type), ggml_type_name(src1->type));
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 }
 
@@ -484,6 +484,6 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
     } else {
         fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__,
                 ggml_type_name(src0->type), ggml_type_name(src1->type));
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 }
diff --git a/ggml/src/ggml-cuda/dmmv.cu b/ggml/src/ggml-cuda/dmmv.cu
index 174489e0..96a5adef 100644
--- a/ggml/src/ggml-cuda/dmmv.cu
+++ b/ggml/src/ggml-cuda/dmmv.cu
@@ -500,7 +500,7 @@ static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, cons
 }
 
 static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     // the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead
     const dim3 block_nums(block_num_y, 1, 1);
@@ -510,7 +510,7 @@ static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y,
 }
 
 static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
@@ -519,7 +519,7 @@ static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y,
 }
 
 static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
@@ -528,7 +528,7 @@ static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y,
 }
 
 static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
@@ -537,7 +537,7 @@ static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y,
 }
 
 static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
@@ -588,7 +588,7 @@ static void dequantize_mul_mat_vec_q6_K_cuda(const void * vx, const float * y, f
 }
 
 static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
@@ -662,7 +662,7 @@ void ggml_cuda_op_dequantize_mul_mat_vec(
             convert_mul_mat_vec_f16_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
             break;
         default:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
             break;
     }
 
@@ -672,3 +672,12 @@ void ggml_cuda_op_dequantize_mul_mat_vec(
     GGML_UNUSED(src1_ncols);
     GGML_UNUSED(src1_padded_row_size);
 }
+
+bool ggml_cuda_dmmv_type_supported(ggml_type src0_type) {
+    return src0_type == GGML_TYPE_Q4_0 || src0_type == GGML_TYPE_Q4_1 ||
+        src0_type == GGML_TYPE_Q5_0 || src0_type == GGML_TYPE_Q5_1 ||
+        src0_type == GGML_TYPE_Q8_0 || src0_type == GGML_TYPE_Q2_K ||
+        src0_type == GGML_TYPE_Q3_K || src0_type == GGML_TYPE_Q4_K ||
+        src0_type == GGML_TYPE_Q5_K || src0_type == GGML_TYPE_Q6_K ||
+        src0_type == GGML_TYPE_F16;
+}
diff --git a/ggml/src/ggml-cuda/dmmv.cuh b/ggml/src/ggml-cuda/dmmv.cuh
index 4c5ebd47..e727eb97 100644
--- a/ggml/src/ggml-cuda/dmmv.cuh
+++ b/ggml/src/ggml-cuda/dmmv.cuh
@@ -16,3 +16,5 @@ void ggml_cuda_op_dequantize_mul_mat_vec(
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
     const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
     const int64_t src1_padded_row_size, cudaStream_t stream);
+
+bool ggml_cuda_dmmv_type_supported(ggml_type src0_type);
diff --git a/ggml/src/ggml-cuda/fattn-common.cuh b/ggml/src/ggml-cuda/fattn-common.cuh
index f24312dd..950fd93d 100644
--- a/ggml/src/ggml-cuda/fattn-common.cuh
+++ b/ggml/src/ggml-cuda/fattn-common.cuh
@@ -564,7 +564,7 @@ static void on_no_fattn_vec_case(const int D) {
         fprintf(stderr, "Unsupported KV type combination for head_size 64.\n");
         fprintf(stderr, "By default only f16 KV cache is supported.\n");
         fprintf(stderr, "Compile with GGML_CUDA_FA_ALL_QUANTS for V cache quantization support.\n");
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     } else if (D == 128) {
         fprintf(stderr, "Unsupported KV type combination for head_size 128.\n");
         fprintf(stderr, "Supported combinations:\n");
@@ -572,11 +572,11 @@ static void on_no_fattn_vec_case(const int D) {
         fprintf(stderr, "  - K == q8_0, V == q8_0,  8.50 BPV\n");
         fprintf(stderr, "  - K == f16,  V == f16,  16.00 BPV\n");
         fprintf(stderr, "Compile with GGML_CUDA_FA_ALL_QUANTS for all combinations of q4_0, q4_1, q5_0, q5_1, q8_0, and f16.\n");
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     } else {
         fprintf(stderr, "Unsupported KV type combination for head_size 256.\n");
         fprintf(stderr, "Only f16 is supported.\n");
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 }
 
diff --git a/ggml/src/ggml-cuda/fattn-tile-f16.cu b/ggml/src/ggml-cuda/fattn-tile-f16.cu
index c6c35134..1b2fd500 100644
--- a/ggml/src/ggml-cuda/fattn-tile-f16.cu
+++ b/ggml/src/ggml-cuda/fattn-tile-f16.cu
@@ -287,7 +287,7 @@ void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
             launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
         } break;
         default: {
-            GGML_ASSERT(false && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
+            GGML_ABORT("FlashAttention without tensor cores only supports head sizes 64 and 128.");
         } break;
     }
 }
diff --git a/ggml/src/ggml-cuda/fattn-tile-f32.cu b/ggml/src/ggml-cuda/fattn-tile-f32.cu
index 15e22f49..f3e68dbf 100644
--- a/ggml/src/ggml-cuda/fattn-tile-f32.cu
+++ b/ggml/src/ggml-cuda/fattn-tile-f32.cu
@@ -284,7 +284,7 @@ void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
             launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
         } break;
         default: {
-            GGML_ASSERT(false && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
+            GGML_ABORT("FlashAttention without tensor cores only supports head sizes 64 and 128.");
         } break;
     }
 }
diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu
index 38d30b21..29f608b0 100644
--- a/ggml/src/ggml-cuda/fattn.cu
+++ b/ggml/src/ggml-cuda/fattn.cu
@@ -38,7 +38,7 @@ static void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, g
                     ggml_cuda_flash_attn_ext_wmma_f16_case<256, cols_per_block, float>(ctx, dst);
                     break;
                 default:
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
                     break;
             }
         } else {
@@ -63,7 +63,7 @@ static void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, g
                 //     ggml_cuda_flash_attn_ext_wmma_f16_case<128, cols_per_block, float>(ctx, dst);
                 //     break;
                 default:
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
                     break;
             }
         }
@@ -86,7 +86,7 @@ static void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, g
                 ggml_cuda_flash_attn_ext_wmma_f16_case<256, cols_per_block, half>(ctx, dst);
                 break;
             default:
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
                 break;
         }
         return;
@@ -114,7 +114,7 @@ static void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, g
                 ggml_cuda_flash_attn_ext_wmma_f16_case<256, cols_per_block, half>(ctx, dst);
                 break;
             default:
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
                 break;
         }
         return;
@@ -141,7 +141,7 @@ static void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, g
             ggml_cuda_flash_attn_ext_wmma_f16_case<256, cols_per_block, half>(ctx, dst);
             break;
         default:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
             break;
     }
 }
diff --git a/ggml/src/ggml-cuda/getrows.cu b/ggml/src/ggml-cuda/getrows.cu
index 55af195f..4c370323 100644
--- a/ggml/src/ggml-cuda/getrows.cu
+++ b/ggml/src/ggml-cuda/getrows.cu
@@ -171,8 +171,7 @@ void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
             break;
         default:
             // TODO: k-quants
-            fprintf(stderr, "%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
-            GGML_ASSERT(false);
+            GGML_ABORT("%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
             break;
     }
 }
diff --git a/ggml/src/ggml-cuda/iqk_mmvq.cu b/ggml/src/ggml-cuda/iqk_mmvq.cu
index 29721cdd..c567ad1a 100644
--- a/ggml/src/ggml-cuda/iqk_mmvq.cu
+++ b/ggml/src/ggml-cuda/iqk_mmvq.cu
@@ -466,7 +466,6 @@ __device__ __forceinline__ float vec_dot_iq3_k_q8_1(
 
     const int * q8;
     int sumi[4] = {0, 0, 0, 0};
-    uint16_t v1, v2;
     int v;
     for (int i = 0; i < 2; ++i) {
         uint32_t vl = ql[2*i+0] | (ql[2*i+1] << 16);
diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu
index 84f6387e..78d70cd7 100644
--- a/ggml/src/ggml-cuda/mmq.cu
+++ b/ggml/src/ggml-cuda/mmq.cu
@@ -84,7 +84,7 @@ void ggml_cuda_op_mul_mat_q(
             mul_mat_q_case<GGML_TYPE_IQ4_NL>(ctx, args, stream);
             break;
         default:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
             break;
     }
 
diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh
index f08a4758..e8a95744 100644
--- a/ggml/src/ggml-cuda/mmq.cuh
+++ b/ggml/src/ggml-cuda/mmq.cuh
@@ -75,7 +75,7 @@ static mmq_q8_1_ds_layout mmq_get_q8_1_ds_layout(const ggml_type type_x) {
         case GGML_TYPE_IQ4_NL:
             return MMQ_Q8_1_DS_LAYOUT_D4;
         default:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
             break;
     }
 }
@@ -2898,7 +2898,7 @@ void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cuda
             break;
         default:
             fprintf(stderr, "mmq_x_best=%d\n", mmq_x_best);
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
             break;
     }
 }
diff --git a/ggml/src/ggml-cuda/mmvq.cu b/ggml/src/ggml-cuda/mmvq.cu
index 9eb3fa4f..2586ab7e 100644
--- a/ggml/src/ggml-cuda/mmvq.cu
+++ b/ggml/src/ggml-cuda/mmvq.cu
@@ -165,7 +165,7 @@ static void mul_mat_vec_q_cuda(
                 rows_per_cuda_block = 2;
                 break;
             default:
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
                 break;
         }
     }
@@ -199,7 +199,7 @@ static void mul_mat_vec_q_cuda(
             mul_mat_vec_q<type, 8><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
             break;
         default:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
             break;
     }
 }
@@ -454,7 +454,7 @@ void ggml_cuda_op_mul_mat_vec_q(
             mul_mat_vec_iq3_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
             break;
         default:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
             break;
     }
 
diff --git a/ggml/src/ggml-cuda/norm.cu b/ggml/src/ggml-cuda/norm.cu
index 30866d51..133e219f 100644
--- a/ggml/src/ggml-cuda/norm.cu
+++ b/ggml/src/ggml-cuda/norm.cu
@@ -142,8 +142,7 @@ static void norm_f32_cuda(const float * x, float * dst, const int ncols, const i
     }
 }
 
-static void group_norm_f32_cuda(const float * x, float * dst, const int num_groups, const int group_size, const int ne_elements, cudaStream_t stream) {
-    static const float eps = 1e-6f;
+static void group_norm_f32_cuda(const float * x, float * dst, const int num_groups, const float eps, const int group_size, const int ne_elements, cudaStream_t stream) {
     if (group_size < 1024) {
         const dim3 block_dims(WARP_SIZE, 1, 1);
         group_norm_f32<WARP_SIZE><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
@@ -196,8 +195,12 @@ void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
     int num_groups = dst->op_params[0];
+
+    float eps;
+    memcpy(&eps, dst->op_params + 1, sizeof(float));
+
     int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
-    group_norm_f32_cuda(src0_d, dst_d, num_groups * src0->ne[3], group_size, ggml_nelements(src0), stream);
+    group_norm_f32_cuda(src0_d, dst_d, num_groups * src0->ne[3], eps, group_size, ggml_nelements(src0), stream);
 }
 
 void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
diff --git a/ggml/src/ggml-cuda/quantize.cu b/ggml/src/ggml-cuda/quantize.cu
index aa7f1eff..45408ce8 100644
--- a/ggml/src/ggml-cuda/quantize.cu
+++ b/ggml/src/ggml-cuda/quantize.cu
@@ -163,7 +163,7 @@ void quantize_mmq_q8_1_cuda(
                 <<<num_blocks, block_size, 0, stream>>>(x, vy, kx0, kx1, kx0_padded);
             break;
         default:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
             break;
     }
 }
diff --git a/ggml/src/ggml-cuda/rope.cu b/ggml/src/ggml-cuda/rope.cu
index 596fb7c1..99ec1dd9 100644
--- a/ggml/src/ggml-cuda/rope.cu
+++ b/ggml/src/ggml-cuda/rope.cu
@@ -251,7 +251,7 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
                 attn_factor, corr_dims, freq_factors, stream
             );
         } else {
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
     } else {
         if (src0->type == GGML_TYPE_F32) {
@@ -265,7 +265,7 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
                 attn_factor, corr_dims, freq_factors, stream
             );
         } else {
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
     }
 }
diff --git a/ggml/src/ggml-cuda/vendors/cuda.h b/ggml/src/ggml-cuda/vendors/cuda.h
new file mode 100644
index 00000000..db9f6a16
--- /dev/null
+++ b/ggml/src/ggml-cuda/vendors/cuda.h
@@ -0,0 +1,14 @@
+#pragma once
+
+#include <cuda_runtime.h>
+#include <cuda.h>
+#include <cublas_v2.h>
+#include <cuda_fp16.h>
+
+#if CUDART_VERSION < 11020
+#define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED
+#define CUBLAS_TF32_TENSOR_OP_MATH CUBLAS_TENSOR_OP_MATH
+#define CUBLAS_COMPUTE_16F CUDA_R_16F
+#define CUBLAS_COMPUTE_32F CUDA_R_32F
+#define cublasComputeType_t cudaDataType_t
+#endif // CUDART_VERSION < 11020
diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.h
new file mode 100644
index 00000000..d0c37725
--- /dev/null
+++ b/ggml/src/ggml-cuda/vendors/hip.h
@@ -0,0 +1,177 @@
+#pragma once
+
+#include <hip/hip_runtime.h>
+#include <hipblas/hipblas.h>
+#include <hip/hip_fp16.h>
+#ifdef __HIP_PLATFORM_AMD__
+// for rocblas_initialize()
+#include "rocblas/rocblas.h"
+#endif // __HIP_PLATFORM_AMD__
+#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F
+#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
+#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
+#define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
+#define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
+#define CUBLAS_OP_N HIPBLAS_OP_N
+#define CUBLAS_OP_T HIPBLAS_OP_T
+#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
+#define CUBLAS_TF32_TENSOR_OP_MATH 0
+#define CUDA_R_16F  HIPBLAS_R_16F
+#define CUDA_R_32F  HIPBLAS_R_32F
+#define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
+#define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6
+#define cublasCreate hipblasCreate
+#define cublasDestroy hipblasDestroy
+#define cublasGemmEx hipblasGemmEx
+#define cublasGemmBatchedEx hipblasGemmBatchedEx
+#define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx
+#define cublasHandle_t hipblasHandle_t
+#define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS
+#define cublasSetStream hipblasSetStream
+#define cublasSgemm hipblasSgemm
+#define cublasStatus_t hipblasStatus_t
+#define cudaDataType_t hipblasDatatype_t //deprecated, new hipblasDatatype not in 5.6
+#define cudaDeviceCanAccessPeer hipDeviceCanAccessPeer
+#define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
+#define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
+#define cudaDeviceProp hipDeviceProp_t
+#define cudaDeviceSynchronize hipDeviceSynchronize
+#define cudaError_t hipError_t
+#define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled
+#define cudaErrorPeerAccessNotEnabled hipErrorPeerAccessNotEnabled
+#define cudaEventCreateWithFlags hipEventCreateWithFlags
+#define cudaEventDisableTiming hipEventDisableTiming
+#define cudaEventRecord hipEventRecord
+#define cudaEventSynchronize hipEventSynchronize
+#define cudaEvent_t hipEvent_t
+#define cudaEventDestroy hipEventDestroy
+#define cudaFree hipFree
+#define cudaFreeHost hipHostFree
+#define cudaGetDevice hipGetDevice
+#define cudaGetDeviceCount hipGetDeviceCount
+#define cudaGetDeviceProperties hipGetDeviceProperties
+#define cudaGetErrorString hipGetErrorString
+#define cudaGetLastError hipGetLastError
+#define cudaHostRegister hipHostRegister
+#define cudaHostRegisterPortable hipHostRegisterPortable
+#define cudaHostRegisterReadOnly hipHostRegisterReadOnly
+#define cudaHostUnregister hipHostUnregister
+#define cudaLaunchHostFunc hipLaunchHostFunc
+#define cudaMalloc hipMalloc
+#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
+#define cudaMemcpy hipMemcpy
+#define cudaMemcpyAsync hipMemcpyAsync
+#define cudaMemcpyPeerAsync hipMemcpyPeerAsync
+#define cudaMemcpy2DAsync hipMemcpy2DAsync
+#define cudaMemcpyDeviceToDevice hipMemcpyDeviceToDevice
+#define cudaMemcpyDeviceToHost hipMemcpyDeviceToHost
+#define cudaMemcpyHostToDevice hipMemcpyHostToDevice
+#define cudaMemcpyKind hipMemcpyKind
+#define cudaMemset hipMemset
+#define cudaMemsetAsync hipMemsetAsync
+#define cudaMemGetInfo hipMemGetInfo
+#define cudaOccupancyMaxPotentialBlockSize hipOccupancyMaxPotentialBlockSize
+#define cudaSetDevice hipSetDevice
+#define cudaStreamCreateWithFlags hipStreamCreateWithFlags
+#define cudaStreamDestroy hipStreamDestroy
+#define cudaStreamFireAndForget hipStreamFireAndForget
+#define cudaStreamNonBlocking hipStreamNonBlocking
+#define cudaStreamPerThread hipStreamPerThread
+#define cudaStreamSynchronize hipStreamSynchronize
+#define cudaStreamWaitEvent(stream, event, flags) hipStreamWaitEvent(stream, event, flags)
+#define cudaStream_t hipStream_t
+#define cudaSuccess hipSuccess
+#define __trap() do { abort(); __builtin_unreachable(); } while(0)
+#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
+#define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED
+#define CUBLAS_STATUS_ALLOC_FAILED HIPBLAS_STATUS_ALLOC_FAILED
+#define CUBLAS_STATUS_INVALID_VALUE HIPBLAS_STATUS_INVALID_VALUE
+#define CUBLAS_STATUS_ARCH_MISMATCH HIPBLAS_STATUS_ARCH_MISMATCH
+#define CUBLAS_STATUS_MAPPING_ERROR HIPBLAS_STATUS_MAPPING_ERROR
+#define CUBLAS_STATUS_EXECUTION_FAILED HIPBLAS_STATUS_EXECUTION_FAILED
+#define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR
+#define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED
+
+#define __CUDA_ARCH__ 1300
+
+#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
+    defined(__gfx1150__) || defined(__gfx1151__)
+#define RDNA3
+#endif
+
+#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || defined(__gfx1033__) || \
+    defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || defined(__gfx1037__)
+#define RDNA2
+#endif
+
+#if defined(__gfx1010__) || defined(__gfx1012__)
+#define RDNA1
+#endif
+
+#ifndef __has_builtin
+    #define __has_builtin(x) 0
+#endif
+
+typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
+typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
+static __device__ __forceinline__ int __vsubss4(const int a, const int b) {
+    const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
+    const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
+#if __has_builtin(__builtin_elementwise_sub_sat)
+    const int8x4_t c = __builtin_elementwise_sub_sat(va, vb);
+    return reinterpret_cast<const int &>(c);
+#else
+    int8x4_t c;
+    int16_t tmp;
+#pragma unroll
+    for (int i = 0; i < 4; i++) {
+        tmp = va[i] - vb[i];
+        if(tmp > std::numeric_limits<int8_t>::max()) tmp = std::numeric_limits<int8_t>::max();
+        if(tmp < std::numeric_limits<int8_t>::min()) tmp = std::numeric_limits<int8_t>::min();
+        c[i] = tmp;
+    }
+    return reinterpret_cast<int &>(c);
+#endif // __has_builtin(__builtin_elementwise_sub_sat)
+}
+
+static __device__ __forceinline__ int __vsub4(const int a, const int b) {
+    return __vsubss4(a, b);
+}
+
+static __device__ __forceinline__ unsigned int __vcmpeq4(unsigned int a, unsigned int b) {
+    const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
+    const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
+    unsigned int c;
+    uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
+#pragma unroll
+    for (int i = 0; i < 4; ++i) {
+        vc[i] = va[i] == vb[i] ? 0xff : 0x00;
+    }
+    return c;
+}
+
+static __device__ __forceinline__ unsigned int __vcmpne4(unsigned int a, unsigned int b) {
+    const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
+    const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
+    unsigned int c;
+    uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
+#pragma unroll
+    for (int i = 0; i < 4; ++i) {
+        vc[i] = va[i] == vb[i] ? 0x00 : 0xff;
+    }
+    return c;
+}
+
+#if defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
+// __shfl_xor() for half2 was added in ROCm 5.6
+static __device__ __forceinline__ half2 __shfl_xor(half2 var, int laneMask, int width) {
+    typedef union half2_b32 {
+        half2 val;
+        int   b32;
+    } half2_b32_t;
+    half2_b32_t tmp;
+    tmp.val = var;
+    tmp.b32 = __shfl_xor(tmp.b32, laneMask, width);
+    return tmp.val;
+}
+#endif // defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
diff --git a/ggml/src/ggml-cuda/vendors/musa.h b/ggml/src/ggml-cuda/vendors/musa.h
new file mode 100644
index 00000000..e50a103a
--- /dev/null
+++ b/ggml/src/ggml-cuda/vendors/musa.h
@@ -0,0 +1,171 @@
+#pragma once
+
+#include <musa_runtime.h>
+#include <musa.h>
+#include <mublas.h>
+#include <musa_fp16.h>
+#define CUBLAS_COMPUTE_16F CUDA_R_16F
+#define CUBLAS_COMPUTE_32F CUDA_R_32F
+#define CUBLAS_COMPUTE_32F_FAST_16F MUBLAS_COMPUTE_32F_FAST_16F
+#define CUBLAS_GEMM_DEFAULT MUBLAS_GEMM_DEFAULT
+#define CUBLAS_GEMM_DEFAULT_TENSOR_OP MUBLAS_GEMM_DEFAULT
+#define CUBLAS_OP_N MUBLAS_OP_N
+#define CUBLAS_OP_T MUBLAS_OP_T
+#define CUBLAS_STATUS_SUCCESS MUBLAS_STATUS_SUCCESS
+#define CUBLAS_TF32_TENSOR_OP_MATH MUBLAS_MATH_MODE_DEFAULT
+#define CUDA_R_16F  MUSA_R_16F
+#define CUDA_R_32F  MUSA_R_32F
+#define cublasComputeType_t cudaDataType_t
+#define cublasCreate mublasCreate
+#define cublasDestroy mublasDestroy
+#define cublasGemmEx mublasGemmEx
+#define cublasGemmBatchedEx mublasGemmBatchedEx
+#define cublasGemmStridedBatchedEx mublasGemmStridedBatchedEx
+#define cublasHandle_t mublasHandle_t
+#define cublasSetMathMode mublasSetMathMode
+#define cublasSetStream mublasSetStream
+#define cublasSgemm mublasSgemm
+#define cublasStatus_t mublasStatus_t
+#define cublasGetStatusString mublasStatus_to_string
+#define cudaDataType_t musaDataType_t
+#define cudaDeviceCanAccessPeer musaDeviceCanAccessPeer
+#define cudaDeviceDisablePeerAccess musaDeviceDisablePeerAccess
+#define cudaDeviceEnablePeerAccess musaDeviceEnablePeerAccess
+#define cudaDeviceProp musaDeviceProp
+#define cudaDeviceSynchronize musaDeviceSynchronize
+#define cudaError_t musaError_t
+#define cudaErrorPeerAccessAlreadyEnabled musaErrorPeerAccessAlreadyEnabled
+#define cudaErrorPeerAccessNotEnabled musaErrorPeerAccessNotEnabled
+#define cudaEventCreateWithFlags musaEventCreateWithFlags
+#define cudaEventDisableTiming musaEventDisableTiming
+#define cudaEventRecord musaEventRecord
+#define cudaEventSynchronize musaEventSynchronize
+#define cudaEvent_t musaEvent_t
+#define cudaEventDestroy musaEventDestroy
+#define cudaFree musaFree
+#define cudaFreeHost musaFreeHost
+#define cudaGetDevice musaGetDevice
+#define cudaGetDeviceCount musaGetDeviceCount
+#define cudaGetDeviceProperties musaGetDeviceProperties
+#define cudaGetErrorString musaGetErrorString
+#define cudaGetLastError musaGetLastError
+#define cudaHostRegister musaHostRegister
+#define cudaHostRegisterPortable musaHostRegisterPortable
+#define cudaHostRegisterReadOnly musaHostRegisterReadOnly
+#define cudaHostUnregister musaHostUnregister
+#define cudaLaunchHostFunc musaLaunchHostFunc
+#define cudaMalloc musaMalloc
+#define cudaMallocHost musaMallocHost
+#define cudaMemcpy musaMemcpy
+#define cudaMemcpyAsync musaMemcpyAsync
+#define cudaMemcpyPeerAsync musaMemcpyPeerAsync
+#define cudaMemcpy2DAsync musaMemcpy2DAsync
+#define cudaMemcpyDeviceToDevice musaMemcpyDeviceToDevice
+#define cudaMemcpyDeviceToHost musaMemcpyDeviceToHost
+#define cudaMemcpyHostToDevice musaMemcpyHostToDevice
+#define cudaMemcpyKind musaMemcpyKind
+#define cudaMemset musaMemset
+#define cudaMemsetAsync musaMemsetAsync
+#define cudaMemGetInfo musaMemGetInfo
+#define cudaOccupancyMaxPotentialBlockSize musaOccupancyMaxPotentialBlockSize
+#define cudaSetDevice musaSetDevice
+#define cudaStreamCreateWithFlags musaStreamCreateWithFlags
+#define cudaStreamDestroy musaStreamDestroy
+#define cudaStreamFireAndForget musaStreamFireAndForget
+#define cudaStreamNonBlocking musaStreamNonBlocking
+#define cudaStreamPerThread musaStreamPerThread
+#define cudaStreamSynchronize musaStreamSynchronize
+#define cudaStreamWaitEvent musaStreamWaitEvent
+#define cudaStream_t musaStream_t
+#define cudaSuccess musaSuccess
+
+// Additional mappings for MUSA virtual memory pool
+#define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED MU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED
+#define CU_MEM_ACCESS_FLAGS_PROT_READWRITE MU_MEM_ACCESS_FLAGS_PROT_READWRITE
+#define CU_MEM_ALLOC_GRANULARITY_RECOMMENDED MU_MEM_ALLOC_GRANULARITY_RECOMMENDED
+#define CU_MEM_ALLOCATION_TYPE_PINNED MU_MEM_ALLOCATION_TYPE_PINNED
+#define CU_MEM_LOCATION_TYPE_DEVICE MU_MEM_LOCATION_TYPE_DEVICE
+#define CUdevice MUdevice
+#define CUdeviceptr MUdeviceptr
+#define CUmemAccessDesc MUmemAccessDesc
+#define CUmemAllocationProp MUmemAllocationProp
+#define CUmemGenericAllocationHandle MUmemGenericAllocationHandle
+#define cuDeviceGet muDeviceGet
+#define cuDeviceGetAttribute muDeviceGetAttribute
+#define cuMemAddressFree muMemAddressFree
+#define cuMemAddressReserve muMemAddressReserve
+#define cuMemCreate muMemCreate
+#define cuMemGetAllocationGranularity muMemGetAllocationGranularity
+#define cuMemMap muMemMap
+#define cuMemRelease muMemRelease
+#define cuMemSetAccess muMemSetAccess
+#define cuMemUnmap muMemUnmap
+#define cudaFuncAttributeMaxDynamicSharedMemorySize musaFuncAttributeMaxDynamicSharedMemorySize
+#define cudaFuncSetAttribute musaFuncSetAttribute
+#define cudaMemcpy3DPeerParms musaMemcpy3DPeerParms
+#define make_cudaExtent make_musaExtent
+#define make_cudaPitchedPtr make_musaPitchedPtr
+
+// Additional mappings for MUSA graphs
+#define CUDA_SUCCESS MUSA_SUCCESS
+#define CUresult MUresult
+#define cuGetErrorString muGetErrorString
+#define cudaErrorGraphExecUpdateFailure musaErrorGraphExecUpdateFailure
+#define cudaErrorInvalidDeviceFunction musaErrorInvalidDeviceFunction
+#define cudaGraphDestroy musaGraphDestroy
+#define cudaGraphExecDestroy musaGraphExecDestroy
+#define cudaGraphExec_t musaGraphExec_t
+#define cudaGraphExecUpdate musaGraphExecUpdate
+#define cudaGraphExecUpdateResultInfo musaGraphExecUpdateResult
+#define cudaGraphGetNodes musaGraphGetNodes
+#define cudaGraphInstantiate musaGraphInstantiate
+#define cudaGraphKernelNodeGetParams musaGraphKernelNodeGetParams
+#define cudaGraphKernelNodeSetParams musaGraphKernelNodeSetParams
+#define cudaGraphLaunch musaGraphLaunch
+#define cudaGraphNodeGetType musaGraphNodeGetType
+#define cudaGraphNode_t musaGraphNode_t
+#define cudaGraphNodeType musaGraphNodeType
+#define cudaGraphNodeTypeKernel musaGraphNodeTypeKernel
+#define cudaGraph_t musaGraph_t
+#define cudaKernelNodeParams musaKernelNodeParams
+#define cudaStreamCaptureModeRelaxed musaStreamCaptureModeRelaxed
+#define cudaStreamEndCapture musaStreamEndCapture
+
+// XXX: Clang builtins mapping
+#define __vsub4   __vsub4_musa
+#define __vcmpeq4 __vcmpeq4_musa
+#define __vcmpne4 __vcmpne4_musa
+
+#ifndef __has_builtin
+    #define __has_builtin(x) 0
+#endif
+
+typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
+
+static __device__ __forceinline__ int __vsub4_musa(const int a, const int b) {
+    return __vsubss4(a, b);
+}
+
+static __device__ __forceinline__ unsigned int __vcmpeq4_musa(unsigned int a, unsigned int b) {
+    const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
+    const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
+    unsigned int c;
+    uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
+#pragma unroll
+    for (int i = 0; i < 4; ++i) {
+        vc[i] = va[i] == vb[i] ? 0xff : 0x00;
+    }
+    return c;
+}
+
+static __device__ __forceinline__ unsigned int __vcmpne4_musa(unsigned int a, unsigned int b) {
+    const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
+    const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
+    unsigned int c;
+    uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
+#pragma unroll
+    for (int i = 0; i < 4; ++i) {
+        vc[i] = va[i] == vb[i] ? 0x00 : 0xff;
+    }
+    return c;
+}
diff --git a/ggml/src/ggml-impl.h b/ggml/src/ggml-impl.h
index a2c8dbec..190af081 100644
--- a/ggml/src/ggml-impl.h
+++ b/ggml/src/ggml-impl.h
@@ -80,8 +80,9 @@ static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
 /**
  * Converts float32 to brain16.
  *
- * This function is binary identical to AMD Zen4 VCVTNEPS2BF16.
- * Subnormals shall be flushed to zero, and NANs will be quiet.
+ * This is binary identical with Google Brain float conversion.
+ * Floats shall round to nearest even, and NANs shall be quiet.
+ * Subnormals aren't flushed to zero, except perhaps when used.
  * This code should vectorize nicely if using modern compilers.
  */
 static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
@@ -95,10 +96,6 @@ static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
         h.bits = (u.i >> 16) | 64; /* force to quiet */
         return h;
     }
-    if (!(u.i & 0x7f800000)) { /* subnormal */
-        h.bits = (u.i & 0x80000000) >> 16; /* flush to zero */
-        return h;
-    }
     h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
     return h;
 }
@@ -146,6 +143,7 @@ extern "C" {
 
 #if defined(__ARM_FEATURE_SVE)
 #include <arm_sve.h>
+#include <sys/prctl.h>
 #endif
 
 // 16-bit float
@@ -634,21 +632,121 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
 #endif
 
-#define GGML_HASHTABLE_FULL ((size_t)-1)
-#define GGML_HASHTABLE_ALREADY_EXISTS ((size_t)-2)
+// bitset
+
+static_assert(sizeof(ggml_bitset_t) == 4, "bitset_t constants must be updated");
+#define BITSET_SHR 5 // log2(sizeof(ggml_bitset_t)*8)
+#define BITSET_MASK (sizeof(ggml_bitset_t)*8 - 1)
+
+static size_t ggml_bitset_size(size_t n) {
+    return (n + BITSET_MASK) >> BITSET_SHR;
+}
+
+static inline bool ggml_bitset_get(const ggml_bitset_t * bitset, size_t i) {
+    return !!(bitset[i >> BITSET_SHR] & (1u << (i & BITSET_MASK)));
+}
+
+static inline void ggml_bitset_set(ggml_bitset_t * bitset, size_t i) {
+    bitset[i >> BITSET_SHR] |= (1u << (i & BITSET_MASK));
+}
+
+static inline void ggml_bitset_clear(ggml_bitset_t * bitset, size_t i) {
+    bitset[i >> BITSET_SHR] &= ~(1u << (i & BITSET_MASK));
+}
+
+// hash set
+
+#define GGML_HASHSET_FULL ((size_t)-1)
+#define GGML_HASHSET_ALREADY_EXISTS ((size_t)-2)
 
 struct ggml_hash_set ggml_hash_set_new(size_t size);
+void                 ggml_hash_set_free(struct ggml_hash_set * hash_set);
 
-bool   ggml_hash_contains      (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
+// returns the minimum size for a hash set that can hold min_sz elements
+size_t ggml_hash_size(size_t min_sz);
 
-// returns GGML_HASHTABLE_FULL if table is full, otherwise the current index of the key or where it should be inserted
-size_t ggml_hash_find          (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
+// remove all elements from the hash set
+void ggml_hash_set_reset(struct ggml_hash_set * hash_set);
 
-// returns GGML_HASHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
-size_t ggml_hash_insert        (      struct ggml_hash_set hash_set, struct ggml_tensor * key);
+// returns true if key is in the hash set
+static bool ggml_hash_contains(const struct ggml_hash_set * hash_set, struct ggml_tensor * key);
+
+// returns GGML_HASHSET_FULL if table is full, otherwise the current index of the key or where it should be inserted
+static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, struct ggml_tensor * key);
+
+// returns GGML_HASHSET_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
+static size_t ggml_hash_insert(struct ggml_hash_set * hash_set, struct ggml_tensor * key);
 
 // return index, asserts if table is full
-size_t ggml_hash_find_or_insert(      struct ggml_hash_set hash_set, struct ggml_tensor * key);
+static size_t ggml_hash_find_or_insert(struct ggml_hash_set * hash_set, struct ggml_tensor * key);
+
+// hash function for ggml_tensor
+static inline size_t ggml_hash(const struct ggml_tensor * p) {
+    // the last 4 bits are always zero due to alignment
+    return (size_t)(uintptr_t)p >> 4;
+}
+
+static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, struct ggml_tensor * key) {
+    size_t h = ggml_hash(key) % hash_set->size;
+
+    // linear probing
+    size_t i = h;
+    while (ggml_bitset_get(hash_set->used, i) && hash_set->keys[i] != key) {
+        i = (i + 1) % hash_set->size;
+        if (i == h) {
+            // visited all hash table entries -> not found
+            return GGML_HASHSET_FULL;
+        }
+    }
+    return i;
+}
+
+static bool ggml_hash_contains(const struct ggml_hash_set * hash_set, struct ggml_tensor * key) {
+    size_t i = ggml_hash_find(hash_set, key);
+    return i != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, i);
+}
+
+static size_t ggml_hash_insert(struct ggml_hash_set * hash_set, struct ggml_tensor * key) {
+    size_t h = ggml_hash(key) % hash_set->size;
+
+    // linear probing
+    size_t i = h;
+    do {
+        if (!ggml_bitset_get(hash_set->used, i)) {
+            ggml_bitset_set(hash_set->used, i);
+            hash_set->keys[i] = key;
+            return i;
+        }
+        if (hash_set->keys[i] == key) {
+            return GGML_HASHSET_ALREADY_EXISTS;
+        }
+        i = (i + 1) % hash_set->size;
+    } while (i != h);
+
+    // visited all hash table entries -> not found
+    GGML_ABORT("fatal error");
+}
+
+static size_t ggml_hash_find_or_insert(struct ggml_hash_set * hash_set, struct ggml_tensor * key) {
+    size_t h = ggml_hash(key) % hash_set->size;
+
+    // linear probing
+    size_t i = h;
+    do {
+        if (!ggml_bitset_get(hash_set->used, i)) {
+            ggml_bitset_set(hash_set->used, i);
+            hash_set->keys[i] = key;
+            return i;
+        }
+        if (hash_set->keys[i] == key) {
+            return i;
+        }
+        i = (i + 1) % hash_set->size;
+    } while (i != h);
+
+    // visited all hash table entries -> not found
+    GGML_ABORT("fatal error");
+}
 
 #ifdef __cplusplus
 }
diff --git a/ggml/src/ggml-kompute.cpp b/ggml/src/ggml-kompute.cpp
index ed5f2e34..41ac63fa 100644
--- a/ggml/src/ggml-kompute.cpp
+++ b/ggml/src/ggml-kompute.cpp
@@ -566,7 +566,7 @@ uint32_t safe_divide(uint32_t a, uint32_t b) {
     }
     if ((a % b) != 0) {
         fprintf(stderr, "((%u %% %u) == %u) != 0\n", a, b, a % b);
-        GGML_ASSERT(!"safe_divide result would've had remainder");
+        GGML_ABORT("safe_divide result would've had remainder");
     }
     return a / b;
 }
@@ -1460,7 +1460,7 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
 
             if (!ggml_vk_supports_op(dst)) {
                  fprintf(stderr, "%s: error: unsupported op '%s'\n", __func__, ggml_op_desc(dst));
-                 GGML_ASSERT(!"unsupported op");
+                 GGML_ABORT("unsupported op");
             }
 
             const int32_t ne00 = src0 ? src0->ne[0] : 0;
@@ -1562,7 +1562,7 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
                             default:
                                 {
                                     fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
-                                    GGML_ASSERT(false);
+                                    GGML_ABORT("fatal error");
                                 }
                         }
                     } break;
@@ -1745,7 +1745,7 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
             continue;
             not_implemented: {}
             fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
-            //GGML_ASSERT(false);
+            //GGML_ABORT("fatal error");
         }
 
         // Evaluate sequence
diff --git a/ggml/src/ggml-metal.m b/ggml/src/ggml-metal.m
index 7d592c22..292f9ac7 100644
--- a/ggml/src/ggml-metal.m
+++ b/ggml/src/ggml-metal.m
@@ -260,7 +260,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_COUNT
 };
 
-struct ggml_metal_context {
+struct ggml_backend_metal_context {
     int n_cb;
 
     id<MTLDevice>       device;
@@ -274,6 +274,10 @@ struct ggml_metal_context {
     bool support_simdgroup_mm;
 
     bool should_capture_next_compute;
+
+    // abort ggml_metal_graph_compute if callback returns true
+    ggml_abort_callback abort_callback;
+    void *              abort_callback_data;
 };
 
 // MSL code
@@ -339,7 +343,7 @@ static void * ggml_metal_host_malloc(size_t n) {
     return data;
 }
 
-static struct ggml_metal_context * ggml_metal_init(int n_cb) {
+static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) {
     GGML_METAL_LOG_INFO("%s: allocating\n", __func__);
 
 #if TARGET_OS_OSX && !GGML_METAL_NDEBUG
@@ -356,7 +360,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
     GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]);
 
     // Configure context
-    struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
+    struct ggml_backend_metal_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_context));
     ctx->device = device;
     ctx->n_cb   = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
     ctx->queue  = [ctx->device newCommandQueue];
@@ -761,7 +765,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
     return ctx;
 }
 
-static void ggml_metal_free(struct ggml_metal_context * ctx) {
+static void ggml_metal_free(struct ggml_backend_metal_context * ctx) {
     GGML_METAL_LOG_INFO("%s: deallocating\n", __func__);
 
     for (int i = 0; i < GGML_METAL_KERNEL_TYPE_COUNT; ++i) {
@@ -827,7 +831,7 @@ static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_tensor * t, size_t * offs
     return nil;
 }
 
-static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const struct ggml_tensor * op) {
+static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx, const struct ggml_tensor * op) {
     for (size_t i = 0, n = 3; i < n; ++i) {
         if (op->src[i] != NULL && op->src[i]->type == GGML_TYPE_BF16) {
             return false;
@@ -938,7 +942,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
 }
 
 static enum ggml_status ggml_metal_graph_compute(
-        struct ggml_metal_context * ctx,
+        struct ggml_backend_metal_context * ctx,
                struct ggml_cgraph * gf) {
 
     @autoreleasepool {
@@ -962,7 +966,7 @@ static enum ggml_status ggml_metal_graph_compute(
         NSError * error = nil;
         if (![[MTLCaptureManager sharedCaptureManager] startCaptureWithDescriptor:descriptor error:&error]) {
             GGML_METAL_LOG_ERROR("%s: error: unable to start capture '%s'\n", __func__, [[error localizedDescription] UTF8String]);
-            GGML_ASSERT(!"capture failed");
+            GGML_ABORT("capture failed");
         }
     }
 
@@ -971,8 +975,11 @@ static enum ggml_status ggml_metal_graph_compute(
         id<MTLCommandBuffer> command_buffer  = [ctx->queue commandBufferWithUnretainedReferences];
         command_buffer_builder[cb_idx] = command_buffer;
 
-        // enqueue the command buffers in order to specify their execution order
-        [command_buffer enqueue];
+        // always enqueue the first two command buffers
+        // enqueue all of the command buffers if we don't need to abort
+        if (cb_idx < 2 || ctx->abort_callback == NULL) {
+            [command_buffer enqueue];
+        }
     }
 
     const id<MTLCommandBuffer> *command_buffers = command_buffer_builder;
@@ -1024,7 +1031,7 @@ static enum ggml_status ggml_metal_graph_compute(
 
             if (!ggml_metal_supports_op(ctx, dst)) {
                 GGML_METAL_LOG_ERROR("%s: error: unsupported op '%s'\n", __func__, ggml_op_desc(dst));
-                GGML_ASSERT(!"unsupported op");
+                GGML_ABORT("unsupported op");
             }
 
             if (should_capture) {
@@ -1207,7 +1214,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                 case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ROW].pipeline; break;
                                 case GGML_OP_MUL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_ROW].pipeline; break;
                                 case GGML_OP_DIV: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIV_ROW].pipeline; break;
-                                default: GGML_ASSERT(false);
+                                default: GGML_ABORT("fatal error");
                             }
 
                             bcast_row = true;
@@ -1216,7 +1223,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                 case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD].pipeline; break;
                                 case GGML_OP_MUL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL].pipeline; break;
                                 case GGML_OP_DIV: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIV].pipeline; break;
-                                default: GGML_ASSERT(false);
+                                default: GGML_ABORT("fatal error");
                             }
                         }
 
@@ -1270,7 +1277,7 @@ static enum ggml_status ggml_metal_graph_compute(
                             case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_F16].pipeline; break;
                             case GGML_TYPE_I32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_I32].pipeline; break;
                             case GGML_TYPE_I16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_I16].pipeline; break;
-                            default: GGML_ASSERT(false);
+                            default: GGML_ABORT("fatal error");
                         }
 
                         [encoder setComputePipelineState:pipeline];
@@ -1526,7 +1533,7 @@ static enum ggml_status ggml_metal_graph_compute(
                         default:
                             {
                                 GGML_METAL_LOG_WARN("%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
-                                GGML_ASSERT(false);
+                                GGML_ABORT("fatal error");
                             }
                     } break;
                 case GGML_OP_SQR:
@@ -1756,7 +1763,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                 case GGML_TYPE_IQ4_K:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_K_F32  ].pipeline; break;
                                 case GGML_TYPE_IQ5_K:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ5_K_F32  ].pipeline; break;
                                 case GGML_TYPE_IQ6_K:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ6_K_F32  ].pipeline; break;
-                                default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
+                                default: GGML_ABORT("MUL MAT-MAT not implemented");
                             }
 
                             [encoder setComputePipelineState:pipeline];
@@ -1977,7 +1984,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                 default:
                                     {
                                         GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
-                                        GGML_ASSERT(false && "not implemented");
+                                        GGML_ABORT("not implemented");
                                     }
                             };
 
@@ -2117,7 +2124,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                 case GGML_TYPE_IQ4_K:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_K_F32  ].pipeline; break;
                                 case GGML_TYPE_IQ5_K:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ5_K_F32  ].pipeline; break;
                                 case GGML_TYPE_IQ6_K:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ6_K_F32  ].pipeline; break;
-                                default: GGML_ASSERT(false && "MUL_MAT_ID not implemented");
+                                default: GGML_ABORT("MUL_MAT_ID not implemented");
                             }
 
                             [encoder setComputePipelineState:pipeline];
@@ -2332,7 +2339,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                 default:
                                     {
                                         GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src2t);
-                                        GGML_ASSERT(false && "not implemented");
+                                        GGML_ABORT("not implemented");
                                     }
                             };
 
@@ -2443,7 +2450,7 @@ static enum ggml_status ggml_metal_graph_compute(
                             case GGML_TYPE_IQ5_K:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ5_K  ].pipeline; break;
                             case GGML_TYPE_IQ6_K:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ6_K  ].pipeline; break;
                             case GGML_TYPE_I32:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_I32    ].pipeline; break;
-                            default: GGML_ASSERT(false && "not implemented");
+                            default: GGML_ABORT("not implemented");
                         }
 
                         [encoder setComputePipelineState:pipeline];
@@ -2494,10 +2501,8 @@ static enum ggml_status ggml_metal_graph_compute(
                         GGML_ASSERT(ne00 % 4 == 0);
                         GGML_ASSERT(ggml_is_contiguous(src0));
 
-                        //float eps;
-                        //memcpy(&eps, dst->op_params, sizeof(float));
-
-                        const float eps = 1e-6f; // TODO: temporarily hardcoded
+                        float eps;
+                        memcpy(&eps, dst->op_params + 1, sizeof(float));
 
                         const int32_t n_groups = ((int32_t *) dst->op_params)[0];
 
@@ -2581,13 +2586,13 @@ static enum ggml_status ggml_metal_graph_compute(
                             switch (src0->type) {
                                 case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32].pipeline; break;
                                 case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16].pipeline; break;
-                                default: GGML_ASSERT(false);
+                                default: GGML_ABORT("fatal error");
                             };
                         } else {
                             switch (src0->type) {
                                 case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32].pipeline; break;
                                 case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16].pipeline; break;
-                                default: GGML_ASSERT(false);
+                                default: GGML_ABORT("fatal error");
                             };
                         }
 
@@ -2664,7 +2669,7 @@ static enum ggml_status ggml_metal_graph_compute(
                         switch (dst->type) {
                             case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F32].pipeline; break;
                             case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F16].pipeline; break;
-                            default: GGML_ASSERT(false);
+                            default: GGML_ABORT("fatal error");
                         };
 
                         [encoder setComputePipelineState:pipeline];
@@ -2821,7 +2826,7 @@ static enum ggml_status ggml_metal_graph_compute(
                         switch (order) {
                             case GGML_SORT_ORDER_ASC:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC].pipeline;  break;
                             case GGML_SORT_ORDER_DESC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC].pipeline; break;
-                            default: GGML_ASSERT(false);
+                            default: GGML_ABORT("fatal error");
                         };
 
                         [encoder setComputePipelineState:pipeline];
@@ -2910,7 +2915,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                           {
                                               GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00);
                                               GGML_METAL_LOG_ERROR("add template specialization for this size\n");
-                                              GGML_ASSERT(false && "add template specialization for this size");
+                                              GGML_ABORT("add template specialization for this size");
                                           }
                             }
                         } else {
@@ -2923,7 +2928,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                           {
                                               GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00);
                                               GGML_METAL_LOG_ERROR("add template specialization for this size\n");
-                                              GGML_ASSERT(false && "add template specialization for this size");
+                                              GGML_ABORT("add template specialization for this size");
                                           }
                             }
                         }
@@ -3044,7 +3049,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                         case GGML_TYPE_Q5_0:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0].pipeline; break;
                                         case GGML_TYPE_Q5_1:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1].pipeline; break;
                                         case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL].pipeline; break;
-                                        default: GGML_ASSERT(false && "not implemented");
+                                        default: GGML_ABORT("not implemented");
                                     };
                                 } break;
                             case GGML_TYPE_F16:
@@ -3052,10 +3057,10 @@ static enum ggml_status ggml_metal_graph_compute(
                                     switch (dstt) {
                                         case GGML_TYPE_F32:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F16_F32].pipeline; break;
                                         case GGML_TYPE_F16:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F16_F16].pipeline; break;
-                                        default: GGML_ASSERT(false && "not implemented");
+                                        default: GGML_ABORT("not implemented");
                                     };
                                 } break;
-                            default: GGML_ASSERT(false && "not implemented");
+                            default: GGML_ABORT("not implemented");
                         }
 
                         [encoder setComputePipelineState:pipeline];
@@ -3083,7 +3088,7 @@ static enum ggml_status ggml_metal_graph_compute(
                 default:
                     {
                         GGML_METAL_LOG_ERROR("%s: error: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
-                        GGML_ASSERT(false);
+                        GGML_ABORT("fatal error");
                     }
             }
 
@@ -3094,7 +3099,9 @@ static enum ggml_status ggml_metal_graph_compute(
 
         [encoder endEncoding];
 
-        [command_buffer commit];
+        if (cb_idx < 2 || ctx->abort_callback == NULL) {
+            [command_buffer commit];
+        }
     });
 
     // Wait for completion and check status of each command buffer
@@ -3114,6 +3121,23 @@ static enum ggml_status ggml_metal_graph_compute(
 
             return GGML_STATUS_FAILED;
         }
+
+        id<MTLCommandBuffer> next_buffer = (i + 1 < n_cb ? command_buffers[i + 1] : nil);
+        if (!next_buffer) {
+            continue;
+        }
+
+        bool next_queued = ([next_buffer status] != MTLCommandBufferStatusNotEnqueued);
+        if (next_queued) {
+            continue;
+        }
+
+        if (ctx->abort_callback && ctx->abort_callback(ctx->abort_callback_data)) {
+            GGML_METAL_LOG_INFO("%s: command buffer %d aborted", __func__, i);
+            return GGML_STATUS_ABORTED;
+        }
+
+        [next_buffer commit];
     }
 
     if (should_capture) {
@@ -3417,7 +3441,7 @@ GGML_CALL static const char * ggml_backend_metal_name(ggml_backend_t backend) {
 }
 
 GGML_CALL static void ggml_backend_metal_free(ggml_backend_t backend) {
-    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
+    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
     ggml_metal_free(ctx);
     free(backend);
 }
@@ -3429,13 +3453,13 @@ GGML_CALL static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffe
 }
 
 GGML_CALL static enum ggml_status ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
+    struct ggml_backend_metal_context * metal_ctx = (struct ggml_backend_metal_context *)backend->context;
 
     return ggml_metal_graph_compute(metal_ctx, cgraph);
 }
 
 GGML_CALL static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
+    struct ggml_backend_metal_context * metal_ctx = (struct ggml_backend_metal_context *)backend->context;
 
     return ggml_metal_supports_op(metal_ctx, op);
 }
@@ -3480,9 +3504,9 @@ static ggml_guid_t ggml_backend_metal_guid(void) {
 }
 
 ggml_backend_t ggml_backend_metal_init(void) {
-    struct ggml_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
-
+    struct ggml_backend_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
     if (ctx == NULL) {
+        GGML_METAL_LOG_ERROR("%s: error: failed to allocate context\n", __func__);
         return NULL;
     }
 
@@ -3504,15 +3528,24 @@ bool ggml_backend_is_metal(ggml_backend_t backend) {
 void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {
     GGML_ASSERT(ggml_backend_is_metal(backend));
 
-    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
+    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
 
     ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
 }
 
+void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data) {
+    GGML_ASSERT(ggml_backend_is_metal(backend));
+
+    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
+
+    ctx->abort_callback = abort_callback;
+    ctx->abort_callback_data = user_data;
+}
+
 bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) {
     GGML_ASSERT(ggml_backend_is_metal(backend));
 
-    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
+    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
 
     return [ctx->device supportsFamily:(MTLGPUFamilyApple1 + family - 1)];
 }
@@ -3520,7 +3553,7 @@ bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) {
 void ggml_backend_metal_capture_next_compute(ggml_backend_t backend) {
     GGML_ASSERT(ggml_backend_is_metal(backend));
 
-    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
+    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
     ctx->should_capture_next_compute = true;
 }
 
diff --git a/ggml/src/ggml-quants.c b/ggml/src/ggml-quants.c
index 99bd682f..41362dee 100644
--- a/ggml/src/ggml-quants.c
+++ b/ggml/src/ggml-quants.c
@@ -3952,7 +3952,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     float sumf = 0;
 
 #if defined(__ARM_FEATURE_SVE)
-    if (svcntb() == QK8_0) {
+    if (ggml_sve_cnt_b == QK8_0) {
         const svbool_t ptrueh = svptrue_pat_b8(SV_VL16);
         const svbool_t ptruel = svnot_b_z(svptrue_b8(), ptrueh);
 
@@ -4324,15 +4324,18 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     sumf = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
 #endif
     for (; ib < nb; ++ib) {
-        int sumi = 0;
+        int sumi0 = 0;
+        int sumi1 = 0;
 
         for (int j = 0; j < qk/2; ++j) {
             const int v0 = (x[ib].qs[j] & 0x0F) - 8;
             const int v1 = (x[ib].qs[j] >>   4) - 8;
 
-            sumi += (v0 * y[ib].qs[j]) + (v1 * y[ib].qs[j + qk/2]);
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
         }
 
+        int sumi = sumi0 + sumi1;
         sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
     }
 
@@ -4613,15 +4616,18 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
     sumf = hsum_float_8(acc) + summs;
 #endif
     for (; ib < nb; ++ib) {
-        int sumi = 0;
+        int sumi0 = 0;
+        int sumi1 = 0;
 
         for (int j = 0; j < qk/2; ++j) {
             const int v0 = (x[ib].qs[j] & 0x0F);
             const int v1 = (x[ib].qs[j] >>   4);
 
-            sumi += (v0 * y[ib].qs[j]) + (v1 * y[ib].qs[j + qk/2]);
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
         }
 
+        int sumi = sumi0 + sumi1;
         sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
     }
 
@@ -4967,18 +4973,21 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         uint32_t qh;
         memcpy(&qh, x[ib].qh, sizeof(qh));
 
-        int sumi = 0;
+        int sumi0 = 0;
+        int sumi1 = 0;
 
         for (int j = 0; j < qk/2; ++j) {
             const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
             const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
 
-            const int32_t x0 = ((x[ib].qs[j] & 0x0F) | xh_0) - 16;
-            const int32_t x1 = ((x[ib].qs[j] >>   4) | xh_1) - 16;
+            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
+            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
 
-            sumi += (x0 * y[ib].qs[j]) + (x1 * y[ib].qs[j + qk/2]);
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
         }
 
+        int sumi = sumi0 + sumi1;
         sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
     }
 
@@ -5343,7 +5352,8 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
         uint32_t qh;
         memcpy(&qh, x[ib].qh, sizeof(qh));
 
-        int sumi = 0;
+        int sumi0 = 0;
+        int sumi1 = 0;
 
         for (int j = 0; j < qk/2; ++j) {
             const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
@@ -5352,9 +5362,11 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
             const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
             const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
 
-            sumi += (x0 * y[ib].qs[j]) + (x1 * y[ib].qs[j + qk/2]);
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
         }
 
+        int sumi = sumi0 + sumi1;
         sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
     }
 
@@ -5445,7 +5457,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     float sumf = 0;
 
 #if defined(__ARM_FEATURE_SVE)
-    if (svcntb() == QK8_0) {
+    if (ggml_sve_cnt_b == QK8_0) {
         svfloat32_t sumv0 = svdup_n_f32(0.0f);
         svfloat32_t sumv1 = svdup_n_f32(0.0f);
 
@@ -6591,22 +6603,22 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             // compute mask for subtraction
             vuint8m1_t qh_m0 = __riscv_vand_vx_u8m1(vqh, m, vl);
             vbool8_t vmask_0 = __riscv_vmseq_vx_u8m1_b8(qh_m0, 0, vl);
-            vint8m1_t q3_m0 = __riscv_vsub_vx_i8m1_m(vmask_0, q3_0, 0x4, vl);
+            vint8m1_t q3_m0 = __riscv_vsub_vx_i8m1_mu(vmask_0, q3_0, q3_0, 0x4, vl);
             m <<= 1;
 
             vuint8m1_t qh_m1 = __riscv_vand_vx_u8m1(vqh, m, vl);
             vbool8_t vmask_1 = __riscv_vmseq_vx_u8m1_b8(qh_m1, 0, vl);
-            vint8m1_t q3_m1 = __riscv_vsub_vx_i8m1_m(vmask_1, q3_1, 0x4, vl);
+            vint8m1_t q3_m1 = __riscv_vsub_vx_i8m1_mu(vmask_1, q3_1, q3_1, 0x4, vl);
             m <<= 1;
 
             vuint8m1_t qh_m2 = __riscv_vand_vx_u8m1(vqh, m, vl);
             vbool8_t vmask_2 = __riscv_vmseq_vx_u8m1_b8(qh_m2, 0, vl);
-            vint8m1_t q3_m2 = __riscv_vsub_vx_i8m1_m(vmask_2, q3_2, 0x4, vl);
+            vint8m1_t q3_m2 = __riscv_vsub_vx_i8m1_mu(vmask_2, q3_2, q3_2, 0x4, vl);
             m <<= 1;
 
             vuint8m1_t qh_m3 = __riscv_vand_vx_u8m1(vqh, m, vl);
             vbool8_t vmask_3 = __riscv_vmseq_vx_u8m1_b8(qh_m3, 0, vl);
-            vint8m1_t q3_m3 = __riscv_vsub_vx_i8m1_m(vmask_3, q3_3, 0x4, vl);
+            vint8m1_t q3_m3 = __riscv_vsub_vx_i8m1_mu(vmask_3, q3_3, q3_3, 0x4, vl);
             m <<= 1;
 
             // load Q8 and take product with Q3
@@ -7862,13 +7874,13 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             vint8m1_t q5_a = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q5_x, 0x0F, vl));
             vuint8m1_t qh_m1 = __riscv_vand_vx_u8m1(vqh, m, vl);
             vbool8_t vmask_1 = __riscv_vmsne_vx_u8m1_b8(qh_m1, 0, vl);
-            vint8m1_t q5_m1 = __riscv_vadd_vx_i8m1_m(vmask_1, q5_a, 16, vl);
+            vint8m1_t q5_m1 = __riscv_vadd_vx_i8m1_mu(vmask_1, q5_a, q5_a, 16, vl);
             m <<= 1;
 
             vint8m1_t q5_l = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(q5_x, 0x04, vl));
             vuint8m1_t qh_m2 = __riscv_vand_vx_u8m1(vqh, m, vl);
             vbool8_t vmask_2 = __riscv_vmsne_vx_u8m1_b8(qh_m2, 0, vl);
-            vint8m1_t q5_m2 = __riscv_vadd_vx_i8m1_m(vmask_2, q5_l, 16, vl);
+            vint8m1_t q5_m2 = __riscv_vadd_vx_i8m1_mu(vmask_2, q5_l, q5_l, 16, vl);
             m <<= 1;
 
             vint16m2_t v0 = __riscv_vwmul_vv_i16m2(q5_m1, q8_y1, vl);
@@ -12851,7 +12863,7 @@ static void quantize_row_iq2_xxs_impl(const float * restrict x, void * restrict
                     printf("Oops: found point %u not on grid:", u);
                     for (int i = 0; i < 8; ++i) printf(" %d", L[8*k+i]);
                     printf("\n");
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
                 }
                 q2[2*ib+0] |= ((uint32_t) grid_index << 8*k);
                 q2[2*ib+1] |= (block_signs[k] << 7*k);
@@ -13030,7 +13042,7 @@ static void quantize_row_iq2_xs_impl(const float * restrict x, void * restrict v
                     printf("Oops: found point %u not on grid:", u);
                     for (int i = 0; i < 8; ++i) printf(" %d", L[8*k+i]);
                     printf("\n");
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
                 }
                 q2[2*ib+k] = grid_index | (block_signs[k] << 9);
             }
@@ -13473,7 +13485,7 @@ static void quantize_row_iq3_xxs_impl(int grid_size, const float * restrict x, v
                     printf("Oops: found point %u not on grid:", u);
                     for (int i = 0; i < 4; ++i) printf(" %d", L[4*k+i]);
                     printf("\n");
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
                 }
                 if (grid_size == 256) {
                     q3[8*ib+k] = grid_index;
@@ -13686,7 +13698,7 @@ static void quantize_row_iq3_s_impl(int block_size, const float * restrict x, vo
                     printf("Oops: found point %u not on grid:", u);
                     for (int i = 0; i < 4; ++i) printf(" %d", L[4*k+i]);
                     printf("\n");
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
                 }
                 qs[k] = grid_index & 255;
                 qh[(ib*bs4+k)/8] |= ((grid_index >> 8) << ((ib*bs4+k)%8));
@@ -14662,7 +14674,7 @@ static void quantize_row_iq2_s_impl(const float * restrict x, void * restrict vy
                     printf("Oops: found point %u not on grid:", u);
                     for (int i = 0; i < 8; ++i) printf(" %d", L[8*k+i]);
                     printf("\n");
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
                 }
                 const int i8 = 2*ib + k;
                 y[ibl].qs[i8] = grid_index & 255;
@@ -14782,7 +14794,7 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte
     }
 
     if (nbytes % ggml_type_size(type) != 0) {
-        fprintf(stderr, "%s: invalid size %zu for type %d\n", __func__, nbytes, type);
+        fprintf(stderr, "%s: invalid size %zu for type %s (type size = %zu)\n", __func__, nbytes, ggml_type_name(type), ggml_type_size(type));
         return false;
     }
 
diff --git a/ggml/src/ggml-quants.h b/ggml/src/ggml-quants.h
index 91063633..775aa875 100644
--- a/ggml/src/ggml-quants.h
+++ b/ggml/src/ggml-quants.h
@@ -146,6 +146,10 @@ void iq2xs_free_impl(enum ggml_type type);
 void iq3xs_init_impl(int grid_size);
 void iq3xs_free_impl(int grid_size);
 
+#if defined(__ARM_FEATURE_SVE)
+extern int ggml_sve_cnt_b;
+#endif
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/ggml/src/ggml-rpc.cpp b/ggml/src/ggml-rpc.cpp
index b01ad267..7757615f 100644
--- a/ggml/src/ggml-rpc.cpp
+++ b/ggml/src/ggml-rpc.cpp
@@ -197,6 +197,10 @@ static std::shared_ptr<socket_t> create_server_socket(const char * host, int por
         fprintf(stderr, "Failed to set SO_REUSEADDR\n");
         return nullptr;
     }
+    if (inet_addr(host) == INADDR_NONE) {
+        fprintf(stderr, "Invalid host address: %s\n", host);
+        return nullptr;
+    }
     struct sockaddr_in serv_addr;
     serv_addr.sin_family = AF_INET;
     serv_addr.sin_addr.s_addr = inet_addr(host);
@@ -879,6 +883,14 @@ ggml_tensor * rpc_server::deserialize_tensor(struct ggml_context * ctx, const rp
     if (result->buffer && buffers.find(result->buffer) == buffers.end()) {
         return nullptr;
     }
+
+    // require that the tensor data does not go beyond the buffer end
+    uint64_t tensor_size = (uint64_t) ggml_nbytes(result);
+    uint64_t buffer_start = (uint64_t) ggml_backend_buffer_get_base(result->buffer);
+    uint64_t buffer_size = (uint64_t) ggml_backend_buffer_get_size(result->buffer);
+    GGML_ASSERT(tensor->data + tensor_size >= tensor->data); // check for overflow
+    GGML_ASSERT(tensor->data >= buffer_start && tensor->data + tensor_size <= buffer_start + buffer_size);
+
     result->op = (ggml_op) tensor->op;
     for (uint32_t i = 0; i < GGML_MAX_OP_PARAMS / sizeof(int32_t); i++) {
         result->op_params[i] = tensor->op_params[i];
@@ -898,7 +910,7 @@ bool rpc_server::set_tensor(const std::vector<uint8_t> & input) {
     const rpc_tensor * in_tensor = (const rpc_tensor *)input.data();
     uint64_t offset;
     memcpy(&offset, input.data() + sizeof(rpc_tensor), sizeof(offset));
-    size_t size = input.size() - sizeof(rpc_tensor) - sizeof(offset);
+    const size_t size = input.size() - sizeof(rpc_tensor) - sizeof(offset);
 
     struct ggml_init_params params {
         /*.mem_size   =*/ ggml_tensor_overhead(),
@@ -913,6 +925,17 @@ bool rpc_server::set_tensor(const std::vector<uint8_t> & input) {
         return false;
     }
     GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %zu\n", __func__, (void*)tensor->buffer, tensor->data, offset, size);
+
+    // sanitize tensor->data
+    {
+        const size_t p0 = (size_t) ggml_backend_buffer_get_base(tensor->buffer);
+        const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer);
+
+        if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) {
+            GGML_ABORT("[%s] tensor->data out of bounds\n", __func__);
+        }
+    }
+
     const void * data = input.data() + sizeof(rpc_tensor) + sizeof(offset);
     ggml_backend_tensor_set(tensor, data, offset, size);
     ggml_free(ctx);
@@ -943,6 +966,17 @@ bool rpc_server::get_tensor(const std::vector<uint8_t> & input, std::vector<uint
         return false;
     }
     GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %" PRIu64 "\n", __func__, (void*)tensor->buffer, tensor->data, offset, size);
+
+    // sanitize tensor->data
+    {
+        const size_t p0 = (size_t) ggml_backend_buffer_get_base(tensor->buffer);
+        const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer);
+
+        if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) {
+            GGML_ABORT("[%s] tensor->data out of bounds\n", __func__);
+        }
+    }
+
     // output serialization format: | data (size bytes) |
     output.resize(size, 0);
     ggml_backend_tensor_get(tensor, output.data(), offset, size);
diff --git a/ggml/src/ggml-sycl.cpp b/ggml/src/ggml-sycl.cpp
index 36518ff9..d8eb86c2 100644
--- a/ggml/src/ggml-sycl.cpp
+++ b/ggml/src/ggml-sycl.cpp
@@ -1723,7 +1723,7 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
                 });
         });
     } else {
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 }
 
@@ -2075,8 +2075,8 @@ static dpct::err0 ggml_sycl_cpy_tensor_2d(void *dst,
         // GGML_SYCL_DEBUG("current device index %d\n", id);
         src_ptr = (char *) extra->data_device[id];
     } else {
-        // GGML_SYCL_DEBUG("GGML_ASSERT(false)\n");
-        GGML_ASSERT(false);
+        // GGML_SYCL_DEBUG("GGML_ABORT("fatal error")\n");
+        GGML_ABORT("fatal error");
     }
     char * dst_ptr = (char *) dst;
 
@@ -2163,7 +2163,7 @@ static void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, const ggml_te
         default:
             // TODO: k-quants
             fprintf(stderr, "%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
             break;
     }
 }
@@ -2192,7 +2192,7 @@ inline void ggml_sycl_op_bin_bcast(ggml_backend_sycl_context & ctx, const ggml_t
     } else {
         fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__,
             ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type));
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 }
 
@@ -2476,7 +2476,7 @@ static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_SYC
         case GGML_TYPE_Q6_K:
             return 64;
         default:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
     }
 
 }
@@ -3101,7 +3101,7 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
                     SYCL_CHECK(ggml_sycl_cpy_tensor_2d(
                                    src1_ddf_i, src1, i03, i02, src1_col_0, src1_col_0+src1_ncols, stream));
                 } else {
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
                 }
 
                 if (convert_src1_to_q8_1 && !src1_is_contiguous) {
@@ -3896,7 +3896,7 @@ static void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor *sr
     } else {
         fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__,
                 ggml_type_name(src0->type), ggml_type_name(src1->type));
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 
     (void) dst;
@@ -3981,6 +3981,9 @@ bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tens
     ggml_sycl_func_t func;
 
     switch (tensor->op) {
+        case GGML_OP_CONV_TRANSPOSE_1D:
+            func = ggml_sycl_op_conv_transpose_1d;
+            break;
         case GGML_OP_REPEAT:
             func = ggml_sycl_repeat;
             break;
@@ -4105,6 +4108,9 @@ bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tens
         case GGML_OP_ARGSORT:
             func = ggml_sycl_argsort;
             break;
+        case GGML_OP_TIMESTEP_EMBEDDING:
+            func = ggml_sycl_op_timestep_embedding;
+            break;
         default:
             return false;
     }
@@ -5090,6 +5096,15 @@ GGML_CALL static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t back
 
 GGML_CALL static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
     switch (op->op) {
+        case GGML_OP_CONV_TRANSPOSE_1D:
+            {
+                ggml_type src0_type = op->src[0]->type;
+                ggml_type src1_type = op->src[1]->type;
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
+                    return true;
+                }
+                return false;
+            } break;
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
                 case GGML_UNARY_OP_GELU:
@@ -5213,6 +5228,7 @@ GGML_CALL static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, cons
         case GGML_OP_UPSCALE:
         case GGML_OP_PAD:
         case GGML_OP_LEAKY_RELU:
+        case GGML_OP_TIMESTEP_EMBEDDING:
             return true;
         default:
             return false;
diff --git a/ggml/src/ggml-sycl/backend.hpp b/ggml/src/ggml-sycl/backend.hpp
index 067181de..58dd9c9a 100644
--- a/ggml/src/ggml-sycl/backend.hpp
+++ b/ggml/src/ggml-sycl/backend.hpp
@@ -15,6 +15,7 @@
 
 #include "concat.hpp"
 #include "common.hpp"
+#include "conv.hpp"
 #include "convert.hpp"
 #include "dequantize.hpp"
 #include "dmmv.hpp"
@@ -23,5 +24,6 @@
 #include "rope.hpp"
 #include "norm.hpp"
 #include "softmax.hpp"
+#include "tsembd.hpp"
 
 #endif // GGML_SYCL_BACKEND_HPP
diff --git a/ggml/src/ggml-sycl/common.hpp b/ggml/src/ggml-sycl/common.hpp
index 397bd98d..86d8b40e 100644
--- a/ggml/src/ggml-sycl/common.hpp
+++ b/ggml/src/ggml-sycl/common.hpp
@@ -100,7 +100,7 @@ static void crash() {
     const char* msg) {
   fprintf(stderr, "SYCL error: %s: %s\n", stmt, msg);
   fprintf(stderr, "  in function %s at %s:%d\n", func, file, line);
-  GGML_ASSERT(!"SYCL error");
+  GGML_ABORT("SYCL error");
 }
 
 #define SYCL_CHECK(err)                     \
diff --git a/ggml/src/ggml-sycl/conv.cpp b/ggml/src/ggml-sycl/conv.cpp
new file mode 100644
index 00000000..bc4ab1dd
--- /dev/null
+++ b/ggml/src/ggml-sycl/conv.cpp
@@ -0,0 +1,99 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#include "conv.hpp"
+
+static  void conv_transpose_1d_kernel(
+        const int s0, const int output_size,
+        const int src0_ne0, const int src0_ne1, const int src0_ne2,
+        const int src1_ne0, const int dst_ne0,
+        const float * src0, const float * src1,  float * dst,
+        const sycl::nd_item<3> &item_ct1) {
+    int global_index = item_ct1.get_local_id(2) +
+                       item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    if (global_index >= output_size) {
+        return;
+    }
+
+    int out_index = global_index / dst_ne0;
+
+    float accumulator = 0;
+
+    for (int c = 0; c < src0_ne2; c++) {
+        int idx = global_index % dst_ne0;
+
+        int kernel_offset = (src0_ne0 * src0_ne1 * c) + (out_index * src0_ne0);
+        int input_offset = src1_ne0 * c;
+
+        for (int i = 0; i < src1_ne0; i++) {
+            if (!(idx >= i*s0 && idx < i*s0 + src0_ne0)) {
+                continue;
+            }
+            int weight_idx = idx - i*s0;
+
+            float kernel_weight = src0[kernel_offset + weight_idx];
+            float input_value =  src1[input_offset+i];
+
+            accumulator += kernel_weight * input_value;
+        }
+    }
+    dst[global_index] = accumulator;
+}
+
+static void conv_transpose_1d_f32_f32_sycl(
+    const int s0, const int output_size,
+    const int src0_ne0, const int src0_ne1, const int src0_ne2,
+    const int src1_ne0, const int dst_ne0,
+    const float *src0, const float *src1, float *dst,
+    const queue_ptr& stream) {
+
+    const int num_blocks = (output_size + SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE - 1) / SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE;
+    const sycl::range<3> block_dims(1, 1, SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE);
+    const sycl::range<3> block_nums(1, 1, num_blocks);
+    stream->parallel_for(
+        sycl::nd_range<3>(
+            block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) {
+            conv_transpose_1d_kernel(
+                s0, output_size,
+                src0_ne0, src0_ne1, src0_ne2,
+                src1_ne0, dst_ne0,
+                src0, src1, dst, item_ct1);
+        });
+}
+
+void ggml_sycl_op_conv_transpose_1d(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+    const ggml_tensor *src1, ggml_tensor *dst) {
+    const float * src0_d = (const float *)src0->data;
+    const float * src1_d = (const float *)src1->data;
+
+    float * dst_d = (float *)dst->data;
+    dpct::queue_ptr stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src1));
+
+    const int32_t * opts = (const int32_t *)dst->op_params;
+
+    const int s0 = opts[0];
+
+    const int64_t output_size = ggml_nelements(dst);
+
+    conv_transpose_1d_f32_f32_sycl(s0, output_size,
+        src0->ne[0], src0->ne[1], src0->ne[2],
+        src1->ne[0], dst->ne[0],
+        src0_d, src1_d, dst_d, stream);
+}
+
diff --git a/ggml/src/ggml-sycl/conv.hpp b/ggml/src/ggml-sycl/conv.hpp
new file mode 100644
index 00000000..eb20730f
--- /dev/null
+++ b/ggml/src/ggml-sycl/conv.hpp
@@ -0,0 +1,21 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_CONV_HPP
+#define GGML_SYCL_CONV_HPP
+
+#include "common.hpp"
+
+void ggml_sycl_op_conv_transpose_1d(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+  const ggml_tensor *src1, ggml_tensor *dst);
+
+#endif // GGML_SYCL_CONV_HPP
diff --git a/ggml/src/ggml-sycl/dmmv.cpp b/ggml/src/ggml-sycl/dmmv.cpp
index 70a94fc1..ae45630e 100644
--- a/ggml/src/ggml-sycl/dmmv.cpp
+++ b/ggml/src/ggml-sycl/dmmv.cpp
@@ -1011,7 +1011,7 @@ void ggml_sycl_op_dequantize_mul_mat_vec(
             break;
         default:
             printf("ggml_sycl_op_dequantize_mul_mat_vec unsupported GGML_TYPE %d\n", src0->type);
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
             break;
     }
 
diff --git a/ggml/src/ggml-sycl/dpct/helper.hpp b/ggml/src/ggml-sycl/dpct/helper.hpp
index 4aaa76bf..fe4a8f74 100644
--- a/ggml/src/ggml-sycl/dpct/helper.hpp
+++ b/ggml/src/ggml-sycl/dpct/helper.hpp
@@ -874,7 +874,7 @@ namespace dpct
         inline std::string get_preferred_gpu_platform_name() {
             std::string result;
 
-            std::string filter = "level-zero";
+            std::string filter = "";
             char* env = getenv("ONEAPI_DEVICE_SELECTOR");
             if (env) {
                 if (std::strstr(env, "level_zero")) {
@@ -892,11 +892,24 @@ namespace dpct
                 else {
                     throw std::runtime_error("invalid device filter: " + std::string(env));
                 }
+            } else {
+                auto default_device = sycl::device(sycl::default_selector_v);
+                auto default_platform_name = default_device.get_platform().get_info<sycl::info::platform::name>();
+
+                if (std::strstr(default_platform_name.c_str(), "Level-Zero") || default_device.is_cpu()) {
+                    filter = "level-zero";
+                }
+                else if (std::strstr(default_platform_name.c_str(), "CUDA")) {
+                    filter = "cuda";
+                }
+                else if (std::strstr(default_platform_name.c_str(), "HIP")) {
+                    filter = "hip";
+                }
             }
 
-            auto plaform_list = sycl::platform::get_platforms();
+            auto platform_list = sycl::platform::get_platforms();
 
-            for (const auto& platform : plaform_list) {
+            for (const auto& platform : platform_list) {
                 auto devices = platform.get_devices();
                 auto gpu_dev = std::find_if(devices.begin(), devices.end(), [](const sycl::device& d) {
                     return d.is_gpu();
@@ -975,7 +988,7 @@ namespace dpct
             if (backend == "opencl:cpu") return 4;
             if (backend == "opencl:acc") return 5;
             printf("convert_backend_index: can't handle backend=%s\n", backend.c_str());
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
         static bool compare_backend(std::string &backend1, std::string &backend2) {
             return convert_backend_index(backend1) < convert_backend_index(backend2);
diff --git a/ggml/src/ggml-sycl/mmq.cpp b/ggml/src/ggml-sycl/mmq.cpp
index 3107ba91..e952533d 100644
--- a/ggml/src/ggml-sycl/mmq.cpp
+++ b/ggml/src/ggml-sycl/mmq.cpp
@@ -1799,7 +1799,7 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy,
         mmq_y  =  MMQ_Y_Q4_0_PASCAL;
         nwarps = NWARPS_Q4_0_PASCAL;
     } else {
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 
     const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -1914,7 +1914,7 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy,
         mmq_y  =  MMQ_Y_Q4_1_PASCAL;
         nwarps = NWARPS_Q4_1_PASCAL;
     } else {
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 
     const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -2029,7 +2029,7 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy,
         mmq_y  =  MMQ_Y_Q5_0_PASCAL;
         nwarps = NWARPS_Q5_0_PASCAL;
     } else {
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 
     const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -2144,7 +2144,7 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy,
         mmq_y  =  MMQ_Y_Q5_1_PASCAL;
         nwarps = NWARPS_Q5_1_PASCAL;
     } else {
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 
     const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -2259,7 +2259,7 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy,
         mmq_y  =  MMQ_Y_Q8_0_PASCAL;
         nwarps = NWARPS_Q8_0_PASCAL;
     } else {
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 
     const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -2374,7 +2374,7 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy,
         mmq_y  =  MMQ_Y_Q2_K_PASCAL;
         nwarps = NWARPS_Q2_K_PASCAL;
     } else {
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 
     const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -2497,7 +2497,7 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
         mmq_y  =  MMQ_Y_Q3_K_PASCAL;
         nwarps = NWARPS_Q3_K_PASCAL;
     } else {
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 
     const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -2625,7 +2625,7 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy,
         mmq_y  =  MMQ_Y_Q4_K_PASCAL;
         nwarps = NWARPS_Q4_K_PASCAL;
     } else {
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 
     const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -2746,7 +2746,7 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy,
         mmq_y  =  MMQ_Y_Q5_K_PASCAL;
         nwarps = NWARPS_Q5_K_PASCAL;
     } else {
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 
     const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -2867,7 +2867,7 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy,
         mmq_y  =  MMQ_Y_Q6_K_PASCAL;
         nwarps = NWARPS_Q6_K_PASCAL;
     } else {
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 
     const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -3016,7 +3016,7 @@ void ggml_sycl_op_mul_mat_q(
             ggml_mul_mat_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
             break;
         default:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
             break;
     }
 
diff --git a/ggml/src/ggml-sycl/mmvq.cpp b/ggml/src/ggml-sycl/mmvq.cpp
index 3fbc4dd6..1b96925e 100644
--- a/ggml/src/ggml-sycl/mmvq.cpp
+++ b/ggml/src/ggml-sycl/mmvq.cpp
@@ -902,7 +902,7 @@ static void mul_mat_vec_iq4_nl_q8_1_sycl(const void *vx, const void *vy,
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
                     [[intel::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q_iq4_nl_q8_1<QK4_NL, QI4_NL, block_iq4_nl, 1>(
+                        mul_mat_vec_q_iq4_nl_q8_1<QK4_NL, QI4_NL, block_iq4_nl, 2>(
                             vx, vy, dst, ncols, nrows, item_ct1);
                     });
         });
@@ -1017,7 +1017,7 @@ void ggml_sycl_op_mul_mat_vec_q(
             mul_mat_vec_iq4_xs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
             break;
         default:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
             break;
         }
     }
diff --git a/ggml/src/ggml-sycl/norm.cpp b/ggml/src/ggml-sycl/norm.cpp
index cccf87d0..b3159b9d 100644
--- a/ggml/src/ggml-sycl/norm.cpp
+++ b/ggml/src/ggml-sycl/norm.cpp
@@ -225,9 +225,8 @@ static void norm_f32_sycl(const float* x, float* dst, const int ncols,
 }
 
 static void group_norm_f32_sycl(const float* x, float* dst,
-    const int num_groups, const int group_size,
+    const int num_groups, const float eps, const int group_size,
     const int ne_elements, queue_ptr stream, int device) {
-    static const float eps = 1e-6f;
     if (group_size < 1024) {
         const sycl::range<3> block_dims(1, 1, WARP_SIZE);
         stream->submit([&](sycl::handler& cgh) {
@@ -343,8 +342,12 @@ void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, const ggml_tensor*
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
 
     int num_groups = dst->op_params[0];
+
+    float eps;
+    memcpy(&eps, dst->op_params + 1, sizeof(float));
+
     int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
-    group_norm_f32_sycl(src0_dd, dst_dd, num_groups, group_size, src0->ne[0] * src0->ne[1] * src0->ne[2], main_stream, ctx.device);
+    group_norm_f32_sycl(src0_dd, dst_dd, num_groups, eps, group_size, src0->ne[0] * src0->ne[1] * src0->ne[2], main_stream, ctx.device);
 
     (void)src1;
     (void)dst;
diff --git a/ggml/src/ggml-sycl/presets.hpp b/ggml/src/ggml-sycl/presets.hpp
index 15ddcac1..340ab8e9 100644
--- a/ggml/src/ggml-sycl/presets.hpp
+++ b/ggml/src/ggml-sycl/presets.hpp
@@ -41,6 +41,8 @@
 #define SYCL_ACC_BLOCK_SIZE 256
 #define SYCL_IM2COL_BLOCK_SIZE 256
 #define SYCL_POOL2D_BLOCK_SIZE 256
+#define SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE 256
+#define SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE 256
 
 // dmmv = dequantize_mul_mat_vec
 #ifndef GGML_SYCL_DMMV_X
diff --git a/ggml/src/ggml-sycl/rope.cpp b/ggml/src/ggml-sycl/rope.cpp
index 6f507941..c7545bcc 100644
--- a/ggml/src/ggml-sycl/rope.cpp
+++ b/ggml/src/ggml-sycl/rope.cpp
@@ -251,7 +251,7 @@ void ggml_sycl_op_rope(
                 attn_factor, corr_dims, freq_factors, main_stream
             );
         } else {
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
     } else {
         if (src0->type == GGML_TYPE_F32) {
@@ -265,7 +265,7 @@ void ggml_sycl_op_rope(
                 attn_factor, corr_dims, freq_factors, main_stream
             );
         } else {
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
     }
 
diff --git a/ggml/src/ggml-sycl/tsembd.cpp b/ggml/src/ggml-sycl/tsembd.cpp
new file mode 100644
index 00000000..d5c227cd
--- /dev/null
+++ b/ggml/src/ggml-sycl/tsembd.cpp
@@ -0,0 +1,71 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#include "tsembd.hpp"
+
+static void timestep_embedding_f32(
+        const float * timesteps, float * dst, const int nb1,
+        const int dim, const int max_period, const sycl::nd_item<3> &item_ct1) {
+    // item_ct1.get_group(1)(blockIDx.y): idx of timesteps->ne[0]
+    // item_ct1.get_group(2) (blockIDx.x): idx of ((dim + 1) / 2) / BLOCK_SIZE
+    int i = item_ct1.get_group(1);
+    int j = item_ct1.get_local_id(2) + item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    float * embed_data = (float *)((char *)dst +  i*nb1);
+
+    if (dim % 2 != 0 && j == ((dim + 1) / 2)) {
+        embed_data[dim] = 0.f;
+    }
+
+    int half = dim / 2;
+    if (j >= half) {
+        return;
+    }
+
+    float timestep = timesteps[i];
+    float freq = (float)sycl::native::exp(-(sycl::log((float)max_period)) * j / half);
+    float arg = timestep * freq;
+    embed_data[j] = sycl::cos(arg);
+    embed_data[j + half] = sycl::sin(arg);
+}
+
+static void timestep_embedding_f32_sycl(
+        const float * x, float * dst, const int ne00, const int nb1,
+        const int dim, const int max_period, const queue_ptr& stream) {
+    // As the kernel returns when thread.idx is larger than dim/2, the half_ceil does not need to pad
+    int half_ceil = dim / 2;
+    int num_blocks = (half_ceil + SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE - 1) / SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE;
+    sycl::range<3> block_dims(1, 1, SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE);
+    sycl::range<3> gridDim(1, ne00, num_blocks);
+    stream->parallel_for(
+        sycl::nd_range<3>(
+            gridDim * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) {
+            timestep_embedding_f32(
+                x, dst, nb1, dim, max_period, item_ct1
+            );
+        });
+}
+
+void ggml_sycl_op_timestep_embedding(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+    const ggml_tensor *src1, ggml_tensor * dst) {
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    dpct::queue_ptr stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    const int dim = dst->op_params[0];
+    const int max_period = dst->op_params[1];
+
+    timestep_embedding_f32_sycl(src0_d, dst_d, src0->ne[0], dst->nb[1], dim, max_period, stream);
+}
diff --git a/ggml/src/ggml-sycl/tsembd.hpp b/ggml/src/ggml-sycl/tsembd.hpp
new file mode 100644
index 00000000..ff854c33
--- /dev/null
+++ b/ggml/src/ggml-sycl/tsembd.hpp
@@ -0,0 +1,21 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_TSEMBD_HPP
+#define GGML_SYCL_TSEMBD_HPP
+
+#include "common.hpp"
+
+void ggml_sycl_op_timestep_embedding(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+    const ggml_tensor *src1, ggml_tensor * dst);
+
+#endif // GGML_SYCL_TSEMBD_HPP
diff --git a/ggml/src/ggml-vulkan.cpp b/ggml/src/ggml-vulkan.cpp
index 6bcd81a7..86732837 100644
--- a/ggml/src/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan.cpp
@@ -177,24 +177,33 @@ struct vk_device_struct {
     vk_pipeline pipeline_mul_mat_vec_nc_f16_f32;
     vk_pipeline pipeline_get_rows[GGML_TYPE_COUNT];
     vk_pipeline pipeline_get_rows_f32[GGML_TYPE_COUNT];
+    vk_pipeline pipeline_add_f32, pipeline_add_f16_f32_f16;
     vk_pipeline pipeline_mul_f32;
     vk_pipeline pipeline_div_f32;
-    vk_pipeline pipeline_add_f32;
+    vk_pipeline pipeline_concat_f32, pipeline_concat_f16, pipeline_concat_i32;
+    vk_pipeline pipeline_upscale_f32;
     vk_pipeline pipeline_scale_f32;
     vk_pipeline pipeline_sqr_f32;
     vk_pipeline pipeline_clamp_f32;
+    vk_pipeline pipeline_pad_f32;
     vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16;
     vk_pipeline pipeline_norm_f32;
+    vk_pipeline pipeline_group_norm_f32;
     vk_pipeline pipeline_rms_norm_f32;
     vk_pipeline pipeline_gelu_f32;
+    vk_pipeline pipeline_gelu_quick_f32;
     vk_pipeline pipeline_silu_f32;
     vk_pipeline pipeline_relu_f32;
+    vk_pipeline pipeline_leaky_relu_f32;
+    vk_pipeline pipeline_tanh_f32;
     vk_pipeline pipeline_diag_mask_inf_f32;
     vk_pipeline pipeline_soft_max_f32, pipeline_soft_max_f32_f16;
     vk_pipeline pipeline_rope_norm_f32, pipeline_rope_norm_f16;
     vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16;
     vk_pipeline pipeline_argsort_f32;
     vk_pipeline pipeline_sum_rows_f32;
+    vk_pipeline pipeline_im2col_f32, pipeline_im2col_f32_f16;
+    vk_pipeline pipeline_timestep_embedding_f32;
 
     std::vector<vk_pipeline_ref> pipelines;
 
@@ -236,8 +245,8 @@ struct vk_device_struct {
 };
 
 struct vk_buffer_struct {
-    vk::Buffer buffer;
-    vk::DeviceMemory device_memory;
+    vk::Buffer buffer = VK_NULL_HANDLE;
+    vk::DeviceMemory device_memory = VK_NULL_HANDLE;
     vk::MemoryPropertyFlags memory_property_flags;
     void * ptr;
     size_t size = 0;
@@ -259,6 +268,10 @@ struct vk_subbuffer {
     vk_buffer buffer;
     uint64_t offset;
     uint64_t size;
+
+    operator vk::DescriptorBufferInfo() const {
+        return { buffer->buffer, offset, size };
+    }
 };
 
 struct vk_semaphore {
@@ -320,7 +333,7 @@ struct vk_op_binary_push_constants {
     uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13; uint32_t nb10; uint32_t nb11; uint32_t nb12; uint32_t nb13;
     uint32_t ne20; uint32_t ne21; uint32_t ne22; uint32_t ne23; uint32_t nb20; uint32_t nb21; uint32_t nb22; uint32_t nb23;
     uint32_t d_offset;
-    float param1; float param2;
+    float param1; float param2; int32_t param3;
 };
 
 struct vk_op_diag_mask_push_constants {
@@ -358,6 +371,25 @@ struct vk_op_argsort_push_constants {
     int32_t order;
 };
 
+struct vk_op_im2col_push_constants {
+    uint32_t batch_offset; uint32_t offset_delta;
+    uint32_t IC;
+    uint32_t IW; uint32_t IH;
+    uint32_t OW; uint32_t OH;
+    uint32_t KW; uint32_t KH;
+    uint32_t pelements;
+    uint32_t CHW;
+    int32_t s0; int32_t s1;
+    int32_t p0; int32_t p1;
+    int32_t d0; int32_t d1;
+};
+
+struct vk_op_timestep_embedding_push_constants {
+    uint32_t nb1;
+    uint32_t dim;
+    uint32_t max_period;
+};
+
 // Allow pre-recording command buffers
 struct vk_staging_memcpy {
     vk_staging_memcpy(void * _dst, const void * _src, size_t _n) : dst(_dst), src(_src), n(_n) {}
@@ -367,28 +399,32 @@ struct vk_staging_memcpy {
     size_t n;
 };
 
-struct vk_context {
-    size_t idx;
+struct vk_op_upscale_push_constants {
+    uint32_t ne; uint32_t d_offset;
+    uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
+    uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13;
+    float sf0; float sf1; float sf2; float sf3;
+};
 
+struct vk_context_struct {
     vk_submission * s;
     std::vector<vk_sequence> seqs;
 
-    ggml_tensor * exit_tensor;
+    int exit_tensor_idx;
 
     std::vector<vk_staging_memcpy> in_memcpys;
     std::vector<vk_staging_memcpy> out_memcpys;
 
     vk_queue * q;
 };
+typedef std::shared_ptr<vk_context_struct> vk_context;
+typedef std::weak_ptr<vk_context_struct> vk_context_ref;
 
 struct ggml_tensor_extra_gpu {
-    size_t ctx_idx;
-
     vk_buffer_ref buffer_gpu;
     uint64_t offset;
 
     void reset() {
-        ctx_idx = 0;
         buffer_gpu.reset();
         offset = 0;
     }
@@ -459,8 +495,10 @@ struct ggml_backend_vk_context {
 
     vk_buffer buffer_pool[MAX_VK_BUFFERS];
 
-    vk_context * compute_ctx;
-    vk_context * transfer_ctx;
+    vk_context_ref compute_ctx;
+    vk_context_ref transfer_ctx;
+
+    std::vector<vk_context_ref> tensor_ctxs;
 };
 
 #ifdef GGML_VULKAN_MEMORY_DEBUG
@@ -510,12 +548,12 @@ static vk_instance_t vk_instance;
 static size_t vk_skip_checks;
 static size_t vk_output_tensor;
 
-static void ggml_vk_print_tensor(ggml_backend * ctx, const ggml_tensor * tensor, const char * name);
-static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor * tensor);
-static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_tensor * tensor);
+static void ggml_vk_print_tensor(const ggml_tensor * tensor, const char * name);
+static void ggml_vk_check_results_0(ggml_tensor * tensor);
+static void ggml_vk_check_results_1(ggml_tensor * tensor);
 #endif
 
-typedef void (*ggml_vk_func_t)(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+typedef void (*ggml_vk_func_t)(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
 
 GGML_CALL static void ggml_backend_vk_free(ggml_backend_t backend);
 
@@ -708,11 +746,11 @@ static vk_submission ggml_vk_create_submission(vk_device& device, vk_queue& q, s
     return s;
 }
 
-static void ggml_vk_submit(vk_context * ctx, vk::Fence fence) {
-    VK_LOG_DEBUG("ggml_vk_submit(" << ctx->seqs.size() << ", " << fence << ")");
+static void ggml_vk_submit(vk_context& ctx, vk::Fence fence) {
     if (ctx->seqs.empty()) {
         return;
     }
+    VK_LOG_DEBUG("ggml_vk_submit(" << ctx << ", " << fence << ")");
 
     std::vector<std::vector<uint64_t>> tl_wait_vals;
     std::vector<std::vector<uint64_t>> tl_signal_vals;
@@ -844,21 +882,17 @@ static void ggml_vk_create_queue(vk_device& device, vk_queue& q, uint32_t queue_
     q.stage_flags = stage_flags;
 }
 
-static vk_context * ggml_vk_create_context(ggml_backend_vk_context * ctx, vk_queue& q) {
-    VK_LOG_DEBUG("ggml_vk_create_context()");
-    ctx->gc.contexts.emplace_back();
-    vk_context * result = &ctx->gc.contexts[ctx->gc.contexts.size() - 1];
-    memset((void *) result, 0, sizeof(vk_context));
-    result->idx = ctx->gc.contexts.size() - 1;
+static vk_context ggml_vk_create_context(ggml_backend_vk_context * ctx, vk_queue& q) {
+    vk_context result = std::make_shared<vk_context_struct>();
+    VK_LOG_DEBUG("ggml_vk_create_context(" << result << ")");
+    ctx->gc.contexts.emplace_back(result);
     result->q = &q;
     return result;
 }
 
-static vk_context * ggml_vk_create_temporary_context(vk_queue& q) {
-    VK_LOG_DEBUG("ggml_vk_create_temporary_context()");
-    vk_context * result = new vk_context;
-    memset((void *) result, 0, sizeof(vk_context));
-    result->idx = 0;
+static vk_context ggml_vk_create_temporary_context(vk_queue& q) {
+    vk_context result = std::make_shared<vk_context_struct>();
+    VK_LOG_DEBUG("ggml_vk_create_temporary_context(" << result << ")");
     result->q = &q;
     return result;
 }
@@ -915,6 +949,10 @@ static uint32_t find_properties(const vk::PhysicalDeviceMemoryProperties* mem_pr
 
 static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::MemoryPropertyFlags req_flags, vk::MemoryPropertyFlags fallback_flags = vk::MemoryPropertyFlags(0)) {
     VK_LOG_DEBUG("ggml_vk_create_buffer(" << device->name << ", " << size << ", " << to_string(req_flags) << ", " << to_string(fallback_flags) << ")");
+    if (size > device->max_memory_allocation_size) {
+        throw vk::OutOfDeviceMemoryError("Requested buffer size exceeds device memory allocation limit");
+    }
+
     std::lock_guard<std::mutex> guard(device->mutex);
 
     vk_buffer buf = std::make_shared<vk_buffer_struct>();
@@ -1027,21 +1065,22 @@ static vk_subbuffer ggml_vk_subbuffer(vk_buffer& buf) {
     return { buf, 0, VK_WHOLE_SIZE };
 }
 
-static void ggml_vk_sync_buffers(vk_context * ctx) {
+static void ggml_vk_sync_buffers(vk_context& ctx) {
     VK_LOG_DEBUG("ggml_vk_sync_buffers()");
-    const std::vector<vk::MemoryBarrier> mem_barriers{ { { vk::AccessFlagBits::eMemoryRead | vk::AccessFlagBits::eMemoryWrite }, { vk::AccessFlagBits::eMemoryRead | vk::AccessFlagBits::eMemoryWrite } } };
-
     ctx->s->buffer.pipelineBarrier(
         ctx->q->stage_flags,
         ctx->q->stage_flags,
         {},
-        mem_barriers,
+        { {
+          {vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eShaderWrite | vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite},
+          {vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eShaderWrite | vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite}
+        } },
         {},
         {}
     );
 }
 
-static void ggml_vk_wait_events(vk_context * ctx, std::vector<vk::Event>&& events) {
+static void ggml_vk_wait_events(vk_context& ctx, std::vector<vk::Event>&& events) {
     VK_LOG_DEBUG("ggml_vk_wait_events()");
     if (events.empty()) {
         return;
@@ -1598,6 +1637,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 7 * sizeof(uint32_t), {1, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_rms_norm_f32, "rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f32, "cpy_f32_f32", cpy_f32_f32_len, cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
@@ -1605,20 +1645,31 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_cpy_f16_f16, "cpy_f16_f16", cpy_f16_f16_len, cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_add_f32, "add_f32", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_add_f16_f32_f16, "add_f16_f32_f16", add_f16_f32_f16_len, add_f16_f32_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_mul_f32, "mul_f32", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
-
     ggml_vk_create_pipeline(device, device->pipeline_div_f32, "div_f32", div_f32_len, div_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
 
+    ggml_vk_create_pipeline(device, device->pipeline_concat_f32, "concat_f32", concat_f32_len, concat_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_concat_f16, "concat_f16", concat_f16_len, concat_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_concat_i32, "concat_i32", concat_i32_len, concat_i32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_upscale_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {}, 1);
+
     ggml_vk_create_pipeline(device, device->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_sqr_f32, "sqr_f32", sqr_f32_len, sqr_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_clamp_f32, "clamp_f32", clamp_f32_len, clamp_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
+    ggml_vk_create_pipeline(device, device->pipeline_pad_f32, "pad_f32", pad_f32_len, pad_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+
     ggml_vk_create_pipeline(device, device->pipeline_gelu_f32, "gelu_f32", gelu_f32_len, gelu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_gelu_quick_f32, "gelu_quick_f32", gelu_quick_f32_len, gelu_quick_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_silu_f32, "silu_f32", silu_f32_len, silu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_relu_f32, "relu_f32", relu_f32_len, relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_leaky_relu_f32, "leaky_relu_f32", leaky_relu_f32_len, leaky_relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_tanh_f32, "tanh_f32", tanh_f32_len, tanh_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {512, 1, 1}, {}, 1);
 
@@ -1634,6 +1685,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_argsort_f32, "argsort_f32", argsort_f32_len, argsort_f32_data, "main", 2, sizeof(vk_op_argsort_push_constants), {1024, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_im2col_f32, "im2col_f32", im2col_f32_len, im2col_f32_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_len, im2col_f32_f16_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_timestep_embedding_f32, "timestep_embedding_f32", timestep_embedding_f32_len, timestep_embedding_f32_data, "main", 2, sizeof(vk_op_timestep_embedding_push_constants), {256, 1, 1}, {}, 1);
 }
 
 static vk_device ggml_vk_get_device(size_t idx) {
@@ -1961,7 +2017,7 @@ void ggml_vk_instance_init() {
         // Make sure at least one device exists
         if (devices.empty()) {
             std::cerr << "ggml_vulkan: Error: No devices found." << std::endl;
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
 
         // Default to using all dedicated GPUs
@@ -2057,9 +2113,9 @@ void ggml_vk_instance_init() {
 }
 
 static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
-    GGML_ASSERT(idx < vk_instance.device_indices.size());
     VK_LOG_DEBUG("ggml_vk_init(" << ctx->name << ", " << idx << ")");
     ggml_vk_instance_init();
+    GGML_ASSERT(idx < vk_instance.device_indices.size());
 
     ctx->name = GGML_VK_NAME + std::to_string(idx);
 
@@ -2077,9 +2133,6 @@ static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
     ctx->staging_size = 0;
     ctx->staging_offset = 0;
 
-    ctx->compute_ctx = nullptr;
-    ctx->transfer_ctx = nullptr;
-
 #ifdef GGML_VULKAN_CHECK_RESULTS
     const char* skip_checks = getenv("GGML_VULKAN_SKIP_CHECKS");
     vk_skip_checks = (skip_checks == NULL ? 0 : atoi(skip_checks));
@@ -2112,7 +2165,7 @@ static vk_pipeline ggml_vk_get_to_fp16(ggml_backend_vk_context * ctx, ggml_type
 }
 
 static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type) {
-    VK_LOG_DEBUG("ggml_vk_get_mul_mat_mat_pipeline()");
+    VK_LOG_DEBUG("ggml_vk_get_mul_mat_mat_pipeline(" << ggml_type_name(src0_type) << ", " << ggml_type_name(src1_type) << ")");
     if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
         return ctx->device->pipeline_matmul_f32;
     }
@@ -2126,7 +2179,9 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
         return ctx->device->pipeline_matmul_f16;
     }
 
-    GGML_ASSERT(src1_type == GGML_TYPE_F32);
+    if (src1_type != GGML_TYPE_F32) {
+        return nullptr;
+    }
 
     switch (src0_type) {
         case GGML_TYPE_Q4_0:
@@ -2370,28 +2425,23 @@ static vk_submission ggml_vk_begin_submission(vk_device& device, vk_queue& q, bo
     return s;
 }
 
-static void ggml_vk_dispatch_pipeline(ggml_backend_vk_context * ctx, vk_context * subctx, vk_pipeline& pipeline, std::vector<vk_subbuffer>&& buffers, size_t push_constant_size, const void* push_constants, std::array<uint32_t, 3> elements) {
+
+
+static void ggml_vk_dispatch_pipeline(ggml_backend_vk_context* ctx, vk_context& subctx, vk_pipeline& pipeline, std::initializer_list<vk::DescriptorBufferInfo> const& descriptor_buffer_infos, size_t push_constant_size, const void* push_constants, std::array<uint32_t, 3> elements) {
     const uint32_t wg0 = CEIL_DIV(elements[0], pipeline->wg_denoms[0]);
     const uint32_t wg1 = CEIL_DIV(elements[1], pipeline->wg_denoms[1]);
     const uint32_t wg2 = CEIL_DIV(elements[2], pipeline->wg_denoms[2]);
     VK_LOG_DEBUG("ggml_vk_dispatch_pipeline(" << pipeline->name << ", {";
-    for (auto& buffer : buffers) {
-        std::cerr << "(" << buffer.buffer << ", " << buffer.offset << ", " << buffer.size << "), ";
+    for (auto& buffer : descriptor_buffer_infos) {
+        std::cerr << "(" << buffer << ", " << buffer.offset << ", " << buffer.size << "), ";
     }
     std::cerr << "}, (" << wg0 << "," << wg1 << "," << wg2 << "))");
-    std::vector<vk::DescriptorBufferInfo> descriptor_buffer_infos;
-    std::vector<vk::WriteDescriptorSet> write_descriptor_sets;
     GGML_ASSERT(pipeline->descriptor_set_idx < pipeline->descriptor_sets.size());
-    GGML_ASSERT(buffers.size() == pipeline->parameter_count);
-    vk::DescriptorSet& descriptor_set = pipeline->descriptor_sets[pipeline->descriptor_set_idx++];
-    for (uint32_t i = 0; i < pipeline->parameter_count; i++) {
-        descriptor_buffer_infos.push_back({buffers[i].buffer->buffer, buffers[i].offset, buffers[i].size});
-    }
-    for (uint32_t i = 0; i < pipeline->parameter_count; i++) {
-        write_descriptor_sets.push_back({descriptor_set, i, 0, 1, vk::DescriptorType::eStorageBuffer, nullptr, &descriptor_buffer_infos[i]});
-    }
+    GGML_ASSERT(descriptor_buffer_infos.size() == pipeline->parameter_count);
 
-    ctx->device->device.updateDescriptorSets(write_descriptor_sets, {});
+    vk::DescriptorSet& descriptor_set = pipeline->descriptor_sets[pipeline->descriptor_set_idx++];
+    vk::WriteDescriptorSet write_descriptor_set{ descriptor_set, 0, 0, pipeline->parameter_count, vk::DescriptorType::eStorageBuffer, nullptr, descriptor_buffer_infos.begin() };
+    ctx->device->device.updateDescriptorSets({ write_descriptor_set }, {});
 
     subctx->s->buffer.pushConstants(pipeline->layout, vk::ShaderStageFlagBits::eCompute, 0, push_constant_size, push_constants);
     subctx->s->buffer.bindPipeline(vk::PipelineBindPoint::eCompute, pipeline->pipeline);
@@ -2410,7 +2460,7 @@ static void ggml_vk_end_submission(vk_submission& s, std::vector<vk_semaphore> w
     s.signal_semaphores = std::move(signal_semaphores);
 }
 
-static void ggml_vk_ctx_end(vk_context * ctx) {
+static void ggml_vk_ctx_end(vk_context& ctx) {
     VK_LOG_DEBUG("ggml_vk_ctx_end(" << ctx << ", " << ctx->seqs.size() << ")");
     if (ctx->s == nullptr) {
         return;
@@ -2420,7 +2470,7 @@ static void ggml_vk_ctx_end(vk_context * ctx) {
     ctx->s = nullptr;
 }
 
-static void ggml_vk_ctx_begin(vk_device& device, vk_context * subctx) {
+static void ggml_vk_ctx_begin(vk_device& device, vk_context& subctx) {
     VK_LOG_DEBUG("ggml_vk_ctx_begin(" << device->name << ")");
     if (subctx->s != nullptr) {
         ggml_vk_ctx_end(subctx);
@@ -2453,13 +2503,13 @@ static void ggml_vk_ensure_sync_staging_buffer(vk_device& device, size_t size) {
     }
 }
 
-static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_context * subctx, vk_buffer& dst, size_t offset, const ggml_tensor * tensor, bool sync_staging = false) {
+static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_context& subctx, vk_buffer& dst, size_t offset, const ggml_tensor * tensor, bool sync_staging = false) {
     VK_LOG_DEBUG("ggml_vk_buffer_write_nc_async(" << tensor << ")");
     GGML_ASSERT(!ggml_is_contiguous(tensor));
     // Buffer is already mapped
     if(dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
         std::cerr << "ggml_vulkan: buffer_write_nc_async dst buffer is host_visible. Use synchronous write." << std::endl;
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
     // Check if src is pinned memory
     vk_buffer buf;
@@ -2527,7 +2577,7 @@ static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_cont
             staging = ctx->device->sync_staging;
             staging_offset = 0;
         } else {
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
     }
 
@@ -2558,12 +2608,12 @@ static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_cont
     }
 }
 
-static void ggml_vk_buffer_write_2d_async(vk_context * subctx, vk_buffer& dst, size_t offset, const void * src, size_t spitch, size_t width, size_t height, vk_buffer staging_buffer, size_t staging_offset, bool sync_staging = false) {
+static void ggml_vk_buffer_write_2d_async(vk_context subctx, vk_buffer& dst, size_t offset, const void * src, size_t spitch, size_t width, size_t height, vk_buffer staging_buffer, size_t staging_offset, bool sync_staging = false) {
     VK_LOG_DEBUG("ggml_vk_buffer_write_2d_async(" << width << ", " << height << ")");
     // Buffer is already mapped
     if(dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
         std::cerr << "ggml_vulkan: buffer_write_async dst buffer is host_visible. Use synchronous write." << std::endl;
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
     // Check if src is pinned memory
     vk_buffer buf = nullptr;
@@ -2602,7 +2652,7 @@ static void ggml_vk_buffer_write_2d_async(vk_context * subctx, vk_buffer& dst, s
             staging_buffer = dst->device->sync_staging;
             staging_offset = 0;
         } else {
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
     }
 
@@ -2623,7 +2673,7 @@ static void ggml_vk_buffer_write_2d_async(vk_context * subctx, vk_buffer& dst, s
     }
 }
 
-static void ggml_vk_buffer_write_async(vk_context * subctx, vk_buffer& dst, size_t offset, const void * src, size_t size, vk_buffer staging_buffer, size_t staging_offset, bool sync_staging = false) {
+static void ggml_vk_buffer_write_async(vk_context subctx, vk_buffer& dst, size_t offset, const void * src, size_t size, vk_buffer staging_buffer, size_t staging_offset, bool sync_staging = false) {
     VK_LOG_DEBUG("ggml_vk_buffer_write_async(" << size << ")");
     return ggml_vk_buffer_write_2d_async(subctx, dst, offset, src, size, size, 1, staging_buffer, staging_offset, sync_staging);
 }
@@ -2638,7 +2688,7 @@ static void ggml_vk_buffer_write_2d(vk_buffer& dst, size_t offset, const void *
             memcpy((uint8_t *)dst->ptr + offset + i * width, (const uint8_t *) src + i * spitch, width);
         }
     } else {
-        vk_context * subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue);
+        vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue);
         ggml_vk_ctx_begin(dst->device, subctx);
         ggml_vk_buffer_write_2d_async(subctx, dst, offset, src, spitch, width, height, nullptr, 0, true);
         ggml_vk_ctx_end(subctx);
@@ -2650,8 +2700,6 @@ static void ggml_vk_buffer_write_2d(vk_buffer& dst, size_t offset, const void *
         ggml_vk_submit(subctx, dst->device->fence);
         VK_CHECK(dst->device->device.waitForFences({ dst->device->fence }, true, UINT64_MAX), "vk_buffer_write_2d waitForFences");
         dst->device->device.resetFences({ dst->device->fence });
-
-        delete subctx;
     }
 }
 
@@ -2660,12 +2708,14 @@ static void ggml_vk_buffer_write(vk_buffer& dst, size_t offset, const void * src
     ggml_vk_buffer_write_2d(dst, offset, src, 0, size, 1);
 }
 
-static void ggml_vk_buffer_read_2d_async(vk_context * subctx, vk_buffer& src, size_t offset, void * dst, size_t spitch, size_t dpitch, size_t width, size_t height, vk_buffer staging_buffer, size_t staging_offset, bool sync_staging = false) {
+static void ggml_vk_buffer_read_2d_async(vk_context subctx, vk_buffer& src, size_t offset, void * dst, size_t spitch, size_t dpitch, size_t width, size_t height, vk_buffer staging_buffer, size_t staging_offset, bool sync_staging = false) {
     VK_LOG_DEBUG("ggml_vk_buffer_read_2d_async(offset=" << offset << ", width=" << width << ", height=" << height << ")");
     GGML_ASSERT(width > 0);
     GGML_ASSERT(height > 0);
     GGML_ASSERT(src != nullptr);
 
+    // TODO: staging_offset is not used
+
     // Check if dst is pinned memory
     vk_buffer buf = nullptr;
     size_t buf_offset;
@@ -2704,7 +2754,7 @@ static void ggml_vk_buffer_read_2d_async(vk_context * subctx, vk_buffer& src, si
 
             staging_buffer = src->device->sync_staging;
         } else {
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
     }
 
@@ -2714,18 +2764,18 @@ static void ggml_vk_buffer_read_2d_async(vk_context * subctx, vk_buffer& src, si
     deferred_memcpy(dst, staging_buffer->ptr, copy_size, &subctx->out_memcpys);
 }
 
-static void ggml_vk_buffer_read_async(vk_context * subctx, vk_buffer& src, size_t offset, void * dst, size_t size, vk_buffer staging_buffer, size_t staging_offset, bool sync_staging = false) {
+static void ggml_vk_buffer_read_async(vk_context subctx, vk_buffer& src, size_t offset, void * dst, size_t size, vk_buffer staging_buffer, size_t staging_offset, bool sync_staging = false) {
     return ggml_vk_buffer_read_2d_async(subctx, src, offset, dst, size, size, size, 1, staging_buffer, staging_offset, sync_staging);
 }
 
 static void ggml_vk_buffer_read(vk_buffer& src, size_t offset, void * dst, size_t size) {
-    VK_LOG_DEBUG("ggml_vk_buffer_read(" << offset << ", " << size << ")");
+    VK_LOG_DEBUG("ggml_vk_buffer_read(" << src->buffer << ", " << offset << ", " << size << ")");
     if(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
         GGML_ASSERT(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostCoherent);
 
         memcpy(dst, (uint8_t *) src->ptr + offset, size);
     } else {
-        vk_context * subctx = ggml_vk_create_temporary_context(src->device->transfer_queue);
+        vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue);
         ggml_vk_ctx_begin(src->device, subctx);
         ggml_vk_buffer_read_async(subctx, src, offset, dst, size, nullptr, 0, true);
         ggml_vk_ctx_end(subctx);
@@ -2737,12 +2787,10 @@ static void ggml_vk_buffer_read(vk_buffer& src, size_t offset, void * dst, size_
         for (auto& cpy : subctx->out_memcpys) {
             memcpy(cpy.dst, cpy.src, cpy.n);
         }
-
-        delete subctx;
     }
 }
 
-static void ggml_vk_buffer_copy_async(vk_context * ctx, vk_buffer& dst, size_t dst_offset, vk_buffer& src, size_t src_offset, size_t size) {
+static void ggml_vk_buffer_copy_async(vk_context& ctx, vk_buffer& dst, size_t dst_offset, vk_buffer& src, size_t src_offset, size_t size) {
     VK_LOG_DEBUG("ggml_vk_buffer_copy_async(" << size << ")");
     // Make sure both buffers are on same device
     GGML_ASSERT(src->device == dst->device);
@@ -2756,15 +2804,13 @@ static void ggml_vk_buffer_copy(vk_buffer& dst, size_t dst_offset, vk_buffer& sr
     if (src->device == dst->device) {
         VK_LOG_DEBUG("ggml_vk_buffer_copy(SINGLE_DEVICE, " << size << ")");
         // Copy within the device
-        vk_context * subctx = ggml_vk_create_temporary_context(src->device->transfer_queue);
+        vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue);
         ggml_vk_ctx_begin(src->device, subctx);
         ggml_vk_buffer_copy_async(subctx, dst, dst_offset, src, src_offset, size);
         ggml_vk_ctx_end(subctx);
         ggml_vk_submit(subctx, src->device->fence);
         VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX), "vk_buffer_copy waitForFences");
         src->device->device.resetFences({ src->device->fence });
-
-        delete subctx;
     } else {
         VK_LOG_DEBUG("ggml_vk_buffer_copy(MULTI_DEVICE, " << size << ")");
         // Copy device to device
@@ -2783,7 +2829,7 @@ static void ggml_vk_buffer_copy(vk_buffer& dst, size_t dst_offset, vk_buffer& sr
 static void ggml_vk_buffer_memset(vk_buffer& dst, size_t offset, uint32_t c, size_t size) {
     VK_LOG_DEBUG("ggml_vk_buffer_memset(" << offset << ", " << c << ", " << size << ")");
 
-    vk_context * subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue);
+    vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue);
     ggml_vk_ctx_begin(dst->device, subctx);
     subctx->s->buffer.fillBuffer(dst->buffer, offset, size, c);
     ggml_vk_ctx_end(subctx);
@@ -2791,8 +2837,6 @@ static void ggml_vk_buffer_memset(vk_buffer& dst, size_t offset, uint32_t c, siz
     ggml_vk_submit(subctx, dst->device->fence);
     VK_CHECK(dst->device->device.waitForFences({ dst->device->fence }, true, UINT64_MAX), "vk_memset waitForFences");
     dst->device->device.resetFences({ dst->device->fence });
-
-    delete subctx;
 }
 
 static uint32_t ggml_vk_guess_split_k(int m, int n, int k) {
@@ -2855,7 +2899,7 @@ static uint32_t ggml_vk_guess_matmul_pipeline_align(ggml_backend_vk_context * ct
 }
 
 static void ggml_vk_matmul(
-        ggml_backend_vk_context * ctx, vk_context * subctx, vk_pipeline& pipeline,
+        ggml_backend_vk_context * ctx, vk_context& subctx, vk_pipeline& pipeline,
         vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& split_k_buffer,
         uint32_t m, uint32_t n, uint32_t k, uint32_t stride_a, uint32_t stride_b, uint32_t stride_d,
         uint32_t batch_stride_a, uint32_t batch_stride_b, uint32_t batch_stride_d,
@@ -2879,7 +2923,7 @@ static void ggml_vk_matmul(
 }
 
 static void ggml_vk_matmul_id(
-        ggml_backend_vk_context * ctx, vk_context * subctx, vk_pipeline& pipeline,
+        ggml_backend_vk_context * ctx, vk_context& subctx, vk_pipeline& pipeline,
         vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& ids,
         uint32_t m, uint32_t n, uint32_t k, uint32_t stride_a, uint32_t stride_b, uint32_t stride_d,
         uint32_t batch_stride_a, uint32_t batch_stride_b, uint32_t batch_stride_d,
@@ -2913,10 +2957,10 @@ static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, ggml_
     }
 
     std::cerr << "Missing CPY op for types: " << ggml_type_name(from) << " " << ggml_type_name(to) << std::endl;
-    GGML_ASSERT(false);
+    GGML_ABORT("fatal error");
 }
 
-static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context * subctx, vk_pipeline pipeline, const ggml_tensor * tensor, vk_subbuffer&& in, vk_subbuffer&& out) {
+static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context& subctx, vk_pipeline pipeline, const ggml_tensor * tensor, vk_subbuffer&& in, vk_subbuffer&& out) {
     VK_LOG_DEBUG("ggml_vk_cpy_to_contiguous((" << tensor << ", type=" << tensor->type << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << "), ";
     std::cerr << "buffer in size=" << in.buffer->size << ", buffer out size=" << out.buffer->size << ")");
     const int tensor_type_size = ggml_type_size(tensor->type);
@@ -2934,7 +2978,7 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, sizeof(vk_op_unary_push_constants), &pc, { ne, 1, 1 });
 }
 
-static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     VK_LOG_DEBUG("ggml_vk_mul_mat_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
     std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
     std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)");
@@ -3079,7 +3123,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su
     } else if (qx_needs_dequant) {
         const std::vector<uint32_t> pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) };
         ggml_vk_sync_buffers(subctx);
-        ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, { { d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, { d_X, 0, x_sz * ne02 * ne03 } }, pc.size() * sizeof(uint32_t), pc.data(), { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
+        ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc.size() * sizeof(uint32_t), pc.data(), { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
     }
     if (y_non_contig) {
         ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
@@ -3107,7 +3151,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su
     );  // NOLINT
 }
 
-static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     VK_LOG_DEBUG("ggml_vk_mul_mat_vec_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
     std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
     std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)");
@@ -3268,11 +3312,11 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context
     };
     ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
-                              { { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz * ne12 * ne13 }, { d_D, d_buf_offset, d_sz * ne22 * ne23} },
+                              { vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 }, vk_subbuffer{ d_Y, y_buf_offset, y_sz * ne12 * ne13 }, vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23} },
                               sizeof(vk_mat_vec_push_constants), &pc, { groups_x, (uint32_t)(ne12 * ne13), groups_z });
 }
 
-static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     VK_LOG_DEBUG("ggml_vk_mul_mat_p021_f16_f32(" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
     std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
     std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)");
@@ -3340,10 +3384,10 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
     // compute
     const std::array<uint32_t, 6> pc = { (uint32_t)ne00, (uint32_t)ne01, (uint32_t)ne02, (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
     ggml_vk_sync_buffers(subctx);
-    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32, { { d_Qx, qx_buf_offset, qx_sz }, { d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, { d_D, d_buffer_offset, d_sz + d_shader_offset } }, 6 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
+    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, 6 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
 }
 
-static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     VK_LOG_DEBUG("ggml_vk_mul_mat_nc_f16_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
     std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
     std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)");
@@ -3415,10 +3459,11 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
     // compute
     const std::array<uint32_t, 7> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, (uint32_t)(ne12 / ne02), (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
     ggml_vk_sync_buffers(subctx);
-    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32, { { d_Qx, qx_buf_offset, qx_sz }, { d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, { d_D, d_buffer_offset, d_sz + d_shader_offset } }, 7 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
+    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32,
+        { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, 7 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
 }
 
-static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     VK_LOG_DEBUG("ggml_vk_mul_mat(" << src0 << ", " << src1 << ", " << dst << ")");
     if (src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && dst->ne[1] == 1) {
         ggml_vk_mul_mat_vec_p021_f16_f32(ctx, subctx, src0, src1, dst);
@@ -3431,7 +3476,7 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context * subctx,
     }
 }
 
-static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
+static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
     VK_LOG_DEBUG("ggml_vk_mul_mat_id_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
     std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
     std::cerr << "), (" << ids << ", name=" << ids->name << ", type=" << ids->type << ", ne0=" << ids->ne[0] << ", ne1=" << ids->ne[1] << ", ne2=" << ids->ne[2] << ", ne3=" << ids->ne[3] << ", nb0=" << ids->nb[0] << ", nb1=" << ids->nb[1] << ", nb2=" << ids->nb[2] << ", nb3=" << ids->nb[3];
@@ -3499,7 +3544,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context *
     const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig;
 
     if (mmp == nullptr) {
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 
     // Not implemented
@@ -3590,7 +3635,8 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context *
     } else if (qx_needs_dequant) {
         const std::vector<uint32_t> pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) };
         ggml_vk_sync_buffers(subctx);
-        ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, { { d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, { d_X, 0, x_sz * ne02 * ne03 } }, pc.size() * sizeof(uint32_t), pc.data(), { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
+        ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0,
+            { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc.size() * sizeof(uint32_t), pc.data(), { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
     }
     if (y_non_contig) {
         ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
@@ -3618,7 +3664,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context *
     );  // NOLINT
 }
 
-static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
+static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
     VK_LOG_DEBUG("ggml_vk_mul_mat_vec_id_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
     std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
     std::cerr << "), (" << ids << ", name=" << ids->name << ", type=" << ids->type << ", ne0=" << ids->ne[0] << ", ne1=" << ids->ne[1] << ", ne2=" << ids->ne[2] << ", ne3=" << ids->ne[3] << ", nb0=" << ids->nb[0] << ", nb1=" << ids->nb[1] << ", nb2=" << ids->nb[2] << ", nb3=" << ids->nb[3];
@@ -3790,11 +3836,12 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
     };
     ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
-        { { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz * ne12 * ne13 }, { d_D, d_buf_offset, d_sz * ne22 * ne23}, { d_ids, ids_buf_offset, ids_sz } },
+        { vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 },
+        vk_subbuffer{ d_Y, y_buf_offset, y_sz * ne12 * ne13 }, vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23}, vk_subbuffer{ d_ids, ids_buf_offset, ids_sz } },
         sizeof(vk_mat_vec_id_push_constants), &pc, { groups_x, (uint32_t)nei0, groups_z });
 }
 
-static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
     VK_LOG_DEBUG("ggml_vk_mul_mat_id(" << src0 << ", " << src1 << ", " << src2 << ", " << dst << ")");
     if (src2->ne[1] == 1 && (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) {
         ggml_vk_mul_mat_vec_id_q_f16(ctx, subctx, src0, src1, src2, dst);
@@ -3803,8 +3850,8 @@ static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context * subct
     }
 }
 
-static void ggml_vk_op_repeat(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    // guaranteed to be an integer due to the check in ggml_can_repeat
+static void ggml_vk_op_repeat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    VK_LOG_DEBUG("ggml_vk_op_repeat(" << src0 << ", " << src1 << ", " << dst << ")");
     const uint64_t ne0 = dst->ne[0];
     const uint64_t ne1 = dst->ne[1];
     const uint64_t ne2 = dst->ne[2];
@@ -3825,6 +3872,7 @@ static void ggml_vk_op_repeat(ggml_backend_vk_context * ctx, vk_context * subctx
     const uint64_t nb02 = src0->nb[2];
     const uint64_t nb03 = src0->nb[3];
 
+    // guaranteed to be an integer due to the check in ggml_can_repeat
     const uint64_t nr0 = ne0/ne00;
     const uint64_t nr1 = ne1/ne01;
     const uint64_t nr2 = ne2/ne02;
@@ -3852,8 +3900,8 @@ static void ggml_vk_op_repeat(ggml_backend_vk_context * ctx, vk_context * subctx
                         for     (uint64_t k1 = 0; k1 < ne01; k1++) {
                             for (uint64_t i0 = 0; i0 < nr0;  i0++) {
                                 copies.push_back({
-                                    src_offset + (i3*ne03 + k3)*nb3  + (i2*ne02 + k2)*nb2  + (i1*ne01 + k1)*nb1  + (i0*ne00)*nb0,
-                                    dst_offset + (          k3)*nb03 + (          k2)*nb02 + (          k1)*nb01,
+                                    src_offset + (          k3)*nb03 + (          k2)*nb02 + (          k1)*nb01,
+                                    dst_offset + (i3*ne03 + k3)*nb3  + (i2*ne02 + k2)*nb2  + (i1*ne01 + k1)*nb1 + (i0*ne00)*nb0,
                                     ne00*nb0,
                                 });
                             }
@@ -3874,11 +3922,6 @@ static void ggml_vk_op_repeat(ggml_backend_vk_context * ctx, vk_context * subctx
 
 static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op) {
     switch (op) {
-    case GGML_OP_ADD:
-        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_add_f32;
-        }
-        return nullptr;
     case GGML_OP_GET_ROWS:
         GGML_ASSERT(src1->type == GGML_TYPE_I32);
         if (dst->type == GGML_TYPE_F16) {
@@ -3888,6 +3931,14 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_get_rows_f32[src0->type];
         }
         return nullptr;
+    case GGML_OP_ADD:
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_add_f32;
+        }
+        if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) {
+            return ctx->device->pipeline_add_f16_f32_f16;
+        }
+        return nullptr;
     case GGML_OP_MUL:
         if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_mul_f32;
@@ -3898,6 +3949,22 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_div_f32;
         }
         return nullptr;
+    case GGML_OP_CONCAT:
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_concat_f32;
+        }
+        if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
+            return ctx->device->pipeline_concat_f16;
+        }
+        if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32 && dst->type == GGML_TYPE_I32) {
+            return ctx->device->pipeline_concat_i32;
+        }
+        return nullptr;
+    case GGML_OP_UPSCALE:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_upscale_f32;
+        }
+        return nullptr;
     case GGML_OP_SCALE:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_scale_f32;
@@ -3913,6 +3980,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_clamp_f32;
         }
         return nullptr;
+    case GGML_OP_PAD:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_pad_f32;
+        }
+        return nullptr;
     case GGML_OP_CPY:
     case GGML_OP_CONT:
     case GGML_OP_DUP:
@@ -3922,6 +3994,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_norm_f32;
         }
         return nullptr;
+    case GGML_OP_GROUP_NORM:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_group_norm_f32;
+        }
+        return nullptr;
     case GGML_OP_RMS_NORM:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_rms_norm_f32;
@@ -3939,11 +4016,21 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
                     return ctx->device->pipeline_gelu_f32;
                 }
                 break;
+            case GGML_UNARY_OP_GELU_QUICK:
+                if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+                    return ctx->device->pipeline_gelu_quick_f32;
+                }
+                break;
             case GGML_UNARY_OP_RELU:
                 if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
                     return ctx->device->pipeline_relu_f32;
                 }
                 break;
+            case GGML_UNARY_OP_TANH:
+                if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+                    return ctx->device->pipeline_tanh_f32;
+                }
+                break;
             default:
                 break;
         }
@@ -3995,6 +4082,24 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_sum_rows_f32;
         }
         return nullptr;
+    case GGML_OP_IM2COL:
+        if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_im2col_f32;
+        }
+        if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) {
+            return ctx->device->pipeline_im2col_f32_f16;
+        }
+        return nullptr;
+    case GGML_OP_TIMESTEP_EMBEDDING:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_timestep_embedding_f32;
+        }
+        return nullptr;
+    case GGML_OP_LEAKY_RELU:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_leaky_relu_f32;
+        }
+        return nullptr;
     default:
         return nullptr;
     }
@@ -4018,9 +4123,12 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
     case GGML_OP_ADD:
     case GGML_OP_MUL:
     case GGML_OP_DIV:
+    case GGML_OP_CONCAT:
+    case GGML_OP_UPSCALE:
     case GGML_OP_SCALE:
     case GGML_OP_SQR:
     case GGML_OP_CLAMP:
+    case GGML_OP_PAD:
         return true;
     default:
         return false;
@@ -4028,7 +4136,7 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
 }
 
 template<typename PC>
-static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op, const PC&& pc) {
+static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op, const PC&& pc) {
     VK_LOG_DEBUG("ggml_vk_op_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
     if (src1 != nullptr) {
         std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
@@ -4078,7 +4186,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
                 std::cerr << " and " << ggml_type_name(src1->type);
             }
             std::cerr << " to " << ggml_type_name(dst->type) << std::endl;
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
 
         op_func(ctx, subctx, src0, src1, dst);
@@ -4124,7 +4232,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
     vk_buffer d_D = extra->buffer_gpu.lock();
 
     // Workaround for tiny tensor inputs on ROPE
-    if (use_src1 && y_sz > d_D->size) {
+    if (op == GGML_OP_ROPE && use_src1 && y_sz > d_D->size) {
         y_sz = VK_WHOLE_SIZE;
     }
 
@@ -4173,13 +4281,26 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
     if (op_supports_incontiguous || (ggml_is_contiguous(src0) && (src1 == nullptr || ggml_is_contiguous(src1)))) {
         ggml_pipeline_allocate_descriptor_sets(ctx->device, pipeline, 1);
 
-        switch (dst->op) {
+        switch (op) {
         case GGML_OP_NORM:
         case GGML_OP_RMS_NORM:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_SUM_ROWS:
-            elements = { (uint32_t)ggml_nrows(src0), 1, 1 };
-            break;
+            {
+                const uint32_t nr = ggml_nrows(src0);
+                if (nr > 262144) {
+                    elements = { 512, 512, CEIL_DIV(nr, 262144) };
+                } else if (nr > 512) {
+                    elements = { 512, CEIL_DIV(nr, 512), 1 };
+                } else {
+                    elements = { nr, 1, 1 };
+                }
+            } break;
+        case GGML_OP_GROUP_NORM:
+            {
+                const uint32_t num_groups = dst->op_params[0];
+                elements = { num_groups * (uint32_t)src0->ne[3], 1, 1 };
+            } break;
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_ROPE:
             elements = { (uint32_t)ggml_nrows(src0), (uint32_t)ne00, 1 };
@@ -4190,6 +4311,49 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
         case GGML_OP_ARGSORT:
             elements = { (uint32_t)ne00, (uint32_t)ggml_nrows(src0), 1 };
             break;
+        case GGML_OP_IM2COL:
+            {
+                const bool is_2D = dst->op_params[6] == 1;
+
+                const uint32_t IC = src1->ne[is_2D ? 2 : 1];
+
+                const uint32_t KH = is_2D ? src0->ne[1] : 1;
+                const uint32_t KW =         src0->ne[0];
+
+                const uint32_t OH = is_2D ? dst->ne[2] : 1;
+                const uint32_t OW =         dst->ne[1];
+
+                const uint32_t batch = src1->ne[3];
+
+                elements = { OW * KW * KH, OH, batch * IC };
+            } break;
+        case GGML_OP_TIMESTEP_EMBEDDING:
+            {
+                const uint32_t dim = dst->op_params[0];
+                uint32_t half_ceil = (dim + 1) / 2;
+                elements = { half_ceil, (uint32_t)src0->ne[0], 1 };
+            } break;
+        case GGML_OP_ADD:
+        case GGML_OP_DIV:
+        case GGML_OP_MUL:
+        case GGML_OP_SCALE:
+        case GGML_OP_SQR:
+        case GGML_OP_CLAMP:
+        case GGML_OP_PAD:
+        case GGML_OP_CPY:
+        case GGML_OP_CONCAT:
+        case GGML_OP_UPSCALE:
+        case GGML_OP_UNARY:
+            {
+                const uint32_t ne = ggml_nelements(dst);
+                if (ne > 262144) {
+                    elements = { 512, 512, CEIL_DIV(ne, 262144) };
+                } else if (ne > 512) {
+                    elements = { 512, CEIL_DIV(ne, 512), 1 };
+                } else {
+                    elements = { ne, 1, 1 };
+                }
+            } break;
         default:
             elements = { (uint32_t)ggml_nelements(src0), 1, 1 };
             break;
@@ -4216,31 +4380,35 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
             if (use_src1) {
                 subbuf_y = { d_Y, y_buf_offset, y_sz };
             } else {
-                subbuf_y = { d_X, 0, d_X->size };
+                subbuf_y = { d_X, 0, x_sz };
             }
 
             ggml_vk_sync_buffers(subctx);
-            ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset, x_sz }, subbuf_y, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+            ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, subbuf_y, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
         } else if (op == GGML_OP_ROPE) {
             // Empty src2 is possible in rope, but the shader needs a buffer
             vk_subbuffer subbuf_z;
             if (use_src2) {
                 subbuf_z = { d_Z, z_buf_offset, z_sz };
             } else {
-                subbuf_z = { d_X, 0, d_X->size };
+                subbuf_z = { d_X, 0, x_sz };
             }
 
             ggml_vk_sync_buffers(subctx);
-            ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset, x_sz }, { d_Y, y_buf_offset, y_sz }, subbuf_z, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+            ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+        } else if (op == GGML_OP_IM2COL) {
+            // im2col uses only src1 and dst buffers
+            ggml_vk_sync_buffers(subctx);
+            ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
         } else if (use_src2) {
             ggml_vk_sync_buffers(subctx);
-            ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset, x_sz }, { d_Y, y_buf_offset, y_sz }, { d_Z, z_buf_offset, z_sz }, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+            ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
         } else if (use_src1) {
             ggml_vk_sync_buffers(subctx);
-            ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset, x_sz }, { d_Y, y_buf_offset, y_sz }, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+            ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
         } else {
             ggml_vk_sync_buffers(subctx);
-            ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset, x_sz }, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+            ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
         }
     } else {
         GGML_ASSERT(op != GGML_OP_SOFT_MAX);
@@ -4249,8 +4417,9 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
 
         ggml_pipeline_allocate_descriptor_sets(ctx->device, pipeline, ne02 * ne03);
 
-        switch (dst->op) {
+        switch (op) {
         case GGML_OP_NORM:
+        case GGML_OP_GROUP_NORM:
         case GGML_OP_RMS_NORM:
             elements = { (uint32_t)ne01, 1, 1 };
             break;
@@ -4276,21 +4445,21 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
 
                 if (use_src1) {
                     ggml_vk_sync_buffers(subctx);
-                    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset + x_offset, x_sz }, { d_Y, y_buf_offset + y_offset, y_sz }, { d_D, d_buf_offset + d_offset, d_sz } }, sizeof(PC), &pc, elements);
+                    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset + x_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset + y_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset + d_offset, d_sz } }, sizeof(PC), &pc, elements);
                 } else {
                     ggml_vk_sync_buffers(subctx);
-                    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset + x_offset, x_sz }, { d_D, d_buf_offset + d_offset, d_sz } }, sizeof(PC), &pc, elements);
+                    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset + x_offset, x_sz }, vk_subbuffer{ d_D, d_buf_offset + d_offset, d_sz } }, sizeof(PC), &pc, elements);
                 }
             }
         }
     }
 }
 
-static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_REPEAT, { (uint32_t)ggml_nelements(src0), (uint32_t)ggml_nelements(src1), 0.0f, 0.0f });
+static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_REPEAT, {});
 }
 
-static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
@@ -4301,11 +4470,11 @@ static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context * subctx,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
-        0.0f, 0.0f,
+        0.0f, 0.0f, 0,
     });
 }
 
-static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
@@ -4316,11 +4485,11 @@ static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context * subctx, cons
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
-        0.0f, 0.0f,
+        0.0f, 0.0f, 0,
     });
 }
 
-static void ggml_vk_mul(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+static void ggml_vk_mul(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
@@ -4331,11 +4500,11 @@ static void ggml_vk_mul(ggml_backend_vk_context * ctx, vk_context * subctx, cons
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
-        0.0f, 0.0f,
+        0.0f, 0.0f, 0,
     });
 }
 
-static void ggml_vk_div(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+static void ggml_vk_div(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
@@ -4346,11 +4515,44 @@ static void ggml_vk_div(ggml_backend_vk_context * ctx, vk_context * subctx, cons
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
-        0.0f, 0.0f,
+        0.0f, 0.0f, 0,
     });
 }
 
-static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+static void ggml_vk_concat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    int * op_params = (int *)dst->op_params;
+
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t src1_type_size = ggml_type_size(src1->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONCAT, {
+        (uint32_t)ggml_nelements(dst),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        0.0f, 0.0f, op_params[0],
+    });
+}
+
+static void ggml_vk_upscale(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+
+    const float sf0 = (float)dst->ne[0] / src0->ne[0];
+    const float sf1 = (float)dst->ne[1] / src0->ne[1];
+    const float sf2 = (float)dst->ne[2] / src0->ne[2];
+    const float sf3 = (float)dst->ne[3] / src0->ne[3];
+
+    ggml_vk_op_f32<vk_op_upscale_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_UPSCALE, {
+        (uint32_t)ggml_nelements(dst), 0,
+        (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t)dst->ne[0], (uint32_t)dst->ne[1], (uint32_t)dst->ne[2],(uint32_t)dst->ne[3],
+        sf0, sf1, sf2, sf3,
+    });
+}
+
+static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     float * op_params = (float *)dst->op_params;
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
@@ -4364,7 +4566,7 @@ static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context * subctx, co
     });
 }
 
-static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
@@ -4377,7 +4579,7 @@ static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context * subctx, cons
     });
 }
 
-static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     float * op_params = (float *)dst->op_params;
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
@@ -4391,7 +4593,20 @@ static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context * subctx, co
     });
 }
 
-static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+static void ggml_vk_pad(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_PAD, {
+        (uint32_t)ggml_nelements(dst),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        0.0f, 0.0f,
+    });
+}
+
+static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra;
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
@@ -4406,27 +4621,37 @@ static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context * subctx, cons
     });
 }
 
-static void ggml_vk_norm(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+static void ggml_vk_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     float * op_params = (float *)dst->op_params;
 
     ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f });
 }
 
-static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+static void ggml_vk_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    int * op_params = (int *)dst->op_params;
+
+    uint32_t num_groups = op_params[0];
+    uint32_t group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
+    static const float eps = 1e-6f;
+
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_GROUP_NORM, { group_size, 0, eps, 0.0f });
+}
+
+static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     float * op_params = (float *)dst->op_params;
     ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_RMS_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f });
 }
 
-static void ggml_vk_unary(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+static void ggml_vk_unary(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_UNARY, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f });
 }
 
-static void ggml_vk_diag_mask_inf(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+static void ggml_vk_diag_mask_inf(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     int32_t * op_params = (int32_t *)dst->op_params;
     ggml_vk_op_f32<vk_op_diag_mask_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_DIAG_MASK_INF, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0] });
 }
 
-static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     float * op_params = (float *)dst->op_params;
 
     float scale = op_params[0];
@@ -4451,7 +4676,7 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context * subctx,
     });
 }
 
-static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
     const int n_dims        = ((int32_t *) dst->op_params)[1];
     // const int mode          = ((int32_t *) dst->op_params)[2];
     // const int n_ctx         = ((int32_t *) dst->op_params)[3];
@@ -4475,7 +4700,7 @@ static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context * subctx, con
     });
 }
 
-static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     int32_t * op_params = (int32_t *)dst->op_params;
 
     uint32_t ncols = src0->ne[0];
@@ -4494,10 +4719,59 @@ static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context * subctx,
     });
 }
 
-static void ggml_vk_sum_rows(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+static void ggml_vk_sum_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SUM_ROWS, { (uint32_t)src0->ne[0], 0, 0.0f, 0.0f });
 }
 
+static void ggml_vk_im2col(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    const int32_t s0 = dst->op_params[0];
+    const int32_t s1 = dst->op_params[1];
+    const int32_t p0 = dst->op_params[2];
+    const int32_t p1 = dst->op_params[3];
+    const int32_t d0 = dst->op_params[4];
+    const int32_t d1 = dst->op_params[5];
+
+    const bool is_2D = dst->op_params[6] == 1;
+
+    const uint32_t IC = src1->ne[is_2D ? 2 : 1];
+    const uint32_t IH = is_2D ? src1->ne[1] : 1;
+    const uint32_t IW =         src1->ne[0];
+
+    const uint32_t KH = is_2D ? src0->ne[1] : 1;
+    const uint32_t KW =         src0->ne[0];
+
+    const uint32_t OH = is_2D ? dst->ne[2] : 1;
+    const uint32_t OW =         dst->ne[1];
+
+    const uint32_t offset_delta = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
+    const uint32_t batch_offset = src1->nb[3] / 4; // nb is byte offset, src is type float32
+
+    const uint32_t pelements = OW * KW * KH;
+
+    ggml_vk_op_f32<vk_op_im2col_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_IM2COL, {
+        batch_offset, offset_delta,
+        IC, IW, IH, OW, OH, KW, KH,
+        pelements,
+        IC * KH * KW,
+        s0, s1, p0, p1, d0, d1,
+    });
+}
+
+static void ggml_vk_timestep_embedding(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    const uint32_t dim = dst->op_params[0];
+    const uint32_t max_period = dst->op_params[1];
+    const uint32_t nb1 = dst->nb[1] / ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_timestep_embedding_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_TIMESTEP_EMBEDDING, {
+        nb1, dim, max_period,
+    });
+}
+
+static void ggml_vk_leaky_relu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    const float * op_params = (const float *)dst->op_params;
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f });
+}
+
 #ifdef GGML_VULKAN_RUN_TESTS
 static void ggml_vk_print_matrix_area(const void * data, ggml_type type, int ne0, int ne1, int i0, int i1, int i2) {
     if (type != GGML_TYPE_F32 && type != GGML_TYPE_F16) {
@@ -4521,7 +4795,7 @@ static void ggml_vk_print_matrix_area(const void * data, ggml_type type, int ne0
                 } else if (type == GGML_TYPE_F16) {
                     val = ggml_fp16_to_fp32(*((const ggml_fp16_t *) data + i2*ne1*ne0 + idx1*ne0 + idx0));
                 } else {
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
                 }
                 fprintf(stderr, "% 7.2f ", val);
             } else {
@@ -4555,7 +4829,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
             p = ctx->device->pipeline_matmul_f16->a_s;
             shname = "F16_ALIGNED_S";
         } else {
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
     } else if (shader_size == 1) {
         if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
@@ -4571,7 +4845,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
             p = ctx->device->pipeline_matmul_f16->a_m;
             shname = "F16_ALIGNED_M";
         } else {
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
     } else if (shader_size == 2) {
         if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
@@ -4587,7 +4861,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
             p = ctx->device->pipeline_matmul_f16->a_l;
             shname = "F16_ALIGNED_L";
         } else {
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
     } else {
         GGML_ASSERT(0);
@@ -4668,7 +4942,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
         } else if (std::is_same<ggml_fp16_t, X_TYPE>()) {
             x[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f);
         } else {
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
     }
     for (size_t i = 0; i < y_ne; i++) {
@@ -4679,14 +4953,14 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
             // y[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f);
             y[i] = ggml_fp32_to_fp16((i % k == i / k) ? 1.0f : 0.0f);
         } else {
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
     }
 
     ggml_vk_buffer_write(d_X, 0, x, sizeof(X_TYPE) * k * m * batch);
     ggml_vk_buffer_write(d_Y, 0, y, sizeof(Y_TYPE) * k * n * batch);
 
-    vk_context * subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+    vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
     for (size_t i = 0; i < num_it; i++) {
         ggml_vk_ctx_begin(ctx->device, subctx);
         ggml_vk_matmul(
@@ -4727,14 +5001,14 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
     } else if (std::is_same<ggml_fp16_t, X_TYPE>()) {
         src0_type = GGML_TYPE_F16;
     } else {
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
     if (std::is_same<float, Y_TYPE>()) {
         src1_type = GGML_TYPE_F32;
     } else if (std::is_same<ggml_fp16_t, Y_TYPE>()) {
         src1_type = GGML_TYPE_F16;
     } else {
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 
     ggml_tensor * src0_ggml = ggml_new_tensor_3d(ggml_ctx, src0_type, k, m, batch);
@@ -4841,7 +5115,7 @@ static void ggml_vk_print_tensor_area(const ggml_tensor * tensor, int i0, int i1
                 } else if (tensor->type == GGML_TYPE_F16) {
                     val = ggml_fp16_to_fp32(*(ggml_fp16_t *) ((char *) tensor->data + i3*tensor->nb[3] + i2*tensor->nb[2] + idx1*tensor->nb[1] + idx0*tensor->nb[0]));
                 } else {
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
                 }
                 fprintf(stderr, "% 7.2f ", val);
             } else {
@@ -4894,7 +5168,7 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_
 
     ggml_vk_buffer_write(qx_buf, 0, qx, qx_sz);
 
-    vk_context * subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+    vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
     ggml_vk_ctx_begin(ctx->device, subctx);
     const std::vector<uint32_t> pc = { 1, (uint32_t)ne, (uint32_t)ne, (uint32_t)ne, (uint32_t)ne };
     ggml_vk_dispatch_pipeline(ctx, subctx, p, { { qx_buf, 0, qx_sz }, { x_buf, 0, x_sz_f16 } }, pc.size() * sizeof(int), pc.data(), { (uint32_t)ne, 1, 1});
@@ -5027,7 +5301,7 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
     ggml_vk_buffer_write(qx_buf, 0, qx, qx_sz);
     ggml_vk_buffer_write(y_buf, 0, y, y_sz);
 
-    vk_context * subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+    vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
     for (size_t i = 0; i < num_it; i++) {
         ggml_vk_ctx_begin(ctx->device, subctx);
         ggml_vk_matmul(
@@ -5175,7 +5449,7 @@ static void ggml_vk_preallocate_buffers_graph(ggml_backend_vk_context * ctx, ggm
 
     const bool y_f32_kernel = use_src1 && src1->type == GGML_TYPE_F32 && !y_non_contig;
 
-    bool mmp = (use_src0 && use_src1 && src1_type == GGML_TYPE_F32) ? ggml_vk_get_mul_mat_mat_pipeline(ctx, src0_type, y_non_contig ? GGML_TYPE_F16 : src1->type) != nullptr : false;
+    bool mmp = (use_src0 && use_src1 && (node->op == GGML_OP_MUL_MAT || node->op == GGML_OP_MUL_MAT_ID)) ? ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type) != nullptr : false;
 
     const bool qx_needs_dequant = use_src0 && (!mmp || x_non_contig);
     const bool qy_needs_dequant = use_src1 && ((src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig);
@@ -5211,24 +5485,33 @@ static void ggml_vk_preallocate_buffers_graph(ggml_backend_vk_context * ctx, ggm
     case GGML_OP_SCALE:
     case GGML_OP_SQR:
     case GGML_OP_CLAMP:
+    case GGML_OP_PAD:
     case GGML_OP_CPY:
     case GGML_OP_CONT:
     case GGML_OP_DUP:
     case GGML_OP_MUL:
     case GGML_OP_DIV:
+    case GGML_OP_CONCAT:
+    case GGML_OP_UPSCALE:
     case GGML_OP_NORM:
+    case GGML_OP_GROUP_NORM:
     case GGML_OP_RMS_NORM:
     case GGML_OP_DIAG_MASK_INF:
     case GGML_OP_SOFT_MAX:
     case GGML_OP_ROPE:
     case GGML_OP_ARGSORT:
     case GGML_OP_SUM_ROWS:
+    case GGML_OP_IM2COL:
+    case GGML_OP_TIMESTEP_EMBEDDING:
+    case GGML_OP_LEAKY_RELU:
         break;
     case GGML_OP_UNARY:
         switch (ggml_get_unary_op(node)) {
         case GGML_UNARY_OP_SILU:
         case GGML_UNARY_OP_GELU:
+        case GGML_UNARY_OP_GELU_QUICK:
         case GGML_UNARY_OP_RELU:
+        case GGML_UNARY_OP_TANH:
             break;
         default:
             return;
@@ -5236,6 +5519,13 @@ static void ggml_vk_preallocate_buffers_graph(ggml_backend_vk_context * ctx, ggm
         break;
     case GGML_OP_MUL_MAT:
     case GGML_OP_MUL_MAT_ID:
+        if (
+                x_sz > ctx->device->max_memory_allocation_size ||
+                y_sz > ctx->device->max_memory_allocation_size ||
+                d_sz > ctx->device->max_memory_allocation_size ||
+                split_k_size > ctx->device->max_memory_allocation_size) {
+            GGML_ABORT("Requested preallocation size is too large");
+        }
         if (ctx->prealloc_size_x < x_sz) {
             ctx->prealloc_size_x = x_sz;
         }
@@ -5391,7 +5681,7 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
         std::cerr << std::endl;
     }
 
-    GGML_ASSERT(false);
+    GGML_ABORT("fatal error");
 #endif
 
     if (ctx->prealloc_x == nullptr || (ctx->prealloc_size_x > 0 && ctx->prealloc_x->size < ctx->prealloc_size_x)) {
@@ -5430,7 +5720,7 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
     }
 }
 
-static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * node, bool last_node){
+static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * node, int node_idx, bool last_node){
     ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) node->extra;
 
     if (ggml_is_empty(node) || extra == nullptr) {
@@ -5457,7 +5747,9 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
         switch (ggml_get_unary_op(node)) {
         case GGML_UNARY_OP_SILU:
         case GGML_UNARY_OP_GELU:
+        case GGML_UNARY_OP_GELU_QUICK:
         case GGML_UNARY_OP_RELU:
+        case GGML_UNARY_OP_TANH:
             break;
         default:
             return;
@@ -5468,13 +5760,17 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_ADD:
     case GGML_OP_MUL:
     case GGML_OP_DIV:
+    case GGML_OP_CONCAT:
+    case GGML_OP_UPSCALE:
     case GGML_OP_SCALE:
     case GGML_OP_SQR:
     case GGML_OP_CLAMP:
+    case GGML_OP_PAD:
     case GGML_OP_CPY:
     case GGML_OP_CONT:
     case GGML_OP_DUP:
     case GGML_OP_NORM:
+    case GGML_OP_GROUP_NORM:
     case GGML_OP_RMS_NORM:
     case GGML_OP_DIAG_MASK_INF:
     case GGML_OP_SOFT_MAX:
@@ -5483,109 +5779,147 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_MUL_MAT_ID:
     case GGML_OP_ARGSORT:
     case GGML_OP_SUM_ROWS:
+    case GGML_OP_IM2COL:
+    case GGML_OP_TIMESTEP_EMBEDDING:
+    case GGML_OP_LEAKY_RELU:
         break;
     default:
         std::cerr << "ggml_vulkan: Error: Missing op: " << ggml_op_name(node->op) << std::endl;
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
         return;
     }
 
-    if (ctx->compute_ctx == nullptr) {
-        ctx->compute_ctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
-        ggml_vk_ctx_begin(ctx->device, ctx->compute_ctx);
+    vk_context compute_ctx;
+
+    if (ctx->compute_ctx.expired()) {
+        compute_ctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+        ctx->compute_ctx = compute_ctx;
+        ggml_vk_ctx_begin(ctx->device, compute_ctx);
+    } else {
+        compute_ctx = ctx->compute_ctx.lock();
     }
 
     switch (node->op) {
     case GGML_OP_REPEAT:
-        ggml_vk_repeat(ctx, ctx->compute_ctx, src0, src1, node);
+        ggml_vk_repeat(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_GET_ROWS:
-        ggml_vk_get_rows(ctx, ctx->compute_ctx, src0, src1, node);
+        ggml_vk_get_rows(ctx, compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_ADD:
-        ggml_vk_add(ctx, ctx->compute_ctx, src0, src1, node);
+        ggml_vk_add(ctx, compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_MUL:
-        ggml_vk_mul(ctx, ctx->compute_ctx, src0, src1, node);
+        ggml_vk_mul(ctx, compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_DIV:
-        ggml_vk_div(ctx, ctx->compute_ctx, src0, src1, node);
+        ggml_vk_div(ctx, compute_ctx, src0, src1, node);
+
+        break;
+    case GGML_OP_CONCAT:
+        ggml_vk_concat(ctx, compute_ctx, src0, src1, node);
+
+        break;
+    case GGML_OP_UPSCALE:
+        ggml_vk_upscale(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_SCALE:
-        ggml_vk_scale(ctx, ctx->compute_ctx, src0, node);
+        ggml_vk_scale(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_SQR:
-        ggml_vk_sqr(ctx, ctx->compute_ctx, src0, node);
+        ggml_vk_sqr(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_CLAMP:
-        ggml_vk_clamp(ctx, ctx->compute_ctx, src0, node);
+        ggml_vk_clamp(ctx, compute_ctx, src0, node);
+
+        break;
+    case GGML_OP_PAD:
+        ggml_vk_pad(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_CPY:
     case GGML_OP_CONT:
     case GGML_OP_DUP:
-        ggml_vk_cpy(ctx, ctx->compute_ctx, src0, node);
+        ggml_vk_cpy(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_NORM:
-        ggml_vk_norm(ctx, ctx->compute_ctx, src0, node);
+        ggml_vk_norm(ctx, compute_ctx, src0, node);
+
+        break;
+    case GGML_OP_GROUP_NORM:
+        ggml_vk_group_norm(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_RMS_NORM:
-        ggml_vk_rms_norm(ctx, ctx->compute_ctx, src0, node);
+        ggml_vk_rms_norm(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_UNARY:
         switch (ggml_get_unary_op(node)) {
         case GGML_UNARY_OP_SILU:
         case GGML_UNARY_OP_GELU:
+        case GGML_UNARY_OP_GELU_QUICK:
         case GGML_UNARY_OP_RELU:
-            ggml_vk_unary(ctx, ctx->compute_ctx, src0, node);
+        case GGML_UNARY_OP_TANH:
+            ggml_vk_unary(ctx, compute_ctx, src0, node);
             break;
         default:
             return;
         }
         break;
     case GGML_OP_DIAG_MASK_INF:
-        ggml_vk_diag_mask_inf(ctx, ctx->compute_ctx, src0, node);
+        ggml_vk_diag_mask_inf(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_SOFT_MAX:
-        ggml_vk_soft_max(ctx, ctx->compute_ctx, src0, src1, node);
+        ggml_vk_soft_max(ctx, compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_ROPE:
-        ggml_vk_rope(ctx, ctx->compute_ctx, src0, src1, src2, node);
+        ggml_vk_rope(ctx, compute_ctx, src0, src1, src2, node);
 
         break;
     case GGML_OP_ARGSORT:
-        ggml_vk_argsort(ctx, ctx->compute_ctx, src0, node);
+        ggml_vk_argsort(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_SUM_ROWS:
-        ggml_vk_sum_rows(ctx, ctx->compute_ctx, src0, node);
+        ggml_vk_sum_rows(ctx, compute_ctx, src0, node);
+
+        break;
+    case GGML_OP_IM2COL:
+        ggml_vk_im2col(ctx, compute_ctx, src0, src1, node);
+
+        break;
+    case GGML_OP_TIMESTEP_EMBEDDING:
+        ggml_vk_timestep_embedding(ctx, compute_ctx, src0, node);
+
+        break;
+    case GGML_OP_LEAKY_RELU:
+        ggml_vk_leaky_relu(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_MUL_MAT:
-        ggml_vk_mul_mat(ctx, ctx->compute_ctx, src0, src1, node);
+        ggml_vk_mul_mat(ctx, compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_MUL_MAT_ID:
-        ggml_vk_mul_mat_id(ctx, ctx->compute_ctx, src0, src1, src2, node);
+        ggml_vk_mul_mat_id(ctx, compute_ctx, src0, src1, src2, node);
 
         break;
     default:
         return;
     }
 
-    extra->ctx_idx = ctx->compute_ctx->idx;
+    ctx->tensor_ctxs[node_idx] = compute_ctx;
 
 #ifdef GGML_VULKAN_CHECK_RESULTS
     // Force context reset on each node so that each tensor ends up in its own context
@@ -5594,13 +5928,13 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
 #endif
 
     if (last_node) {
-        ggml_vk_ctx_end(ctx->compute_ctx);
-        ctx->compute_ctx->exit_tensor = node;
-        ctx->compute_ctx = nullptr;
+        ggml_vk_ctx_end(compute_ctx);
+        compute_ctx->exit_tensor_idx = node_idx;
+        ctx->compute_ctx.reset();
     }
 }
 
-static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * tensor){
+static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * tensor, int tensor_idx){
     ggml_tensor_extra_gpu * extra = nullptr;
 
     switch (tensor->op) {
@@ -5608,13 +5942,17 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
     case GGML_OP_GET_ROWS:
     case GGML_OP_MUL:
     case GGML_OP_DIV:
+    case GGML_OP_CONCAT:
+    case GGML_OP_UPSCALE:
     case GGML_OP_SCALE:
     case GGML_OP_SQR:
     case GGML_OP_CLAMP:
+    case GGML_OP_PAD:
     case GGML_OP_CPY:
     case GGML_OP_CONT:
     case GGML_OP_DUP:
     case GGML_OP_NORM:
+    case GGML_OP_GROUP_NORM:
     case GGML_OP_RMS_NORM:
     case GGML_OP_DIAG_MASK_INF:
     case GGML_OP_SOFT_MAX:
@@ -5626,6 +5964,10 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
     case GGML_OP_NONE:
     case GGML_OP_ARGSORT:
     case GGML_OP_SUM_ROWS:
+    case GGML_OP_IM2COL:
+    case GGML_OP_TIMESTEP_EMBEDDING:
+    case GGML_OP_LEAKY_RELU:
+    case GGML_OP_REPEAT:
         extra = (ggml_tensor_extra_gpu *) tensor->extra;
 
         break;
@@ -5633,7 +5975,9 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
         switch (ggml_get_unary_op(tensor)) {
         case GGML_UNARY_OP_SILU:
         case GGML_UNARY_OP_GELU:
+        case GGML_UNARY_OP_GELU_QUICK:
         case GGML_UNARY_OP_RELU:
+        case GGML_UNARY_OP_TANH:
             extra = (ggml_tensor_extra_gpu *) tensor->extra;
             break;
         default:
@@ -5656,31 +6000,31 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
     VK_LOG_DEBUG("ggml_vk_compute_forward(" << tensor << ", name=" << tensor->name << ", op=" << ggml_op_name(tensor->op) << ", type=" << tensor->type << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << ", view_src=" << tensor->view_src << ", view_offs=" << tensor->view_offs << ")");
 
 #ifdef GGML_VULKAN_CHECK_RESULTS
-    ggml_vk_check_results_0(ctx, tensor);
+    ggml_vk_check_results_0(tensor);
 #endif
 
-    vk_context& subctx = ctx->gc.contexts[extra->ctx_idx];
+    vk_context subctx = ctx->tensor_ctxs[tensor_idx].lock();
 
     // Only run if ctx hasn't been submitted yet
-    if (!subctx.seqs.empty()) {
+    if (!subctx->seqs.empty()) {
         // Do staging buffer copies
-        for (auto& cpy : subctx.in_memcpys) {
+        for (auto& cpy : subctx->in_memcpys) {
             memcpy(cpy.dst, cpy.src, cpy.n);
         }
 
-        ggml_vk_submit(&subctx, ctx->fence);
+        ggml_vk_submit(subctx, ctx->fence);
     }
 
-    if (tensor == subctx.exit_tensor) {
+    if (tensor_idx == subctx->exit_tensor_idx) {
         VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_compute_forward waitForFences");
         ctx->device->device.resetFences({ ctx->fence });
 
         // Do staging buffer copies
-        for (auto& cpy : subctx.out_memcpys) {
+        for (auto& cpy : subctx->out_memcpys) {
             memcpy(cpy.dst, cpy.src, cpy.n);
         }
-        subctx.in_memcpys.clear();
-        subctx.out_memcpys.clear();
+        subctx->in_memcpys.clear();
+        subctx->out_memcpys.clear();
     }
 
     return true;
@@ -5725,8 +6069,7 @@ static void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx) {
 
     ctx->staging_offset = 0;
 
-    ctx->compute_ctx = nullptr;
-    ctx->transfer_ctx = nullptr;
+    ctx->tensor_ctxs.clear();
     ctx->gc.contexts.clear();
 }
 
@@ -6063,15 +6406,20 @@ GGML_CALL static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, g
 
     ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
 
-    if (ctx->transfer_ctx == nullptr) {
+    vk_context transfer_ctx;
+
+    if (ctx->transfer_ctx.expired()) {
         // Initialize new transfer context
-        ctx->transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue);
-        ggml_vk_ctx_begin(ctx->device, ctx->transfer_ctx);
+        transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue);
+        ctx->transfer_ctx = transfer_ctx;
+        ggml_vk_ctx_begin(ctx->device, transfer_ctx);
+    } else {
+        transfer_ctx = ctx->transfer_ctx.lock();
     }
 
     vk_buffer buf = extra->buffer_gpu.lock();
 
-    ggml_vk_buffer_write_async(ctx->transfer_ctx, buf, extra->offset + tensor->view_offs + offset, data, size, ctx->staging, ctx->staging_offset);
+    ggml_vk_buffer_write_async(transfer_ctx, buf, extra->offset + tensor->view_offs + offset, data, size, ctx->staging, ctx->staging_offset);
 }
 
 GGML_CALL static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
@@ -6081,15 +6429,20 @@ GGML_CALL static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, c
 
     ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
 
-    if (ctx->transfer_ctx == nullptr) {
+    vk_context transfer_ctx;
+
+    if (ctx->transfer_ctx.expired()) {
         // Initialize new transfer context
-        ctx->transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue);
-        ggml_vk_ctx_begin(ctx->device, ctx->transfer_ctx);
+        transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue);
+        ctx->transfer_ctx = transfer_ctx;
+        ggml_vk_ctx_begin(ctx->device, transfer_ctx);
+    } else {
+        transfer_ctx = ctx->transfer_ctx.lock();
     }
 
     vk_buffer buf = extra->buffer_gpu.lock();
 
-    ggml_vk_buffer_read_async(ctx->transfer_ctx, buf, extra->offset + tensor->view_offs + offset, data, size, ctx->staging, ctx->staging_offset);
+    ggml_vk_buffer_read_async(transfer_ctx, buf, extra->offset + tensor->view_offs + offset, data, size, ctx->staging, ctx->staging_offset);
 }
 
 GGML_CALL static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, const ggml_tensor * src, ggml_tensor * dst) {
@@ -6099,16 +6452,21 @@ GGML_CALL static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, c
         ggml_tensor_extra_gpu * src_extra = (ggml_tensor_extra_gpu *) src->extra;
         ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
 
-        if (ctx->transfer_ctx == nullptr) {
+        vk_context transfer_ctx;
+
+        if (ctx->transfer_ctx.expired()) {
             // Initialize new transfer context
-            ctx->transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue);
-            ggml_vk_ctx_begin(ctx->device, ctx->transfer_ctx);
+            transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue);
+            ctx->transfer_ctx = transfer_ctx;
+            ggml_vk_ctx_begin(ctx->device, transfer_ctx);
+        } else {
+            transfer_ctx = ctx->transfer_ctx.lock();
         }
 
         vk_buffer src_buf = src_extra->buffer_gpu.lock();
         vk_buffer dst_buf = dst_extra->buffer_gpu.lock();
 
-        ggml_vk_buffer_copy_async(ctx->transfer_ctx, dst_buf, dst_extra->offset + dst->view_offs, src_buf, src_extra->offset + src->view_offs, ggml_nbytes(src));
+        ggml_vk_buffer_copy_async(transfer_ctx, dst_buf, dst_extra->offset + dst->view_offs, src_buf, src_extra->offset + src->view_offs, ggml_nbytes(src));
         return true;
     }
 
@@ -6118,25 +6476,27 @@ GGML_CALL static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, c
 GGML_CALL static void ggml_backend_vk_synchronize(ggml_backend_t backend) {
     VK_LOG_DEBUG("ggml_backend_vk_synchronize()");
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
-    if(ctx->transfer_ctx == nullptr) {
+    if(ctx->transfer_ctx.expired()) {
         return;
     }
 
-    ggml_vk_ctx_end(ctx->transfer_ctx);
+    vk_context transfer_ctx = ctx->transfer_ctx.lock();
+
+    ggml_vk_ctx_end(transfer_ctx);
 
-    for (auto& cpy : ctx->transfer_ctx->in_memcpys) {
+    for (auto& cpy : transfer_ctx->in_memcpys) {
         memcpy(cpy.dst, cpy.src, cpy.n);
     }
 
-    ggml_vk_submit(ctx->transfer_ctx, ctx->fence);
+    ggml_vk_submit(transfer_ctx, ctx->fence);
     VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_backend_vk_synchronize waitForFences");
     ctx->device->device.resetFences({ ctx->fence });
 
-    for (auto& cpy : ctx->transfer_ctx->out_memcpys) {
+    for (auto& cpy : transfer_ctx->out_memcpys) {
         memcpy(cpy.dst, cpy.src, cpy.n);
     }
 
-    ctx->transfer_ctx = nullptr;
+    ctx->transfer_ctx.reset();
 }
 
 static bool ggml_vk_is_empty(ggml_tensor * node) {
@@ -6159,8 +6519,11 @@ GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backen
         last_node -= 1;
     }
 
+    // Reserve tensor context space for all nodes
+    ctx->tensor_ctxs.resize(cgraph->n_nodes);
+
     for (int i = 0; i < cgraph->n_nodes; i++) {
-        ggml_vk_build_graph(ctx,cgraph->nodes[i], i == last_node);
+        ggml_vk_build_graph(ctx, cgraph->nodes[i], i, i == last_node);
     }
 
     for (int i = 0; i < cgraph->n_nodes; i++) {
@@ -6170,13 +6533,17 @@ GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backen
             continue;
         }
 
-        bool ok = ggml_vk_compute_forward(ctx, node);
+        bool ok = ggml_vk_compute_forward(ctx, node, i);
         if (!ok) {
-            fprintf(stderr, "%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
+            if (node->op == GGML_OP_UNARY) {
+                std::cerr << __func__ << ": error: op not supported UNARY " << node->name << " (" << ggml_unary_op_name(static_cast<ggml_unary_op>(node->op_params[0])) << ")" << std::endl;
+            } else {
+                std::cerr << __func__ << ": error: op not supported " << node->name << " (" << ggml_op_name(node->op) << ")" << std::endl;
+            }
         }
 #ifdef GGML_VULKAN_CHECK_RESULTS
         else {
-            ggml_vk_check_results_1(ctx, node);
+            ggml_vk_check_results_1(node);
         }
 #endif
         GGML_ASSERT(ok);
@@ -6196,8 +6563,10 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
                 case GGML_UNARY_OP_GELU:
+                case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_TANH:
                     return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;
@@ -6270,11 +6639,11 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const
                 }
                 return false;
             } break;
-        // case GGML_OP_REPEAT:
-        //     {
-        //         ggml_type src0_type = op->src[0]->type;
-        //         return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16;
-        //     } break;
+        case GGML_OP_REPEAT:
+            {
+                ggml_type src0_type = op->src[0]->type;
+                return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16;
+            } break;
         case GGML_OP_ROPE:
             return ggml_is_contiguous(op->src[0]);
         case GGML_OP_NONE:
@@ -6283,18 +6652,25 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const
         case GGML_OP_PERMUTE:
         case GGML_OP_TRANSPOSE:
         case GGML_OP_NORM:
+        case GGML_OP_GROUP_NORM:
+        case GGML_OP_RMS_NORM:
         case GGML_OP_ADD:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
-        case GGML_OP_RMS_NORM:
+        case GGML_OP_CONCAT:
+        case GGML_OP_UPSCALE:
         case GGML_OP_SCALE:
         case GGML_OP_SQR:
         case GGML_OP_CLAMP:
+        case GGML_OP_PAD:
         case GGML_OP_CONT:
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_ARGSORT:
         case GGML_OP_SUM_ROWS:
+        case GGML_OP_IM2COL:
+        case GGML_OP_TIMESTEP_EMBEDDING:
+        case GGML_OP_LEAKY_RELU:
             return true;
         default:
             return false;
@@ -6498,7 +6874,7 @@ static void ggml_vk_print_tensor_area(const ggml_tensor * tensor, const void * d
                 } else if (tensor->type == GGML_TYPE_I32) {
                     val = *(const int32_t *) ((const char *) data + i3*tensor->nb[3] + i2*tensor->nb[2] + idx1*tensor->nb[1] + idx0*tensor->nb[0]);
                 } else {
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
                 }
                 fprintf(stderr, "% 7.2f ", val);
             } else {
@@ -6509,10 +6885,12 @@ static void ggml_vk_print_tensor_area(const ggml_tensor * tensor, const void * d
     }
 }
 
-static void ggml_vk_print_tensor(ggml_backend_vk_context * ctx, const ggml_tensor * tensor, const char * name) {
+static void ggml_vk_print_tensor(const ggml_tensor * tensor, const char * name) {
     void * tensor_data = tensor->data;
 
-    if (ggml_backend_buffer_is_vk(tensor->buffer)) {
+    const bool is_gpu = tensor->buffer != nullptr && ggml_backend_buffer_is_vk(tensor->buffer);
+
+    if (is_gpu) {
         const size_t tensor_size = ggml_nbytes(tensor);
         tensor_data = malloc(tensor_size);
 
@@ -6533,13 +6911,10 @@ static void ggml_vk_print_tensor(ggml_backend_vk_context * ctx, const ggml_tenso
     std::cerr << std::endl << "Result:" << std::endl;
     ggml_vk_print_tensor_area(tensor, tensor_data, 5, 5, 0, 0);
     std::cerr << std::endl;
-    std::cerr << std::endl << "Result:" << std::endl;
-    ggml_vk_print_tensor_area(tensor, tensor_data, 5, 5, 1, 0);
-    std::cerr << std::endl;
     std::vector<const ggml_tensor *> done;
     ggml_vk_print_graph_origin(tensor, done);
 
-    if (ggml_backend_buffer_is_vk(tensor->buffer)) {
+    if (is_gpu) {
         free(tensor_data);
     }
 }
@@ -6548,8 +6923,8 @@ void * comp_result;
 size_t comp_size;
 size_t comp_nb[GGML_MAX_DIMS];
 size_t check_counter = 0;
-static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor * tensor) {
-        if (tensor->op == GGML_OP_TRANSPOSE) {
+static void ggml_vk_check_results_0(ggml_tensor * tensor) {
+    if (tensor->op == GGML_OP_TRANSPOSE) {
         return;
     }
 
@@ -6565,7 +6940,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor *
     ggml_tensor * src2 = tensor->src[2];
 
     struct ggml_init_params iparams = {
-        /*.mem_size   =*/ 1024*1024*1024,
+        /*.mem_size   =*/ 2ul*1024ul*1024ul*1024ul,
         /*.mem_buffer =*/ NULL,
         /*.no_alloc   =*/ false,
     };
@@ -6620,11 +6995,11 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor *
                 memcpy(src0_clone->nb, src0->nb, sizeof(size_t) * GGML_MAX_DIMS);
             }
         } else {
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
 
         if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
-            ggml_vk_print_tensor(ctx, src0, "src0");
+            ggml_vk_print_tensor(src0, "src0");
         }
     }
     if (src1 != nullptr) {
@@ -6662,27 +7037,11 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor *
                 memcpy(src1_clone->nb, src1->nb, sizeof(size_t) * GGML_MAX_DIMS);
             }
         } else {
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
 
         if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
-            ggml_vk_print_tensor(ctx, src1, "src1");
-            std::cerr << "TENSOR CHECK: " << ggml_op_name(src1_clone->op) << " (check " << check_counter << ")" << std::endl;
-            std::cerr << "src1_clone=" << tensor << " src1_clone->type: " << ggml_type_name(src1_clone->type) << " ne0=" << src1_clone->ne[0] << " nb0=" << src1_clone->nb[0] << " ne1=" << src1_clone->ne[1] << " nb1=" << src1_clone->nb[1] << " ne2=" << src1_clone->ne[2] << " nb2=" << src1_clone->nb[2] << " ne3=" << src1_clone->ne[3] << " nb3=" << src1_clone->nb[3] << std::endl;
-            if (src1->src[0] != nullptr) {
-                std::cerr << "src1->src[0]=" << src1->src[0] << " op=" << ggml_op_name(src1->src[0]->op) << " type=" << ggml_type_name(src1->src[0]->type) << " ne0=" << src1->src[0]->ne[0] << " nb0=" << src1->src[0]->nb[0] << " ne1=" << src1->src[0]->ne[1] << " nb1=" << src1->src[0]->nb[1] << " ne2=" << src1->src[0]->ne[2] << " nb2=" << src1->src[0]->nb[2] << " ne3=" << src1->src[0]->ne[3] << " nb3=" << src1->src[0]->nb[3] << std::endl;
-            }
-            if (src1->src[1] != nullptr) {
-                std::cerr << "src1->src[1]=" << src1->src[1] << " op=" << ggml_op_name(src1->src[1]->op) << " type=" << ggml_type_name(src1->src[1]->type) << " ne0=" << src1->src[1]->ne[0] << " nb0=" << src1->src[1]->nb[0] << " ne1=" << src1->src[1]->ne[1] << " nb1=" << src1->src[1]->nb[1] << " ne2=" << src1->src[1]->ne[2] << " nb2=" << src1->src[1]->nb[2] << " ne3=" << src1->src[1]->ne[3] << " nb3=" << src1->src[1]->nb[3] << std::endl;
-            }
-            std::cerr << std::endl << "Result:" << std::endl;
-            ggml_vk_print_tensor_area(src1_clone, src1_clone->data, 5, 5, 0, 0);
-            std::cerr << std::endl;
-            std::cerr << std::endl << "Result:" << std::endl;
-            ggml_vk_print_tensor_area(src1_clone, src1_clone->data, 5, 5, 1, 0);
-            std::cerr << std::endl;
-            std::vector<const ggml_tensor *> done;
-            ggml_vk_print_graph_origin(src1_clone, done);
+            ggml_vk_print_tensor(src1, "src1");
         }
     }
     if (src2 != nullptr) {
@@ -6720,27 +7079,11 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor *
                 memcpy(src2_clone->nb, src2->nb, sizeof(size_t) * GGML_MAX_DIMS);
             }
         } else {
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
 
         if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
-            ggml_vk_print_tensor(ctx, src2, "src2");
-            std::cerr << "TENSOR CHECK: " << ggml_op_name(src2_clone->op) << " (check " << check_counter << ")" << std::endl;
-            std::cerr << "src2_clone=" << tensor << " src2_clone->type: " << ggml_type_name(src2_clone->type) << " ne0=" << src2_clone->ne[0] << " nb0=" << src2_clone->nb[0] << " ne1=" << src2_clone->ne[1] << " nb1=" << src2_clone->nb[1] << " ne2=" << src2_clone->ne[2] << " nb2=" << src2_clone->nb[2] << " ne3=" << src2_clone->ne[3] << " nb3=" << src2_clone->nb[3] << std::endl;
-            if (src2->src[0] != nullptr) {
-                std::cerr << "src2->src[0]=" << src2->src[0] << " op=" << ggml_op_name(src2->src[0]->op) << " type=" << ggml_type_name(src2->src[0]->type) << " ne0=" << src2->src[0]->ne[0] << " nb0=" << src2->src[0]->nb[0] << " ne1=" << src2->src[0]->ne[1] << " nb1=" << src2->src[0]->nb[1] << " ne2=" << src2->src[0]->ne[2] << " nb2=" << src2->src[0]->nb[2] << " ne3=" << src2->src[0]->ne[3] << " nb3=" << src2->src[0]->nb[3] << std::endl;
-            }
-            if (src2->src[1] != nullptr) {
-                std::cerr << "src2->src[1]=" << src2->src[1] << " op=" << ggml_op_name(src2->src[1]->op) << " type=" << ggml_type_name(src2->src[1]->type) << " ne0=" << src2->src[1]->ne[0] << " nb0=" << src2->src[1]->nb[0] << " ne1=" << src2->src[1]->ne[1] << " nb1=" << src2->src[1]->nb[1] << " ne2=" << src2->src[1]->ne[2] << " nb2=" << src2->src[1]->nb[2] << " ne3=" << src2->src[1]->ne[3] << " nb3=" << src2->src[1]->nb[3] << std::endl;
-            }
-            std::cerr << std::endl << "Result:" << std::endl;
-            ggml_vk_print_tensor_area(src2_clone, src2_clone->data, 5, 5, 0, 0);
-            std::cerr << std::endl;
-            std::cerr << std::endl << "Result:" << std::endl;
-            ggml_vk_print_tensor_area(src2_clone, src2_clone->data, 5, 5, 1, 0);
-            std::cerr << std::endl;
-            std::vector<const ggml_tensor *> done;
-            ggml_vk_print_graph_origin(src2_clone, done);
+            ggml_vk_print_tensor(src2, "src2");
         }
     }
 
@@ -6752,16 +7095,24 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor *
         tensor_clone = ggml_mul(ggml_ctx, src0_clone, src1_clone);
     } else if (tensor->op == GGML_OP_DIV) {
         tensor_clone = ggml_div(ggml_ctx, src0_clone, src1_clone);
+    } else if (tensor->op == GGML_OP_CONCAT) {
+        tensor_clone = ggml_concat(ggml_ctx, src0_clone, src1_clone, *(int *)tensor->op_params);
+    } else if (tensor->op == GGML_OP_UPSCALE) {
+        tensor_clone = ggml_upscale_ext(ggml_ctx, src0_clone, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
     } else if (tensor->op == GGML_OP_SCALE) {
         tensor_clone = ggml_scale(ggml_ctx, src0_clone, ((float *)tensor->op_params)[0]);
     } else if (tensor->op == GGML_OP_SQR) {
         tensor_clone = ggml_sqr(ggml_ctx, src0_clone);
     } else if (tensor->op == GGML_OP_CLAMP) {
         tensor_clone = ggml_clamp(ggml_ctx, src0_clone, ((float *)tensor->op_params)[0], ((float *)tensor->op_params)[1]);
+    } else if (tensor->op == GGML_OP_PAD) {
+        tensor_clone = ggml_pad(ggml_ctx, src0_clone, tensor->ne[0] - src0_clone->ne[0], tensor->ne[1] - src0_clone->ne[1], tensor->ne[2] - src0_clone->ne[2], tensor->ne[3] - src0_clone->ne[3]);
     } else if (tensor->op == GGML_OP_ADD) {
         tensor_clone = ggml_add(ggml_ctx, src0_clone, src1_clone);
     } else if (tensor->op == GGML_OP_NORM) {
         tensor_clone = ggml_norm(ggml_ctx, src0_clone, *(float *)tensor->op_params);
+    } else if (tensor->op == GGML_OP_GROUP_NORM) {
+        tensor_clone = ggml_group_norm(ggml_ctx, src0_clone, *(int *)tensor->op_params);
     } else if (tensor->op == GGML_OP_RMS_NORM) {
         tensor_clone = ggml_rms_norm(ggml_ctx, src0_clone, *(float *)tensor->op_params);
     } else if (tensor->op == GGML_OP_SOFT_MAX) {
@@ -6777,12 +7128,12 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor *
         const int mode        = ((int32_t *) tensor->op_params)[2];
         //const int n_ctx_ggml       = ((int32_t *) tensor->op_params)[3];
         const int n_ctx_orig_ggml  = ((int32_t *) tensor->op_params)[4];
-        float freq_base       = ((float *)   tensor->op_params)[5];
-        float freq_scale      = ((float *)   tensor->op_params)[6];
-        float ext_factor      = ((float *)   tensor->op_params)[7];
-        float attn_factor     = ((float *)   tensor->op_params)[8];
-        float beta_fast       = ((float *)   tensor->op_params)[9];
-        float beta_slow       = ((float *)   tensor->op_params)[10];
+        const float freq_base       = ((float *) tensor->op_params)[5];
+        const float freq_scale      = ((float *) tensor->op_params)[6];
+        const float ext_factor      = ((float *) tensor->op_params)[7];
+        const float attn_factor     = ((float *) tensor->op_params)[8];
+        const float beta_fast       = ((float *) tensor->op_params)[9];
+        const float beta_slow       = ((float *) tensor->op_params)[10];
         tensor_clone = ggml_rope_ext(ggml_ctx, src0_clone, src1_clone, src2_clone, n_dims, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
     } else if (tensor->op == GGML_OP_UNARY) {
         switch (ggml_get_unary_op(tensor)) {
@@ -6792,12 +7143,18 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor *
         case GGML_UNARY_OP_GELU:
             tensor_clone = ggml_gelu(ggml_ctx, src0_clone);
             break;
+        case GGML_UNARY_OP_GELU_QUICK:
+            tensor_clone = ggml_gelu_quick(ggml_ctx, src0_clone);
+            break;
         case GGML_UNARY_OP_RELU:
             tensor_clone = ggml_relu(ggml_ctx, src0_clone);
             break;
+        case GGML_UNARY_OP_TANH:
+            tensor_clone = ggml_tanh(ggml_ctx, src0_clone);
+            break;
         default:
             std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
     } else if (tensor->op == GGML_OP_CPY || tensor->op == GGML_OP_DUP) {
         if (src1 == nullptr) {
@@ -6823,9 +7180,26 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor *
         tensor_clone = ggml_argsort(ggml_ctx, src0_clone, (ggml_sort_order) *(int *)tensor->op_params);
     } else if (tensor->op == GGML_OP_SUM_ROWS) {
         tensor_clone = ggml_sum_rows(ggml_ctx, src0_clone);
+    } else if (tensor->op == GGML_OP_IM2COL) {
+        const int32_t s0 = tensor->op_params[0];
+        const int32_t s1 = tensor->op_params[1];
+        const int32_t p0 = tensor->op_params[2];
+        const int32_t p1 = tensor->op_params[3];
+        const int32_t d0 = tensor->op_params[4];
+        const int32_t d1 = tensor->op_params[5];
+
+        const bool is_2D = tensor->op_params[6] == 1;
+        tensor_clone = ggml_im2col(ggml_ctx, src0_clone, src1_clone, s0, s1, p0, p1, d0, d1, is_2D, tensor->type);
+    } else if (tensor->op == GGML_OP_TIMESTEP_EMBEDDING) {
+        const int32_t dim = tensor->op_params[0];
+        const int32_t max_period = tensor->op_params[1];
+        tensor_clone = ggml_timestep_embedding(ggml_ctx, src0_clone, dim, max_period);
+    } else if (tensor->op == GGML_OP_LEAKY_RELU) {
+        const float * op_params = (const float *)tensor->op_params;
+        tensor_clone = ggml_leaky_relu(ggml_ctx, src0_clone, op_params[0], false);
     } else {
         std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 
     ggml_cgraph * cgraph = ggml_new_graph(ggml_ctx);
@@ -6834,7 +7208,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor *
     ggml_graph_compute_with_ctx(ggml_ctx, cgraph, 8);
 
     if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
-        ggml_vk_print_tensor(ctx, tensor_clone, "tensor_clone");
+        ggml_vk_print_tensor(tensor_clone, "tensor_clone");
     }
 
     comp_size = ggml_nbytes(tensor_clone);
@@ -6851,9 +7225,11 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor *
     }
 
     ggml_free(ggml_ctx);
+
+    VK_LOG_DEBUG("END ggml_vk_check_results_0(" << tensor->name << ")");
 }
 
-static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_tensor * tensor) {
+static void ggml_vk_check_results_1(ggml_tensor * tensor) {
     if (tensor->op == GGML_OP_TRANSPOSE) {
         return;
     }
@@ -6912,7 +7288,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_tensor *
                         }
                     } else {
                         std::cerr << "Missing debug code for type " << ggml_type_name(tensor->type) << std::endl;
-                        GGML_ASSERT(false);
+                        GGML_ABORT("fatal error");
                     }
 
                     if ((std::isnan(correct) != std::isnan(result)) || (std::isinf(correct) != std::isinf(result)) || !buffer_size_fit) {
@@ -6935,7 +7311,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_tensor *
                         std::cerr << std::endl;
                         std::vector<const ggml_tensor *> done;
                         ggml_vk_print_graph_origin(tensor, done);
-                        GGML_ASSERT(false);
+                        GGML_ABORT("fatal error");
                     }
                     if (first_error[0] == -1 && std::fabs(correct - result) > 0.1f) {
                         first_error[0] = i0;
@@ -6977,11 +7353,6 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_tensor *
         std::cerr << std::endl << "Correct:" << std::endl;
         ggml_vk_print_tensor_area(tensor, comp_result, 5, 5, 0, 0);
         std::cerr << std::endl;
-        std::cerr << std::endl << "Result:" << std::endl;
-        ggml_vk_print_tensor_area(tensor, tensor_data, 5, 5, 1, 0);
-        std::cerr << std::endl << "Correct:" << std::endl;
-        ggml_vk_print_tensor_area(tensor, comp_result, 5, 5, 1, 0);
-        std::cerr << std::endl;
         std::vector<const ggml_tensor *> done;
         ggml_vk_print_graph_origin(tensor, done);
     }
@@ -7006,7 +7377,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_tensor *
         std::cerr << std::endl;
         std::vector<const ggml_tensor *> done;
         ggml_vk_print_graph_origin(tensor, done);
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     } else {
         std::cerr << check_counter << " " << tensor->name << " op=" << ggml_op_name(tensor->op) << " avg_err=" << avg_err << std::endl;
     }
@@ -7018,5 +7389,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_tensor *
     if (ggml_backend_buffer_is_vk(tensor->buffer)) {
         free(tensor_data);
     }
+
+    VK_LOG_DEBUG("END ggml_vk_check_results_1(" << tensor->name << ")");
 }
 #endif
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index b5fdb96d..73054bfe 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -47,6 +47,9 @@
 #include <unistd.h>
 #endif
 
+#if defined(__ARM_FEATURE_SVE)
+int ggml_sve_cnt_b = 0;
+#endif
 #if defined(__ARM_FEATURE_SVE) || defined(__ARM_FEATURE_MATMUL_INT8)
 #undef GGML_USE_LLAMAFILE
 #endif
@@ -63,6 +66,9 @@
 // disable POSIX deprecation warnings
 // these functions are never going away, anyway
 #pragma warning(disable: 4996)
+
+// unreachable code because of multiple instances of code after GGML_ABORT
+#pragma warning(disable: 4702)
 #endif
 
 #if defined(_WIN32)
@@ -151,23 +157,69 @@ typedef pthread_t ggml_thread_t;
 
 #include <sys/wait.h>
 
-void ggml_print_backtrace(void) {
-    /*
-    #include <execinfo.h>
-    #include <dlfcn.h>
+#if defined(__ANDROID__)
+#include <unwind.h>
+#include <dlfcn.h>
+#include <stdio.h>
 
-    void * trace[100];
+struct backtrace_state {
+    void ** current;
+    void ** end;
+};
 
-    int nptrs = backtrace(trace, sizeof(trace)/sizeof(trace[0]));
+static _Unwind_Reason_Code unwind_callback(struct _Unwind_Context* context, void* arg) {
+    struct backtrace_state * state = (struct backtrace_state *)arg;
+    uintptr_t pc = _Unwind_GetIP(context);
+    if (pc) {
+        if (state->current == state->end) {
+            return _URC_END_OF_STACK;
+        } else {
+            *state->current++ = (void*)pc;
+        }
+    }
+    return _URC_NO_REASON;
+}
+
+static void ggml_print_backtrace_symbols(void) {
+    const int max = 100;
+    void* buffer[max];
 
+    struct backtrace_state state = {buffer, buffer + max};
+    _Unwind_Backtrace(unwind_callback, &state);
+
+    int count = state.current - buffer;
+
+    for (int idx = 0; idx < count; ++idx) {
+        const void * addr = buffer[idx];
+        const char * symbol = "";
+
+        Dl_info info;
+        if (dladdr(addr, &info) && info.dli_sname) {
+            symbol = info.dli_sname;
+        }
+
+        fprintf(stderr, "%d: %p %s\n", idx, addr, symbol);
+    }
+}
+#elif defined(__linux__) && defined(__GLIBC__)
+#include <execinfo.h>
+static void ggml_print_backtrace_symbols(void) {
+    void * trace[100];
+    int nptrs = backtrace(trace, sizeof(trace)/sizeof(trace[0]));
     backtrace_symbols_fd(trace, nptrs, STDERR_FILENO);
-    */
+}
+#else
+static void ggml_print_backtrace_symbols(void) {
+    // platform not supported
+}
+#endif
 
-    // backtrack_symbols does not show line numbers, use gdb instead
+static void ggml_print_backtrace(void) {
     char attach[32];
     snprintf(attach, sizeof(attach), "attach %d", getpid());
     int pid = fork();
     if (pid == 0) {
+        // try gdb
         execlp("gdb", "gdb", "--batch",
             "-ex", "set style enabled on",
             "-ex", attach,
@@ -175,16 +227,46 @@ void ggml_print_backtrace(void) {
             "-ex", "detach",
             "-ex", "quit",
             (char *) NULL);
+        // try lldb
+        execlp("lldb", "lldb", "--batch",
+            "-o", "bt",
+            "-o", "quit",
+            "-p", attach,
+            (char *) NULL);
+        exit(EXIT_FAILURE);
     } else {
-        waitpid(pid, NULL, 0);
+        int wstatus;
+        waitpid(pid, &wstatus, 0);
+        if (WIFEXITED(wstatus)) {
+            if (WEXITSTATUS(wstatus) == EXIT_FAILURE) {
+                // gdb failed, fallback to backtrace_symbols
+                ggml_print_backtrace_symbols();
+            }
+        }
     }
 }
 #else
-void ggml_print_backtrace(void) {
+static void ggml_print_backtrace(void) {
     // platform not supported
 }
 #endif
 
+void ggml_abort(const char * file, int line, const char * fmt, ...) {
+    fflush(stdout);
+
+    fprintf(stderr, "%s:%d: ", file, line);
+
+    va_list args;
+    va_start(args, fmt);
+    vfprintf(stderr, fmt, args);
+    va_end(args);
+
+    fprintf(stderr, "\n");
+
+    ggml_print_backtrace();
+    abort();
+}
+
 #define GGML_DEBUG 0
 #define GGML_GELU_FP16
 #define GGML_GELU_QUICK_FP16
@@ -256,7 +338,7 @@ inline static void * ggml_aligned_malloc(size_t size) {
                 break;
         }
         GGML_PRINT("%s: %s (attempted to allocate %6.2f MB)\n", __func__, error_desc, size/(1024.0*1024.0));
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
         return NULL;
     }
     return aligned_memory;
@@ -277,7 +359,7 @@ inline static void * ggml_malloc(size_t size) {
     void * result = malloc(size);
     if (result == NULL) {
         GGML_PRINT("%s: failed to allocate %6.2f MB\n", __func__, size/(1024.0*1024.0));
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
     return result;
 }
@@ -291,7 +373,7 @@ inline static void * ggml_calloc(size_t num, size_t size) {
     void * result = calloc(num, size);
     if (result == NULL) {
         GGML_PRINT("%s: failed to allocate %6.2f MB\n", __func__, size/(1024.0*1024.0));
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
     return result;
 }
@@ -414,9 +496,16 @@ void ggml_bf16_to_fp32_row(const ggml_bf16_t * x, float * y, int64_t n) {
     }
 }
 
+void ggml_fp32_to_bf16_row_ref(const float * x, ggml_bf16_t * y, int64_t n) {
+    for (int i = 0; i < n; i++) {
+        y[i] = ggml_compute_fp32_to_bf16(x[i]);
+    }
+}
+
 void ggml_fp32_to_bf16_row(const float * x, ggml_bf16_t * y, int64_t n) {
   int i = 0;
 #if defined(__AVX512BF16__)
+  // subnormals are flushed to zero on this platform
   for (; i + 32 <= n; i += 32) {
         _mm512_storeu_si512(
             (__m512i *)(y + i),
@@ -939,7 +1028,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .is_quantized             = false,
         .to_float                 = (ggml_to_float_t) ggml_bf16_to_fp32_row,
         .from_float               = (ggml_from_float_t) ggml_fp32_to_bf16_row,
-        .from_float_ref           = (ggml_from_float_t) ggml_fp32_to_bf16_row,
+        .from_float_ref           = (ggml_from_float_t) ggml_fp32_to_bf16_row_ref,
         .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_bf16,
         .vec_dot_type             = GGML_TYPE_BF16,
         .nrows                    = 1,
@@ -2339,7 +2428,7 @@ inline static void ggml_vec_abs_f32  (const int n, float * y, const float * x) {
 inline static void ggml_vec_sgn_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : ((x[i] < 0.f) ? -1.f : 0.f); }
 inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : 0.f; }
 //inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = tanhf(x[i]);  }
-inline static void ggml_vec_elu_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expf(x[i])-1; }
+inline static void ggml_vec_elu_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expm1f(x[i]); }
 inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; }
 inline static void ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); }
 inline static void ggml_vec_sigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = 1.f / (1.f + expf(-x[i])); }
@@ -3544,7 +3633,7 @@ static inline int ggml_up(int n, int m) {
 }
 
 // assert that pointer is aligned to GGML_MEM_ALIGN
-#define ggml_assert_aligned(ptr) \
+#define GGML_ASSERT_ALIGNED(ptr) \
     GGML_ASSERT(((uintptr_t) (ptr))%GGML_MEM_ALIGN == 0)
 
 ////////////////////////////////////////////////////////////////////////////////
@@ -3645,7 +3734,13 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
 
     GGML_ASSERT(ctx->mem_buffer != NULL);
 
-    ggml_assert_aligned(ctx->mem_buffer);
+    GGML_ASSERT_ALIGNED(ctx->mem_buffer);
+
+#if defined(__ARM_FEATURE_SVE)
+    if (!ggml_sve_cnt_b) {
+        ggml_sve_cnt_b = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
+    }
+#endif
 
     GGML_PRINT_DEBUG("%s: context initialized\n", __func__);
 
@@ -3777,7 +3872,7 @@ static struct ggml_object * ggml_new_object(struct ggml_context * ctx, enum ggml
         .type = type,
     };
 
-    ggml_assert_aligned(mem_buffer + obj_new->offs);
+    GGML_ASSERT_ALIGNED(mem_buffer + obj_new->offs);
 
     if (obj_cur != NULL) {
         obj_cur->next = obj_new;
@@ -3801,7 +3896,8 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         struct ggml_tensor  * view_src,
         size_t                view_offs) {
 
-    assert(n_dims >= 1 && n_dims <= GGML_MAX_DIMS);
+    GGML_ASSERT(type >= 0 && type < GGML_TYPE_COUNT);
+    GGML_ASSERT(n_dims >= 1 && n_dims <= GGML_MAX_DIMS);
 
     // find the base tensor and absolute offset
     if (view_src != NULL && view_src->view_src != NULL) {
@@ -3878,7 +3974,7 @@ static struct ggml_tensor * ggml_new_tensor_impl(
 #endif
 
     // TODO: this should not be needed as long as we don't rely on aligned SIMD loads
-    //ggml_assert_aligned(result->data);
+    //GGML_ASSERT_ALIGNED(result->data);
 
     for (int i = 0; i < n_dims; i++) {
         result->ne[i] = ne[i];
@@ -4051,8 +4147,8 @@ struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value) {
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 
     return tensor;
@@ -4110,8 +4206,8 @@ struct ggml_tensor * ggml_set_f32(struct ggml_tensor * tensor, float value) {
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 
     return tensor;
@@ -4180,11 +4276,9 @@ int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i) {
             }
         default:
             {
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
             }
     }
-
-    return 0.0f;
 }
 
 void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value) {
@@ -4227,8 +4321,8 @@ void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value) {
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -4248,10 +4342,8 @@ int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i
         case GGML_TYPE_F32:
             return ((float *) data)[0];
         default:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
     }
-
-    return 0.0f;
 }
 
 void ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value) {
@@ -4283,8 +4375,8 @@ void ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2,
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -4321,11 +4413,9 @@ float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i) {
             }
         default:
             {
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
             }
     }
-
-    return 0.0f;
 }
 
 void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value) {
@@ -4362,8 +4452,8 @@ void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value) {
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -4383,10 +4473,8 @@ float ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2,
         case GGML_TYPE_F32:
             return ((float *) data)[0];
         default:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
     }
-
-    return 0.0f;
 }
 
 void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value) {
@@ -4418,8 +4506,8 @@ void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2,
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -4442,8 +4530,11 @@ const char * ggml_get_name(const struct ggml_tensor * tensor) {
 }
 
 struct ggml_tensor * ggml_set_name(struct ggml_tensor * tensor, const char * name) {
-    strncpy(tensor->name, name, sizeof(tensor->name) - 1);
-    tensor->name[sizeof(tensor->name) - 1] = '\0';
+    size_t i;
+    for (i = 0; i < sizeof(tensor->name) - 1 && name[i] != '\0'; i++) {
+        tensor->name[i] = name[i];
+    }
+    tensor->name[i] = '\0';
     return tensor;
 }
 
@@ -5014,7 +5105,7 @@ struct ggml_tensor * ggml_mean(
     bool is_node = false;
 
     if (a->grad) {
-        GGML_ASSERT(false); // TODO: implement
+        GGML_ABORT("fatal error"); // TODO: implement
         is_node = true;
     }
 
@@ -5037,7 +5128,7 @@ struct ggml_tensor * ggml_argmax(
     bool is_node = false;
 
     if (a->grad) {
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
         is_node = true;
     }
 
@@ -5360,7 +5451,7 @@ static struct ggml_tensor * ggml_norm_impl(
     bool is_node = false;
 
     if (!inplace && (a->grad)) {
-        GGML_ASSERT(false); // TODO: implement backward
+        GGML_ABORT("fatal error"); // TODO: implement backward
         is_node = true;
     }
 
@@ -5459,17 +5550,19 @@ static struct ggml_tensor * ggml_group_norm_impl(
     struct ggml_context * ctx,
     struct ggml_tensor * a,
     int n_groups,
+    float eps,
     bool inplace) {
 
     bool is_node = false;
     if (!inplace && (a->grad)) {
-        GGML_ASSERT(false); // TODO: implement backward
+        GGML_ABORT("fatal error"); // TODO: implement backward
         is_node = true;
     }
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    result->op_params[0] = n_groups;
+    ggml_set_op_params_i32(result, 0, n_groups);
+    ggml_set_op_params_f32(result, 1, eps);
 
     result->op = GGML_OP_GROUP_NORM;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -5481,15 +5574,17 @@ static struct ggml_tensor * ggml_group_norm_impl(
 struct ggml_tensor * ggml_group_norm(
     struct ggml_context * ctx,
     struct ggml_tensor * a,
-    int n_groups) {
-    return ggml_group_norm_impl(ctx, a, n_groups, false);
+    int n_groups,
+    float eps) {
+    return ggml_group_norm_impl(ctx, a, n_groups, eps, false);
 }
 
 struct ggml_tensor * ggml_group_norm_inplace(
     struct ggml_context * ctx,
     struct ggml_tensor * a,
-    int n_groups) {
-    return ggml_group_norm_impl(ctx, a, n_groups, true);
+    int n_groups,
+    float eps) {
+    return ggml_group_norm_impl(ctx, a, n_groups, eps, true);
 }
 
 // ggml_mul_mat
@@ -5877,7 +5972,7 @@ struct ggml_tensor * ggml_reshape(
 
     if (b->grad) {
         // gradient propagation is not supported
-        //GGML_ASSERT(false);
+        //GGML_ABORT("fatal error");
     }
 
     struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, GGML_MAX_DIMS, b->ne, a, 0);
@@ -6660,7 +6755,7 @@ struct ggml_tensor * ggml_clamp(
     bool is_node = false;
 
     if (a->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
+        GGML_ABORT("fatal error"); // TODO: implement backward
         is_node = true;
     }
 
@@ -6736,7 +6831,7 @@ GGML_API struct ggml_tensor * ggml_conv_transpose_1d(
     bool is_node = false;
 
     if (a->grad || b->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
+        GGML_ABORT("fatal error"); // TODO: implement backward
         is_node = true;
     }
 
@@ -6808,7 +6903,7 @@ struct ggml_tensor * ggml_im2col(
     bool is_node = false;
 
     if (a->grad || b->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
+        GGML_ABORT("fatal error"); // TODO: implement backward
         is_node = true;
     }
 
@@ -6894,7 +6989,7 @@ struct ggml_tensor * ggml_conv_transpose_2d_p0(
     bool is_node = false;
 
     if (a->grad || b->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
+        GGML_ABORT("fatal error"); // TODO: implement backward
         is_node = true;
     }
 
@@ -6935,7 +7030,7 @@ struct ggml_tensor * ggml_pool_1d(
     bool is_node = false;
 
     if (a->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
+        GGML_ABORT("fatal error"); // TODO: implement backward
         is_node = true;
     }
 
@@ -6973,7 +7068,7 @@ struct ggml_tensor * ggml_pool_2d(
     bool is_node = false;
 
     if (a->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
+        GGML_ABORT("fatal error"); // TODO: implement backward
         is_node = true;
     }
 
@@ -7006,7 +7101,7 @@ static struct ggml_tensor * ggml_upscale_impl(
     bool is_node = false;
 
     if (a->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
+        GGML_ABORT("fatal error"); // TODO: implement backward
         is_node = true;
     }
 
@@ -7056,7 +7151,7 @@ struct ggml_tensor * ggml_pad(
     bool is_node = false;
 
     if (a->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
+        GGML_ABORT("fatal error"); // TODO: implement backward
         is_node = true;
     }
 
@@ -7105,7 +7200,7 @@ struct ggml_tensor * ggml_timestep_embedding(
     bool is_node = false;
 
     if (timesteps->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
+        GGML_ABORT("fatal error"); // TODO: implement backward
         is_node = true;
     }
 
@@ -7231,7 +7326,7 @@ struct ggml_tensor * ggml_flash_attn_back(
         struct ggml_tensor  * v,
         struct ggml_tensor  * d,
         bool                  masked) {
-    GGML_ASSERT(false && "TODO: adapt to ggml_flash_attn_ext() changes");
+    GGML_ABORT("TODO: adapt to ggml_flash_attn_ext() changes");
 
     GGML_ASSERT(ggml_can_mul_mat(k, q));
     // TODO: check if vT can be multiplied by (k*qT)
@@ -7330,7 +7425,7 @@ struct ggml_tensor * ggml_ssm_conv(
     bool is_node = false;
 
     if (s->grad || x->grad || c->grad || sq->grad) {
-        GGML_ASSERT(false); // TODO: implement
+        GGML_ABORT("fatal error"); // TODO: implement
         is_node = true;
     }
 
@@ -7384,7 +7479,7 @@ struct ggml_tensor * ggml_ssm_scan(
     bool is_node = false;
 
     if (s->grad || x->grad || dt->grad || A->grad || B->grad || C->grad || sq->grad) {
-        GGML_ASSERT(false); // TODO: implement
+        GGML_ABORT("fatal error"); // TODO: implement
         is_node = true;
     }
 
@@ -7416,7 +7511,7 @@ struct ggml_tensor * ggml_win_part(
     bool is_node = false;
 
     if (a->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
+        GGML_ABORT("fatal error"); // TODO: implement backward
         is_node = true;
     }
 
@@ -7454,7 +7549,7 @@ struct ggml_tensor * ggml_win_unpart(
     bool is_node = false;
 
     if (a->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
+        GGML_ABORT("fatal error"); // TODO: implement backward
         is_node = true;
     }
 
@@ -7484,7 +7579,7 @@ struct ggml_tensor * ggml_get_rel_pos(
     bool is_node = false;
 
     if (a->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
+        GGML_ABORT("fatal error"); // TODO: implement backward
         is_node = true;
     }
 
@@ -8174,7 +8269,7 @@ static void ggml_compute_forward_dup_f16(
                     }
                 }
             } else {
-                GGML_ASSERT(false); // TODO: implement
+                GGML_ABORT("fatal error"); // TODO: implement
             }
         } else {
             //printf("%s: this is not optimal - fix me\n", __func__);
@@ -8216,7 +8311,7 @@ static void ggml_compute_forward_dup_f16(
                     }
                 }
             } else {
-                GGML_ASSERT(false); // TODO: implement
+                GGML_ABORT("fatal error"); // TODO: implement
             }
         }
         return;
@@ -8333,7 +8428,7 @@ static void ggml_compute_forward_dup_f16(
             }
         }
     } else {
-        GGML_ASSERT(false); // TODO: implement
+        GGML_ABORT("fatal error"); // TODO: implement
     }
 }
 
@@ -8460,7 +8555,7 @@ static void ggml_compute_forward_dup_bf16(
                     }
                 }
             } else {
-                GGML_ASSERT(false); // TODO: implement
+                GGML_ABORT("fatal error"); // TODO: implement
             }
         } else {
             //printf("%s: this is not optimal - fix me\n", __func__);
@@ -8520,7 +8615,7 @@ static void ggml_compute_forward_dup_bf16(
                     }
                 }
             } else {
-                GGML_ASSERT(false); // TODO: implement
+                GGML_ABORT("fatal error"); // TODO: implement
             }
         }
         return;
@@ -8689,7 +8784,7 @@ static void ggml_compute_forward_dup_bf16(
             }
         }
     } else {
-        GGML_ASSERT(false); // TODO: implement
+        GGML_ABORT("fatal error"); // TODO: implement
     }
 }
 
@@ -8775,7 +8870,7 @@ static void ggml_compute_forward_dup_f32(
                     }
                 }
             } else {
-                GGML_ASSERT(false); // TODO: implement
+                GGML_ABORT("fatal error"); // TODO: implement
             }
         } else {
             //printf("%s: this is not optimal - fix me\n", __func__);
@@ -8835,7 +8930,7 @@ static void ggml_compute_forward_dup_f32(
                     }
                 }
             } else {
-                GGML_ASSERT(false); // TODO: implement
+                GGML_ABORT("fatal error"); // TODO: implement
             }
         }
 
@@ -9006,7 +9101,7 @@ static void ggml_compute_forward_dup_f32(
             }
         }
     } else {
-        GGML_ASSERT(false); // TODO: implement
+        GGML_ABORT("fatal error"); // TODO: implement
     }
 }
 
@@ -9184,8 +9279,8 @@ static void ggml_compute_forward_dup(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -9337,7 +9432,7 @@ static void ggml_compute_forward_add_f16_f32(
     }
     else {
         // src1 is not contiguous
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 }
 
@@ -9412,7 +9507,7 @@ static void ggml_compute_forward_add_bf16_f32(
     }
     else {
         // src1 is not contiguous
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 }
 
@@ -9464,7 +9559,7 @@ static void ggml_compute_forward_add_f16_f16(
     }
     else {
         // src1 is not contiguous
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 }
 
@@ -9516,7 +9611,7 @@ static void ggml_compute_forward_add_bf16_bf16(
     }
     else {
         // src1 is not contiguous
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 }
 
@@ -9610,7 +9705,7 @@ static void ggml_compute_forward_add(
                     ggml_compute_forward_add_f32(params, dst);
                 }
                 else {
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
                 }
             } break;
         case GGML_TYPE_F16:
@@ -9622,7 +9717,7 @@ static void ggml_compute_forward_add(
                     ggml_compute_forward_add_f16_f32(params, dst);
                 }
                 else {
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
                 }
             } break;
         case GGML_TYPE_BF16:
@@ -9634,7 +9729,7 @@ static void ggml_compute_forward_add(
                     ggml_compute_forward_add_bf16_f32(params, dst);
                 }
                 else {
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
                 }
             } break;
         case GGML_TYPE_Q4_0:
@@ -9672,8 +9767,8 @@ static void ggml_compute_forward_add(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -10007,7 +10102,7 @@ static void ggml_compute_forward_add1(
                     ggml_compute_forward_add1_f16_f32(params, dst);
                 }
                 else {
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
                 }
             } break;
         case GGML_TYPE_BF16:
@@ -10019,7 +10114,7 @@ static void ggml_compute_forward_add1(
                     ggml_compute_forward_add1_bf16_f32(params, dst);
                 }
                 else {
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
                 }
             } break;
         case GGML_TYPE_Q4_0:
@@ -10058,8 +10153,8 @@ static void ggml_compute_forward_add1(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -10191,8 +10286,8 @@ static void ggml_compute_forward_acc(
         case GGML_TYPE_Q4_0_8_8:
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -10272,8 +10367,8 @@ static void ggml_compute_forward_sub(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -10383,8 +10478,8 @@ static void ggml_compute_forward_mul(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -10474,8 +10569,8 @@ static void ggml_compute_forward_div(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -10519,8 +10614,8 @@ static void ggml_compute_forward_sqr(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -10564,8 +10659,8 @@ static void ggml_compute_forward_sqrt(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -10609,8 +10704,8 @@ static void ggml_compute_forward_log(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -10738,8 +10833,8 @@ static void ggml_compute_forward_sum(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -10791,8 +10886,8 @@ static void ggml_compute_forward_sum_rows(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -10848,8 +10943,8 @@ static void ggml_compute_forward_mean(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -10896,8 +10991,8 @@ static void ggml_compute_forward_argmax(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11014,8 +11109,8 @@ static void ggml_compute_forward_repeat(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11092,8 +11187,8 @@ static void ggml_compute_forward_repeat_back(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11161,8 +11256,8 @@ static void ggml_compute_forward_concat(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11205,8 +11300,8 @@ static void ggml_compute_forward_abs(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11249,8 +11344,8 @@ static void ggml_compute_forward_sgn(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11293,8 +11388,8 @@ static void ggml_compute_forward_neg(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11337,8 +11432,8 @@ static void ggml_compute_forward_step(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11380,8 +11475,8 @@ static void ggml_compute_forward_tanh(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11424,8 +11519,8 @@ static void ggml_compute_forward_elu(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11468,8 +11563,8 @@ static void ggml_compute_forward_relu(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11512,8 +11607,8 @@ static void ggml_compute_forward_sigmoid(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11571,8 +11666,8 @@ static void ggml_compute_forward_gelu(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11630,8 +11725,8 @@ static void ggml_compute_forward_gelu_quick(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11689,8 +11784,8 @@ static void ggml_compute_forward_silu(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 // ggml_compute_forward_leaky_relu
@@ -11738,8 +11833,8 @@ static void ggml_compute_forward_leaky_relu(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11801,8 +11896,8 @@ static void ggml_compute_forward_silu_back(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11843,8 +11938,8 @@ static void ggml_compute_forward_hardswish(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11885,8 +11980,8 @@ static void ggml_compute_forward_hardsigmoid(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11957,8 +12052,8 @@ static void ggml_compute_forward_norm(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -12025,8 +12120,8 @@ static void ggml_compute_forward_rms_norm(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -12198,8 +12293,8 @@ static void ggml_compute_forward_rms_norm_back(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -12220,10 +12315,11 @@ static void ggml_compute_forward_group_norm_f32(
 
     GGML_TENSOR_UNARY_OP_LOCALS
 
-    const float eps = 1e-6f; // TODO: make this a parameter
-
     // TODO: optimize
 
+    float eps;
+    memcpy(&eps, dst->op_params + 1, sizeof(float));
+
     int n_channels = src0->ne[2];
     int n_groups = dst->op_params[0];
     int n_channels_per_group = (n_channels + n_groups - 1) / n_groups;
@@ -12292,8 +12388,8 @@ static void ggml_compute_forward_group_norm(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -12544,6 +12640,8 @@ UseGgmlGemm1:;
 IQK_MulMat_Not_Available2:;
 #endif
 
+    ggml_barrier(params->shared);
+
 #if GGML_USE_LLAMAFILE
     if (src1->type != vec_dot_type) {
         const size_t row_size = ggml_row_size(vec_dot_type, ne10);
@@ -12802,6 +12900,34 @@ IQK_MulMat_Not_Available:;
             continue;
         }
 
+        if (((ggml_n_dims(src0) - 1) == 2) && gemv) {
+            int64_t src0_cur_start = (ith * ne01) / nth;
+            int64_t src0_cur_end   = ((ith + 1) * ne01) / nth;
+            src0_cur_start = (src0_cur_start % matmul_num_cols) ? src0_cur_start + matmul_num_cols - (src0_cur_start % matmul_num_cols): src0_cur_start;
+            src0_cur_end   = (src0_cur_end % matmul_num_cols) ? src0_cur_end + matmul_num_cols - (src0_cur_end % matmul_num_cols): src0_cur_end;
+            if (src0_cur_start >= src0_cur_end) return;
+
+            for (int ir1 = 0; ir1 < nr1; ir1++) {
+                struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, ir1);
+                const int id       = row_mapping.i1; // selected expert index
+
+                const int64_t  i11 = id % ne11;
+                const int64_t  i12 = row_mapping.i2; // row index in src1
+
+                const int64_t  i1 = id;  // selected expert index
+                const int64_t  i2 = i12; // row
+
+                const char * src1_col = (const char *) wdata +
+                    (src1_cont || src1->type != vec_dot_type
+                    ? (i11        + i12 * ne11) * row_size
+                    : (i11 * nb11 + i12 * nb12));
+
+                gemv(ne00, (float *)((char *) dst->data + (i1 * nb1 + i2 * nb2)) + src0_cur_start, ne01,
+                     (const char *) src0_cur + src0_cur_start * nb01, src1_col, 1, src0_cur_end - src0_cur_start);
+            }
+            continue;
+        }
+
         // distribute the thread work across the inner or outer loop based on which one is larger
 
         const int64_t nth0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows
@@ -13119,17 +13245,17 @@ static void ggml_compute_forward_out_prod(
             } break;
         case GGML_TYPE_F16:
             {
-                GGML_ASSERT(false); // todo
+                GGML_ABORT("fatal error"); // todo
                 // ggml_compute_forward_out_prod_f16_f32(params, dst);
-            } break;
+            }
         case GGML_TYPE_F32:
             {
                 ggml_compute_forward_out_prod_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -13188,8 +13314,8 @@ static void ggml_compute_forward_scale(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -13312,8 +13438,8 @@ static void ggml_compute_forward_set(
         case GGML_TYPE_Q4_0_8_8:
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -13598,8 +13724,8 @@ static void ggml_compute_forward_get_rows(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 
     //static bool first = true;
@@ -13706,8 +13832,8 @@ static void ggml_compute_forward_get_rows_back(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 
     //static bool first = true;
@@ -13784,8 +13910,8 @@ static void ggml_compute_forward_diag(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -13854,8 +13980,8 @@ static void ggml_compute_forward_diag_mask_inf(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -13872,8 +13998,8 @@ static void ggml_compute_forward_diag_mask_zero(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -13990,8 +14116,8 @@ static void ggml_compute_forward_soft_max(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -14086,8 +14212,8 @@ static void ggml_compute_forward_soft_max_back(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -14186,8 +14312,8 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_F64:
         case GGML_TYPE_COUNT:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -14516,8 +14642,8 @@ static void ggml_compute_forward_rope(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -14540,8 +14666,8 @@ static void ggml_compute_forward_rope_back(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -14740,8 +14866,8 @@ static void ggml_compute_forward_conv_transpose_1d(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -14912,8 +15038,8 @@ static void ggml_compute_forward_im2col(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -15024,7 +15150,7 @@ static void ggml_compute_forward_pool_1d_sk_p0(
 
     const struct ggml_tensor * src = dst->src[0];
 
-    assert(src->type == GGML_TYPE_F32);
+    assert(src->type == GGML_TYPE_F32 || src->type == GGML_TYPE_F16);
 
     if (params->ith != 0) {
         return;
@@ -15037,28 +15163,27 @@ static void ggml_compute_forward_pool_1d_sk_p0(
     const int64_t rs = dst->ne[0];
 
     while (cdata < data_end) {
-        const float * const srow = (const float *)cdata;
-
+        const void * srow = (const void *)cdata;
         int j = 0;
-
         for (int64_t i = 0; i < rs; ++i) {
             switch (op) {
                 case GGML_OP_POOL_AVG:   drow[i] = 0;        break;
                 case GGML_OP_POOL_MAX:   drow[i] = -FLT_MAX; break;
-                case GGML_OP_POOL_COUNT: GGML_ASSERT(false); break;
+                case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error");
             }
             for (int ki = 0; ki < k; ++ki) {
+                const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]);
                 switch (op) {
-                    case GGML_OP_POOL_AVG:                          drow[i] += srow[j]; break;
-                    case GGML_OP_POOL_MAX:   if (srow[j] > drow[i]) drow[i]  = srow[j]; break;
-                    case GGML_OP_POOL_COUNT:                        GGML_ASSERT(false); break;
+                    case GGML_OP_POOL_AVG:                         drow[i] += srow_j; break;
+                    case GGML_OP_POOL_MAX:   if (srow_j > drow[i]) drow[i]  = srow_j; break;
+                    case GGML_OP_POOL_COUNT:                       GGML_ABORT("fatal error");
                 }
                 ++j;
             }
             switch (op) {
                 case GGML_OP_POOL_AVG:         drow[i] /= k; break;
                 case GGML_OP_POOL_MAX:                       break;
-                case GGML_OP_POOL_COUNT: GGML_ASSERT(false); break;
+                case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error");
             }
         }
 
@@ -15092,7 +15217,7 @@ static void ggml_compute_forward_pool_2d(
 
     const struct ggml_tensor * src = dst->src[0];
 
-    GGML_ASSERT(src->type == GGML_TYPE_F32);
+    assert(src->type == GGML_TYPE_F32 || src->type == GGML_TYPE_F16);
 
     if (params->ith != 0) {
         return;
@@ -15127,7 +15252,7 @@ static void ggml_compute_forward_pool_2d(
                 switch (op) {
                     case GGML_OP_POOL_AVG:     *out = 0;        break;
                     case GGML_OP_POOL_MAX:     *out = -FLT_MAX; break;
-                    case GGML_OP_POOL_COUNT: GGML_ASSERT(false); break;
+                    case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error");
                 }
 
                 const int ix = offset0 + ox * s0;
@@ -15135,21 +15260,22 @@ static void ggml_compute_forward_pool_2d(
 
                 for (int ky = 0; ky < k1; ++ky) {
                     if (iy + ky < 0 || iy + ky >= src->ne[1]) continue;
-                    const float * const srow = (const float *)(cdata + src->nb[1] * (iy + ky));
+                    const void * srow = (const void *)(cdata + src->nb[1] * (iy + ky));
                     for (int kx = 0; kx < k0; ++kx) {
                         int j = ix + kx;
                         if (j < 0 || j >= src->ne[0]) continue;
+                        const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]);
                         switch (op) {
-                            case GGML_OP_POOL_AVG:                     *out += srow[j]; break;
-                            case GGML_OP_POOL_MAX: if (srow[j] > *out) *out  = srow[j]; break;
-                            case GGML_OP_POOL_COUNT:                GGML_ASSERT(false); break;
+                            case GGML_OP_POOL_AVG:                     *out += srow_j; break;
+                            case GGML_OP_POOL_MAX: if (srow_j > *out)  *out  = srow_j; break;
+                            case GGML_OP_POOL_COUNT:               GGML_ABORT("fatal error");
                         }
                     }
                 }
                 switch (op) {
                     case GGML_OP_POOL_AVG:           *out /= ka; break;
                     case GGML_OP_POOL_MAX:                       break;
-                    case GGML_OP_POOL_COUNT: GGML_ASSERT(false); break;
+                    case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error");
                 }
             }
         }
@@ -15213,8 +15339,8 @@ static void ggml_compute_forward_upscale(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -15271,8 +15397,8 @@ static void ggml_compute_forward_pad(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -15312,8 +15438,8 @@ static void ggml_compute_forward_arange(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -15363,8 +15489,8 @@ static void ggml_compute_forward_timestep_embedding(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -15422,8 +15548,8 @@ static void ggml_compute_forward_argsort(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -15645,8 +15771,8 @@ static void ggml_compute_forward_flash_attn_ext(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -15981,8 +16107,8 @@ static void ggml_compute_forward_flash_attn_back(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -16103,8 +16229,8 @@ static void ggml_compute_forward_ssm_conv(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -16224,8 +16350,8 @@ static void ggml_compute_forward_ssm_scan(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -16287,8 +16413,8 @@ static void ggml_compute_forward_win_part(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -16348,8 +16474,8 @@ static void ggml_compute_forward_win_unpart(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -16416,8 +16542,8 @@ static void ggml_compute_forward_unary(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -16463,8 +16589,8 @@ static void ggml_compute_forward_get_rel_pos(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -16544,8 +16670,8 @@ static void ggml_compute_forward_add_rel_pos(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -16590,8 +16716,8 @@ static void ggml_compute_forward_map_unary(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -16639,8 +16765,8 @@ static void ggml_compute_forward_map_binary(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -16838,8 +16964,8 @@ static void ggml_compute_forward_cross_entropy_loss(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -16925,8 +17051,8 @@ static void ggml_compute_forward_cross_entropy_loss_back(
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -17261,14 +17387,32 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             } break;
         case GGML_OP_COUNT:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
 ////////////////////////////////////////////////////////////////////////////////
 
-static size_t ggml_hash_size(size_t min_sz) {
+struct ggml_hash_set ggml_hash_set_new(size_t size) {
+    size = ggml_hash_size(size);
+    struct ggml_hash_set result;
+    result.size = size;
+    result.keys = GGML_MALLOC(sizeof(struct ggml_tensor *) * size);
+    result.used = GGML_CALLOC(ggml_bitset_size(size), sizeof(ggml_bitset_t));
+    return result;
+}
+
+void ggml_hash_set_reset(struct ggml_hash_set * hash_set) {
+    memset(hash_set->used, 0, sizeof(ggml_bitset_t) * ggml_bitset_size(hash_set->size));
+}
+
+void ggml_hash_set_free(struct ggml_hash_set * hash_set) {
+    GGML_FREE(hash_set->used);
+    GGML_FREE(hash_set->keys);
+}
+
+size_t ggml_hash_size(size_t min_sz) {
     // next primes after powers of two
     static const size_t primes[] = {
         2, 3, 5, 11, 17, 37, 67, 131, 257, 521, 1031,
@@ -17279,7 +17423,7 @@ static size_t ggml_hash_size(size_t min_sz) {
     };
     static const size_t n_primes = sizeof(primes)/sizeof(primes[0]);
 
-    // find the smallest prime that is larger or equal to min_sz
+    // find the smallest prime that is larger or equal than min_sz
     size_t l = 0;
     size_t r = n_primes;
     while (l < r) {
@@ -17294,67 +17438,6 @@ static size_t ggml_hash_size(size_t min_sz) {
     return sz;
 }
 
-static size_t ggml_hash(const void * p) {
-    return (size_t)p;
-}
-
-size_t ggml_hash_find(const struct ggml_hash_set hash_set, struct ggml_tensor * key) {
-    size_t h = ggml_hash(key) % hash_set.size;
-
-    // linear probing
-    size_t i = h;
-    while (hash_set.keys[i] != NULL && hash_set.keys[i] != key) {
-        i = (i + 1) % hash_set.size;
-        if (i == h) {
-            // visited all hash table entries -> not found
-            return GGML_HASHTABLE_FULL;
-        }
-    }
-    return i;
-}
-
-bool ggml_hash_contains(struct ggml_hash_set hash_set, struct ggml_tensor * key) {
-    size_t i = ggml_hash_find(hash_set, key);
-    return i != GGML_HASHTABLE_FULL && hash_set.keys[i] == key;
-}
-
-size_t ggml_hash_insert(struct ggml_hash_set hash_set, struct ggml_tensor * key) {
-    size_t i = ggml_hash_find(hash_set, key);
-
-    GGML_ASSERT(i != GGML_HASHTABLE_FULL);
-
-    if (hash_set.keys[i] == key) {
-        return GGML_HASHTABLE_ALREADY_EXISTS;
-    }
-
-    // insert
-    GGML_ASSERT(hash_set.keys[i] == NULL);
-    hash_set.keys[i] = key;
-    return i;
-}
-
-size_t ggml_hash_find_or_insert(struct ggml_hash_set hash_set, struct ggml_tensor * key) {
-    size_t i = ggml_hash_find(hash_set, key);
-
-    GGML_ASSERT(i != GGML_HASHTABLE_FULL);
-
-    hash_set.keys[i] = key;
-    return i;
-}
-
-struct ggml_hash_set ggml_hash_set_new(size_t size) {
-    size = ggml_hash_size(size);
-    struct ggml_hash_set result;
-    result.size = size;
-    result.keys = GGML_MALLOC(sizeof(struct ggml_tensor *) * size);
-    memset(result.keys, 0, sizeof(struct ggml_tensor *) * size);
-    return result;
-}
-
-static void ggml_hash_set_free(struct ggml_hash_set hash_set) {
-    GGML_FREE(hash_set.keys);
-}
-
 struct hash_map {
     struct ggml_hash_set set;
     struct ggml_tensor ** vals;
@@ -17363,13 +17446,12 @@ struct hash_map {
 static struct hash_map * ggml_new_hash_map(size_t size) {
     struct hash_map * result = GGML_MALLOC(sizeof(struct hash_map));
     result->set = ggml_hash_set_new(size);
-    result->vals = GGML_MALLOC(sizeof(struct ggml_tensor *) * result->set.size);
-    memset(result->vals, 0, sizeof(struct ggml_tensor *) * result->set.size);
+    result->vals = GGML_CALLOC(result->set.size, sizeof(struct ggml_tensor *));
     return result;
 }
 
 static void ggml_hash_map_free(struct hash_map * map) {
-    ggml_hash_set_free(map->set);
+    ggml_hash_set_free(&map->set);
     GGML_FREE(map->vals);
     GGML_FREE(map);
 }
@@ -17390,7 +17472,7 @@ static struct ggml_tensor * ggml_recompute_graph_node(
         return node;
     }
 
-    if (!ggml_hash_contains(graph->visited_hash_table, node)) {
+    if (!ggml_hash_contains(&graph->visited_hash_set, node)) {
         return node;
     }
 
@@ -17405,8 +17487,8 @@ static struct ggml_tensor * ggml_recompute_graph_node(
         return node;
     }
 
-    size_t i = ggml_hash_find(replacements->set, node);
-    GGML_ASSERT(i != GGML_HASHTABLE_FULL); // assert that not full
+    size_t i = ggml_hash_find(&replacements->set, node);
+    GGML_ASSERT(i != GGML_HASHSET_FULL); // assert that not full
     if (replacements->set.keys[i] == node) {
         return replacements->vals[i];
     }
@@ -17464,8 +17546,8 @@ void ggml_build_backward_gradient_checkpointing(
 
     // insert checkpoints in replacements
     for (int i = 0; i < n_checkpoints; ++i) {
-        size_t k = ggml_hash_find(replacements->set, checkpoints[i]);
-        GGML_ASSERT(k != GGML_HASHTABLE_FULL); // assert that not full
+        size_t k = ggml_hash_find(&replacements->set, checkpoints[i]);
+        GGML_ASSERT(k != GGML_HASHSET_FULL); // assert that not full
         GGML_ASSERT(replacements->set.keys[k] == NULL); // assert that we don't overwrite
         replacements->set.keys[k] = checkpoints[i];
         replacements->vals[k]     = checkpoints[i];
@@ -17493,7 +17575,7 @@ void ggml_build_backward_gradient_checkpointing(
 
 // functions to change gradients considering the case that input a might be initial gradient with zero value
 
-static struct ggml_tensor * ggml_add_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, struct ggml_hash_set zero_table) {
+static struct ggml_tensor * ggml_add_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, struct ggml_hash_set * zero_table) {
     if (ggml_hash_contains(zero_table, a)) {
         return b;
     } else {
@@ -17501,7 +17583,7 @@ static struct ggml_tensor * ggml_add_or_set(struct ggml_context * ctx, struct gg
     }
 }
 
-static struct ggml_tensor * ggml_acc_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, size_t nb1, size_t nb2, size_t nb3, size_t offset, struct ggml_hash_set zero_table) {
+static struct ggml_tensor * ggml_acc_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, size_t nb1, size_t nb2, size_t nb3, size_t offset, struct ggml_hash_set * zero_table) {
     if (ggml_hash_contains(zero_table, a)) {
         struct ggml_tensor * a_zero = ggml_scale(ctx, a, 0.0f);
         return ggml_acc_impl(ctx, a_zero, b, nb1, nb2, nb3, offset, false);
@@ -17510,7 +17592,7 @@ static struct ggml_tensor * ggml_acc_or_set(struct ggml_context * ctx, struct gg
     }
 }
 
-static struct ggml_tensor * ggml_add1_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, struct ggml_hash_set zero_table) {
+static struct ggml_tensor * ggml_add1_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, struct ggml_hash_set * zero_table) {
     if (ggml_hash_contains(zero_table, a)) {
         return ggml_repeat(ctx, b, a);
     } else {
@@ -17518,7 +17600,7 @@ static struct ggml_tensor * ggml_add1_or_set(struct ggml_context * ctx, struct g
     }
 }
 
-static struct ggml_tensor * ggml_sub_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, struct ggml_hash_set zero_table) {
+static struct ggml_tensor * ggml_sub_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, struct ggml_hash_set * zero_table) {
     if (ggml_hash_contains(zero_table, a)) {
         return ggml_neg(ctx, b);
     } else {
@@ -17526,7 +17608,7 @@ static struct ggml_tensor * ggml_sub_or_set(struct ggml_context * ctx, struct gg
     }
 }
 
-static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor * tensor, struct ggml_hash_set zero_table) {
+static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor * tensor, struct ggml_hash_set * zero_table) {
     struct ggml_tensor * src0 = tensor->src[0];
     struct ggml_tensor * src1 = tensor->src[1];
     struct ggml_tensor * src2 = tensor->src[2];
@@ -17695,8 +17777,8 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
         case GGML_OP_MEAN:
         case GGML_OP_ARGMAX:
             {
-                GGML_ASSERT(false); // TODO: implement
-            } break;
+                GGML_ABORT("fatal error"); // TODO: implement
+            }
         case GGML_OP_REPEAT:
             {
                 // necessary for llama
@@ -17719,16 +17801,16 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             } break;
         case GGML_OP_CONCAT:
             {
-                GGML_ASSERT(false); // TODO: implement
-            } break;
+                GGML_ABORT("fatal error"); // TODO: implement
+            }
         case GGML_OP_SILU_BACK:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_NORM:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_RMS_NORM:
             {
                 // necessary for llama
@@ -17744,12 +17826,12 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             } break;
         case GGML_OP_RMS_NORM_BACK:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_GROUP_NORM:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_MUL_MAT:
             {
                 // https://cs231n.github.io/optimization-2/#staged
@@ -17810,12 +17892,12 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             } break;
         case GGML_OP_MUL_MAT_ID:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_OUT_PROD:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_SCALE:
             {
                 // necessary for llama
@@ -17991,12 +18073,12 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             } break;
         case GGML_OP_GET_ROWS_BACK:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_DIAG:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_DIAG_MASK_INF:
             {
                 // necessary for llama
@@ -18034,8 +18116,8 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             } break;
         case GGML_OP_SOFT_MAX_BACK:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_ROPE:
             {
                 // necessary for llama
@@ -18110,52 +18192,52 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             } break;
         case GGML_OP_CLAMP:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_CONV_TRANSPOSE_1D:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_IM2COL:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_CONV_TRANSPOSE_2D:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_POOL_1D:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_POOL_2D:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_UPSCALE:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_PAD:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_ARANGE:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_TIMESTEP_EMBEDDING:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_ARGSORT:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_LEAKY_RELU:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_FLASH_ATTN_EXT:
             {
                 struct ggml_tensor * flash_grad = NULL;
@@ -18211,13 +18293,13 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             } break;
         case GGML_OP_FLASH_ATTN_BACK:
             {
-                GGML_ASSERT(false); // not supported
-            } break;
+                GGML_ABORT("fatal error"); // not supported
+            }
         case GGML_OP_SSM_CONV:
         case GGML_OP_SSM_SCAN:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_WIN_PART:
         case GGML_OP_WIN_UNPART:
         case GGML_OP_UNARY:
@@ -18255,12 +18337,12 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                         } break;
                     case GGML_UNARY_OP_TANH:
                         {
-                            GGML_ASSERT(false); // TODO: not implemented
-                        } break;
+                            GGML_ABORT("fatal error"); // TODO: not implemented
+                        }
                     case GGML_UNARY_OP_ELU:
                         {
-                            GGML_ASSERT(false); // TODO: not implemented
-                        } break;
+                            GGML_ABORT("fatal error"); // TODO: not implemented
+                        }
                     case GGML_UNARY_OP_RELU:
                         {
                             if (src0->grad) {
@@ -18274,16 +18356,16 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                         } break;
                     case GGML_UNARY_OP_SIGMOID:
                         {
-                            GGML_ASSERT(false); // TODO: not implemented
-                        } break;
+                            GGML_ABORT("fatal error"); // TODO: not implemented
+                        }
                     case GGML_UNARY_OP_GELU:
                         {
-                            GGML_ASSERT(false); // TODO: not implemented
-                        } break;
+                            GGML_ABORT("fatal error"); // TODO: not implemented
+                        }
                     case GGML_UNARY_OP_GELU_QUICK:
                         {
-                            GGML_ASSERT(false); // TODO: not implemented
-                        } break;
+                            GGML_ABORT("fatal error"); // TODO: not implemented
+                        }
                     case GGML_UNARY_OP_SILU:
                         {
                             // necessary for llama
@@ -18295,7 +18377,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                             }
                         } break;
                     default:
-                        GGML_ASSERT(false);
+                        GGML_ABORT("fatal error");
                 }
             } break;
         case GGML_OP_GET_REL_POS:
@@ -18309,8 +18391,8 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
         case GGML_OP_MAP_CUSTOM2:
         case GGML_OP_MAP_CUSTOM3:
             {
-                GGML_ASSERT(false); // not supported
-            } break;
+                GGML_ABORT("fatal error"); // not supported
+            }
         case GGML_OP_CROSS_ENTROPY_LOSS:
             {
                 if (src0->grad) {
@@ -18325,16 +18407,16 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             } break;
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
             {
-                GGML_ASSERT(false); // not supported
-            } break;
+                GGML_ABORT("fatal error"); // not supported
+            }
         case GGML_OP_NONE:
             {
                 // nop
             } break;
         case GGML_OP_COUNT:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 
     for (int i = 0; i < GGML_MAX_SRC; ++i) {
@@ -18354,7 +18436,7 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor *
     }
 
     // check if already visited
-    if (ggml_hash_insert(cgraph->visited_hash_table, node) == GGML_HASHTABLE_ALREADY_EXISTS) {
+    if (ggml_hash_insert(&cgraph->visited_hash_set, node) == GGML_HASHSET_ALREADY_EXISTS) {
         return;
     }
 
@@ -18400,7 +18482,6 @@ static void ggml_build_forward_impl(struct ggml_cgraph * cgraph, struct ggml_ten
     }
 
     const int n0 = cgraph->n_nodes;
-    UNUSED(n0);
 
     ggml_visit_parents(cgraph, tensor);
 
@@ -18436,7 +18517,7 @@ void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph *
     struct ggml_hash_set zero_table = ggml_hash_set_new(gf->size);
     for (int i = 0; i < gf->n_nodes; i++) {
         if (gf->grads[i]) {
-            ggml_hash_insert(zero_table, gf->grads[i]);
+            ggml_hash_insert(&zero_table, gf->grads[i]);
         }
     }
 
@@ -18446,7 +18527,7 @@ void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph *
         // inplace operations to add gradients are not created by ggml_compute_backward
         // use allocator to automatically make inplace operations
         if (node->grad) {
-            ggml_compute_backward(ctx, node, zero_table);
+            ggml_compute_backward(ctx, node, &zero_table);
         }
     }
 
@@ -18459,16 +18540,29 @@ void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph *
         }
     }
 
-    ggml_hash_set_free(zero_table);
+    ggml_hash_set_free(&zero_table);
+}
+
+static void * incr_ptr_aligned(void ** p, size_t size, size_t align) {
+    void * ptr = *p;
+    ptr = (void *) GGML_PAD((uintptr_t) ptr, align);
+    *p = (void *) ((char *) ptr + size);
+    return ptr;
 }
 
 static size_t ggml_graph_nbytes(size_t size, bool grads) {
-    size_t nbytes = sizeof(struct ggml_cgraph);
-    nbytes += size * sizeof(struct ggml_tensor *) * 2; // leafs + nodes
+    size_t hash_size = ggml_hash_size(size * 2);
+    void * p = 0;
+    incr_ptr_aligned(&p, sizeof(struct ggml_cgraph), 1);
+    incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // nodes
+    incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // leafs
+    incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // hash keys
     if (grads) {
-        nbytes += size * sizeof(struct ggml_tensor *); // grads
+        incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // grads
     }
-    nbytes += ggml_hash_size(size * 2) * sizeof(struct ggml_tensor *); // hash set
+    incr_ptr_aligned(&p, ggml_bitset_size(hash_size) * sizeof(ggml_bitset_t), sizeof(ggml_bitset_t));
+
+    size_t nbytes = (size_t) p;
     return nbytes;
 }
 
@@ -18485,19 +18579,19 @@ struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t siz
     struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_TYPE_GRAPH, obj_size);
     struct ggml_cgraph * cgraph = (struct ggml_cgraph *) ((char *) ctx->mem_buffer + obj->offs);
 
-    struct ggml_tensor ** data_start = (struct ggml_tensor **) (cgraph + 1);
-
+    // the size of the hash table is doubled since it needs to hold both nodes and leafs
     size_t hash_size = ggml_hash_size(size * 2);
-    struct ggml_tensor ** nodes_ptr = data_start;
-    struct ggml_tensor ** leafs_ptr = nodes_ptr + size;
-    struct ggml_tensor ** hash_keys_ptr = leafs_ptr + size;
-    struct ggml_tensor ** grads_ptr = grads ? hash_keys_ptr + hash_size : NULL;
 
-    // check that we allocated the correct amount of memory
-    assert(obj_size == (size_t) (
-        (grads ? (char *)(grads_ptr + size) : (char *)(hash_keys_ptr + hash_size)) - (char *)cgraph));
+    void * p = cgraph + 1;
 
-    memset(hash_keys_ptr, 0, hash_size * sizeof(struct ggml_tensor *));
+    struct ggml_tensor ** nodes_ptr = incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
+    struct ggml_tensor ** leafs_ptr = incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
+    struct ggml_tensor ** hash_keys_ptr = incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
+    struct ggml_tensor ** grads_ptr = grads ? incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)) : NULL;
+    ggml_bitset_t * hash_used = incr_ptr_aligned(&p, ggml_bitset_size(hash_size) * sizeof(ggml_bitset_t), sizeof(ggml_bitset_t));
+
+    // check that we allocated the correct amount of memory
+    assert(obj_size == (size_t)((char *)p - (char *)cgraph));
 
     *cgraph = (struct ggml_cgraph) {
         /*.size         =*/ size,
@@ -18506,10 +18600,12 @@ struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t siz
         /*.nodes        =*/ nodes_ptr,
         /*.grads        =*/ grads_ptr,
         /*.leafs        =*/ leafs_ptr,
-        /*.hash_table   =*/ { hash_size, hash_keys_ptr },
+        /*.hash_table   =*/ { hash_size, hash_used, hash_keys_ptr },
         /*.order        =*/ GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT,
     };
 
+    ggml_hash_set_reset(&cgraph->visited_hash_set);
+
     return cgraph;
 }
 
@@ -18525,7 +18621,7 @@ struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph0, int i0, int i1)
         /*.nodes        =*/ cgraph0->nodes + i0,
         /*.grads        =*/ cgraph0->grads ? cgraph0->grads + i0 : NULL,
         /*.leafs        =*/ NULL,
-        /*.hash_table   =*/ { 0, NULL },
+        /*.hash_table   =*/ { 0, NULL, NULL },
         /*.order        =*/ cgraph0->order,
     };
 
@@ -18535,7 +18631,7 @@ struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph0, int i0, int i1)
 void ggml_graph_cpy(struct ggml_cgraph * src, struct ggml_cgraph * dst) {
     GGML_ASSERT(dst->size >= src->n_leafs);
     GGML_ASSERT(dst->size >= src->n_nodes);
-    GGML_ASSERT(dst->visited_hash_table.size >= src->visited_hash_table.size);
+    GGML_ASSERT(dst->visited_hash_set.size >= src->visited_hash_set.size);
 
     dst->n_leafs = src->n_leafs;
     dst->n_nodes = src->n_nodes;
@@ -18556,9 +18652,9 @@ void ggml_graph_cpy(struct ggml_cgraph * src, struct ggml_cgraph * dst) {
         }
     }
 
-    for (size_t i = 0; i < src->visited_hash_table.size; ++i) {
-        if (src->visited_hash_table.keys[i]) {
-            ggml_hash_insert(dst->visited_hash_table, src->visited_hash_table.keys[i]);
+    for (size_t i = 0; i < src->visited_hash_set.size; ++i) {
+        if (src->visited_hash_set.keys[i]) {
+            ggml_hash_insert(&dst->visited_hash_set, src->visited_hash_set.keys[i]);
         }
     }
 }
@@ -18584,7 +18680,7 @@ void ggml_graph_reset(struct ggml_cgraph * cgraph) {
 void ggml_graph_clear(struct ggml_cgraph * cgraph) {
     cgraph->n_leafs = 0;
     cgraph->n_nodes = 0;
-    memset(cgraph->visited_hash_table.keys, 0, cgraph->visited_hash_table.size * sizeof(struct ggml_tensor *));
+    ggml_hash_set_reset(&cgraph->visited_hash_set);
 }
 
 //
@@ -18779,7 +18875,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                         n_tasks = MIN(ggml_nrows(node), n_threads);
                     } break;
                 default:
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
             }
             break;
         case GGML_OP_SILU_BACK:
@@ -18906,8 +19002,8 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             } break;
         case GGML_OP_COUNT:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
         default:
             {
                 fprintf(stderr, "%s: op not implemented: ", __func__);
@@ -18916,8 +19012,8 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 } else {
                     fprintf(stderr, "%d\n", node->op);
                 }
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 
     assert(n_tasks > 0);
@@ -19027,7 +19123,7 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa
                         cur += sizeof(float)*ne00*ne01*ne02;
                         cur += sizeof(float)*ne10*ne11;
                     } else {
-                        GGML_ASSERT(false);
+                        GGML_ABORT("fatal error");
                     }
                 } break;
             case GGML_OP_CONV_TRANSPOSE_2D:
@@ -19073,8 +19169,8 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa
                 } break;
             case GGML_OP_COUNT:
                 {
-                    GGML_ASSERT(false);
-                } break;
+                    GGML_ABORT("fatal error");
+                }
             default:
                 break;
         }
@@ -20308,9 +20404,9 @@ static enum ggml_opt_result linesearch_backtracking(
         (*step) *= width;
     }
 
-    GGML_ASSERT(false && "line search failed");
+    GGML_ABORT("line search failed");
 
-    return GGML_LINESEARCH_FAIL;
+    //return GGML_LINESEARCH_FAIL;
 }
 
 static enum ggml_opt_result ggml_opt_lbfgs(
@@ -20578,9 +20674,9 @@ static enum ggml_opt_result ggml_opt_lbfgs(
         step[0] = 1.0;
     }
 
-    GGML_ASSERT(false && "lbfgs failed");
+    GGML_ABORT("lbfgs failed");
 
-    return GGML_OPT_RESULT_DID_NOT_CONVERGE;
+    //return GGML_OPT_RESULT_DID_NOT_CONVERGE;
 }
 
 struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type) {
@@ -20925,7 +21021,7 @@ size_t ggml_quantize_chunk(
         case GGML_TYPE_BF16:
             {
                 size_t elemsize = sizeof(ggml_bf16_t);
-                ggml_fp32_to_bf16_row(src + start, (ggml_bf16_t *)dst + start, n);
+                ggml_fp32_to_bf16_row_ref(src + start, (ggml_bf16_t *)dst + start, n);
                 result = n * elemsize;
             } break;
         case GGML_TYPE_F32:
@@ -21283,10 +21379,10 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
                                     }
                                 } break;
                             case GGUF_TYPE_ARRAY:
-                            default: GGML_ASSERT(false && "invalid type"); break;
+                            default: GGML_ABORT("invalid type");
                         }
                     } break;
-                default: GGML_ASSERT(false && "invalid type");
+                default: GGML_ABORT("invalid type");
             }
 
             if (!ok) {
@@ -21453,7 +21549,12 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
                 ctx->infos[i].ne[3],
             };
 
-            struct ggml_tensor * cur = ggml_new_tensor(ctx_data, ctx->infos[i].type, ctx->infos[i].n_dims, ne);
+            int n_dims = ctx->infos[i].n_dims;
+            if (n_dims == 0 || n_dims > 4) {
+                n_dims = 4;
+                for (; n_dims > 1; --n_dims) if (ne[n_dims-1] > 1) break;
+            }
+            struct ggml_tensor * cur = ggml_new_tensor(ctx_data, ctx->infos[i].type, n_dims, ne);
 
             ok = ok && cur != NULL;
 
@@ -21867,12 +21968,12 @@ void gguf_set_kv(struct gguf_context * ctx, struct gguf_context * src) {
                         gguf_set_arr_str(ctx, src->kv[i].key.data, data, src->kv[i].value.arr.n);
                         GGML_FREE((void *)data);
                     } else if (src->kv[i].value.arr.type == GGUF_TYPE_ARRAY) {
-                        GGML_ASSERT(false && "nested arrays not supported");
+                        GGML_ABORT("nested arrays not supported");
                     } else {
                         gguf_set_arr_data(ctx, src->kv[i].key.data, src->kv[i].value.arr.type, src->kv[i].value.arr.data, src->kv[i].value.arr.n);
                     }
                 } break;
-            default: GGML_ASSERT(false && "invalid type"); break;
+            default: GGML_ABORT("invalid type");
         }
     }
 }
@@ -21881,7 +21982,7 @@ void gguf_add_tensor(
              struct gguf_context * ctx,
         const struct ggml_tensor * tensor) {
     if (gguf_find_tensor(ctx, tensor->name) != -1) {
-        GGML_ASSERT(false && "duplicated tensor name");
+        GGML_ABORT("duplicated tensor name");
     }
 
     const int idx = ctx->header.n_tensors;
@@ -21914,7 +22015,7 @@ void gguf_add_tensor(
 void gguf_set_tensor_type(struct gguf_context * ctx, const char * name, enum ggml_type type) {
     const int idx = gguf_find_tensor(ctx, name);
     if (idx < 0) {
-        GGML_ASSERT(false && "tensor not found");
+        GGML_ABORT("tensor not found");
     }
 
     ctx->infos[idx].type = type;
@@ -21923,7 +22024,7 @@ void gguf_set_tensor_type(struct gguf_context * ctx, const char * name, enum ggm
 void gguf_set_tensor_data(struct gguf_context * ctx, const char * name, const void * data, size_t size) {
     const int idx = gguf_find_tensor(ctx, name);
     if (idx < 0) {
-        GGML_ASSERT(false && "tensor not found");
+        GGML_ABORT("tensor not found");
     }
 
     ctx->infos[idx].data = data;
@@ -22052,10 +22153,10 @@ static void gguf_write_to_buf(const struct gguf_context * ctx, struct gguf_buf *
                                 }
                             } break;
                         case GGUF_TYPE_ARRAY:
-                        default: GGML_ASSERT(false && "invalid type"); break;
+                        default: GGML_ABORT("invalid type");
                     }
                 } break;
-            default: GGML_ASSERT(false && "invalid type");
+            default: GGML_ABORT("invalid type");
         }
     }
 
@@ -22116,7 +22217,7 @@ static void gguf_write_to_buf(const struct gguf_context * ctx, struct gguf_buf *
 void gguf_write_to_file(const struct gguf_context * ctx, const char * fname, bool only_meta) {
     FILE * file = ggml_fopen(fname, "wb");
     if (!file) {
-        GGML_ASSERT(false && "failed to open file for writing");
+        GGML_ABORT("failed to open file for writing");
     }
 
     struct gguf_buf buf = gguf_buf_init(16*1024);
diff --git a/ggml/src/vulkan-shaders/CMakeLists.txt b/ggml/src/vulkan-shaders/CMakeLists.txt
index 41551e00..10075db3 100644
--- a/ggml/src/vulkan-shaders/CMakeLists.txt
+++ b/ggml/src/vulkan-shaders/CMakeLists.txt
@@ -1,5 +1,7 @@
+find_package (Threads REQUIRED)
 
 set(TARGET vulkan-shaders-gen)
 add_executable(${TARGET} vulkan-shaders-gen.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_link_libraries(vulkan-shaders-gen PUBLIC Threads::Threads)
diff --git a/ggml/src/vulkan-shaders/add.comp b/ggml/src/vulkan-shaders/add.comp
index 8475b011..3974845d 100644
--- a/ggml/src/vulkan-shaders/add.comp
+++ b/ggml/src/vulkan-shaders/add.comp
@@ -4,9 +4,11 @@
 #include "generic_binary_head.comp"
 
 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
         return;
     }
 
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) + FLOAT_TYPE(data_b[src1_idx(gl_GlobalInvocationID.x)]));
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) + FLOAT_TYPE(data_b[src1_idx(idx)]));
 }
diff --git a/ggml/src/vulkan-shaders/clamp.comp b/ggml/src/vulkan-shaders/clamp.comp
index ca272e22..7071302a 100644
--- a/ggml/src/vulkan-shaders/clamp.comp
+++ b/ggml/src/vulkan-shaders/clamp.comp
@@ -4,10 +4,12 @@
 #include "generic_unary_head.comp"
 
 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
         return;
     }
 
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]);
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
 }
diff --git a/ggml/src/vulkan-shaders/concat.comp b/ggml/src/vulkan-shaders/concat.comp
new file mode 100644
index 00000000..08ab5514
--- /dev/null
+++ b/ggml/src/vulkan-shaders/concat.comp
@@ -0,0 +1,35 @@
+#version 450
+
+#include "types.comp"
+#include "generic_binary_head.comp"
+
+void main() {
+    const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+    const int dim = p.param3;
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint i3 = idx / (p.ne22*p.ne21*p.ne20);
+    const uint i3_offset = i3 * p.ne22*p.ne21*p.ne20;
+    const uint i2 = (idx - i3_offset) / (p.ne21*p.ne20);
+    const uint i2_offset = i2*p.ne21*p.ne20;
+    const uint i1 = (idx - i3_offset - i2_offset) / p.ne20;
+    const uint i0 = idx - i3_offset - i2_offset - i1*p.ne20;
+
+    uint o[4] = {0, 0, 0, 0};
+    o[dim] = dim == 0 ? p.ne00 : (dim == 1 ? p.ne01 : (dim == 2 ? p.ne02 : p.ne03));
+
+    const uint src0_idx = i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0*p.nb00;
+    const uint src1_idx = (i3 - o[3])*p.nb13 + (i2 - o[2])*p.nb12 + (i1 - o[1])*p.nb11 + (i0 - o[0])*p.nb10;
+    const uint dst_idx = i3*p.nb23 + i2*p.nb22 + i1*p.nb21 + i0*p.nb20;
+
+    const bool is_src0 = i0 < p.ne00 && i1 < p.ne01 && i2 < p.ne02 && i3 < p.ne03;
+
+#ifndef OPTIMIZATION_ERROR_WORKAROUND
+    data_d[p.d_offset + dst_idx] = D_TYPE(is_src0 ? data_a[src0_idx] : data_b[src1_idx]);
+#else
+    data_d[p.d_offset + dst_idx] = is_src0 ? data_a[src0_idx] : data_b[src1_idx];
+#endif
+}
diff --git a/ggml/src/vulkan-shaders/copy.comp b/ggml/src/vulkan-shaders/copy.comp
index efb55876..c26917c0 100644
--- a/ggml/src/vulkan-shaders/copy.comp
+++ b/ggml/src/vulkan-shaders/copy.comp
@@ -4,13 +4,15 @@
 #include "generic_unary_head.comp"
 
 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
         return;
     }
 
 #ifndef OPTIMIZATION_ERROR_WORKAROUND
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]);
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(data_a[src0_idx(idx)]);
 #else
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = data_a[src0_idx(gl_GlobalInvocationID.x)];
+    data_d[p.d_offset + dst_idx(idx)] = data_a[src0_idx(idx)];
 #endif
 }
diff --git a/ggml/src/vulkan-shaders/div.comp b/ggml/src/vulkan-shaders/div.comp
index 8ee4bfc7..8cfce58b 100644
--- a/ggml/src/vulkan-shaders/div.comp
+++ b/ggml/src/vulkan-shaders/div.comp
@@ -4,9 +4,11 @@
 #include "generic_binary_head.comp"
 
 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
         return;
     }
 
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) / FLOAT_TYPE(data_b[src1_idx(gl_GlobalInvocationID.x)]));
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) / FLOAT_TYPE(data_b[src1_idx(idx)]));
 }
diff --git a/ggml/src/vulkan-shaders/gelu.comp b/ggml/src/vulkan-shaders/gelu.comp
index 9fe807cc..4cc7a68c 100644
--- a/ggml/src/vulkan-shaders/gelu.comp
+++ b/ggml/src/vulkan-shaders/gelu.comp
@@ -13,7 +13,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 void main() {
     const float GELU_COEF_A    = 0.044715f;
     const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
-    const uint i = gl_GlobalInvocationID.x;
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
 
     if (i >= p.KX) {
         return;
diff --git a/ggml/src/vulkan-shaders/gelu_quick.comp b/ggml/src/vulkan-shaders/gelu_quick.comp
new file mode 100644
index 00000000..e6e6fcfd
--- /dev/null
+++ b/ggml/src/vulkan-shaders/gelu_quick.comp
@@ -0,0 +1,23 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const float GELU_QUICK_COEF = -1.702f;
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float x = float(data_a[i]);
+    data_d[i] = D_TYPE(x * (1.0f / (1.0f + exp(GELU_QUICK_COEF * x))));
+}
diff --git a/ggml/src/vulkan-shaders/generic_binary_head.comp b/ggml/src/vulkan-shaders/generic_binary_head.comp
index ab45d256..b6beaff1 100644
--- a/ggml/src/vulkan-shaders/generic_binary_head.comp
+++ b/ggml/src/vulkan-shaders/generic_binary_head.comp
@@ -7,7 +7,7 @@ layout (push_constant) uniform parameter
     uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
     uint ne20; uint ne21; uint ne22; uint ne23; uint nb20; uint nb21; uint nb22; uint nb23;
     uint d_offset;
-    float param1; float param2;
+    float param1; float param2; int param3;
 } p;
 
 layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
@@ -16,6 +16,10 @@ layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
 layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 
+uint get_idx() {
+    return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+}
+
 uint src0_idx(uint idx) {
     const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
     const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
diff --git a/ggml/src/vulkan-shaders/generic_unary_head.comp b/ggml/src/vulkan-shaders/generic_unary_head.comp
index de08de7c..eacdefc7 100644
--- a/ggml/src/vulkan-shaders/generic_unary_head.comp
+++ b/ggml/src/vulkan-shaders/generic_unary_head.comp
@@ -14,6 +14,10 @@ layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 
+uint get_idx() {
+    return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+}
+
 uint src0_idx(uint idx) {
     const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
     const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
diff --git a/ggml/src/vulkan-shaders/group_norm.comp b/ggml/src/vulkan-shaders/group_norm.comp
new file mode 100644
index 00000000..5ad9b28d
--- /dev/null
+++ b/ggml/src/vulkan-shaders/group_norm.comp
@@ -0,0 +1,66 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+#define BLOCK_SIZE 512
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+shared float tmp[BLOCK_SIZE];
+
+void main() {
+    const uint group_size = p.KX;
+    const float eps = p.param1;
+
+    const uint tid = gl_LocalInvocationID.x;
+    const uint start = gl_WorkGroupID.x * group_size + tid;
+    const uint end = start + group_size;
+
+    tmp[tid] = 0.0f;
+
+    // Calculate mean
+    [[unroll]] for (uint col = start; col < end; col += BLOCK_SIZE) {
+        tmp[tid] += float(data_a[col]);
+    }
+
+    // tmp up partial tmps and write back result
+    barrier();
+    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+
+    const float mean = tmp[0] / group_size;
+    barrier();
+    tmp[tid] = 0.0f;
+
+    // Calculate variance
+    [[unroll]] for (uint col = start; col < end; col += BLOCK_SIZE) {
+        const float xi = float(data_a[col]) - mean;
+        data_d[col] = D_TYPE(xi);
+        tmp[tid] += xi * xi;
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+
+    const float variance = tmp[0] / group_size;
+    const float scale = inversesqrt(variance + eps);
+
+    [[unroll]] for (uint col = start; col < end; col += BLOCK_SIZE) {
+        data_d[col] *= D_TYPE(scale);
+    }
+}
diff --git a/ggml/src/vulkan-shaders/im2col.comp b/ggml/src/vulkan-shaders/im2col.comp
new file mode 100644
index 00000000..4d48610a
--- /dev/null
+++ b/ggml/src/vulkan-shaders/im2col.comp
@@ -0,0 +1,57 @@
+#version 450
+
+#extension GL_EXT_shader_16bit_storage : require
+
+layout (push_constant) uniform parameter
+{
+    uint batch_offset; uint offset_delta;
+    uint IC;
+    uint IW; uint IH;
+    uint OW; uint OH;
+    uint KW; uint KH;
+    uint pelements;
+    uint CHW;
+    int s0; int s1;
+    int p0; int p1;
+    int d0; int d1;
+} p;
+
+#include "types.comp"
+
+#define BLOCK_SIZE 256
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.x;
+    if (i >= p.pelements) {
+        return;
+    }
+
+    const uint ksize = p.OW * (p.KH > 1 ? p.KW : 1);
+    const uint kx = i / ksize;
+    const uint kd = kx * ksize;
+    const uint ky = (i - kd) / p.OW;
+    const uint ix = i % p.OW;
+
+    const uint oh = gl_GlobalInvocationID.y;
+    const uint batch = gl_GlobalInvocationID.z / p.IC;
+    const uint ic = gl_GlobalInvocationID.z % p.IC;
+
+    const uint iiw = ix * p.s0 + kx * p.d0 - p.p0;
+    const uint iih = oh * p.s1 + ky * p.d1 - p.p1;
+
+    const uint offset_dst =
+        ((batch * p.OH + oh) * p.OW + ix) * p.CHW +
+        (ic * (p.KW * p.KH) + ky * p.KW + kx);
+
+    if (iih < 0 || iih >= p.IH || iiw < 0 || iiw >= p.IW) {
+        data_d[offset_dst] = D_TYPE(0.0f);
+    } else {
+        const uint offset_src = ic * p.offset_delta + batch * p.batch_offset;
+        data_d[offset_dst] = D_TYPE(data_a[offset_src + iih * p.IW + iiw]);
+    }
+}
diff --git a/ggml/src/vulkan-shaders/leaky_relu.comp b/ggml/src/vulkan-shaders/leaky_relu.comp
new file mode 100644
index 00000000..d90a99ae
--- /dev/null
+++ b/ggml/src/vulkan-shaders/leaky_relu.comp
@@ -0,0 +1,22 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float val = float(data_a[i]);
+    data_d[i] = D_TYPE(max(val, 0.0f) + min(val, 0.0f) * p.param1);
+}
diff --git a/ggml/src/vulkan-shaders/mul.comp b/ggml/src/vulkan-shaders/mul.comp
index bbb0aa1d..bfb61c92 100644
--- a/ggml/src/vulkan-shaders/mul.comp
+++ b/ggml/src/vulkan-shaders/mul.comp
@@ -4,9 +4,11 @@
 #include "generic_binary_head.comp"
 
 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
         return;
     }
 
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) * FLOAT_TYPE(data_b[src1_idx(gl_GlobalInvocationID.x)]));
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) * FLOAT_TYPE(data_b[src1_idx(idx)]));
 }
diff --git a/ggml/src/vulkan-shaders/mul_mat_vec.comp b/ggml/src/vulkan-shaders/mul_mat_vec.comp
index 15d2a806..46a6369b 100644
--- a/ggml/src/vulkan-shaders/mul_mat_vec.comp
+++ b/ggml/src/vulkan-shaders/mul_mat_vec.comp
@@ -16,6 +16,13 @@ void main() {
     const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
     const uint tid = gl_LocalInvocationID.x;
 
+    // There are not enough cols to use all threads
+    if (tid >= p.ncols) {
+        return;
+    }
+
+    const uint block_size = min(p.ncols, BLOCK_SIZE);
+
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
 
@@ -23,8 +30,8 @@ void main() {
 
     tmp[tid] = FLOAT_TYPE(0.0f);
 
-    [[unroll]] for (uint i = 0; i < p.ncols/BLOCK_SIZE; i += 2) {
-        const uint col = i*BLOCK_SIZE + 2*tid;
+    [[unroll]] for (uint i = 0; i < p.ncols/block_size; i += 2) {
+        const uint col = i*block_size + 2*tid;
         const uint ib = (row*p.ncols + col)/QUANT_K; // block index
         const uint iqs = (col%QUANT_K)/QUANT_R; // quant index
         const uint iybs = col - col%QUANT_K; // y block start index
@@ -38,7 +45,7 @@ void main() {
 
     // sum up partial sums and write back result
     barrier();
-    [[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
+    [[unroll]] for (uint s = block_size/2; s > 0; s >>= 1) {
         if (tid < s) {
             tmp[tid] += tmp[tid + s];
         }
diff --git a/ggml/src/vulkan-shaders/norm.comp b/ggml/src/vulkan-shaders/norm.comp
index 803dbdcb..6627a50b 100644
--- a/ggml/src/vulkan-shaders/norm.comp
+++ b/ggml/src/vulkan-shaders/norm.comp
@@ -14,7 +14,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 shared vec2 sum[BLOCK_SIZE];
 
 void main() {
-    const uint row = gl_WorkGroupID.x;
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint tid = gl_LocalInvocationID.x;
 
     sum[tid] = vec2(0.0f, 0.0f);
diff --git a/ggml/src/vulkan-shaders/pad.comp b/ggml/src/vulkan-shaders/pad.comp
new file mode 100644
index 00000000..a465cd52
--- /dev/null
+++ b/ggml/src/vulkan-shaders/pad.comp
@@ -0,0 +1,26 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+void main() {
+    const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint i3 = idx / (p.ne12*p.ne11*p.ne10);
+    const uint i3_offset = i3 * p.ne12*p.ne11*p.ne10;
+    const uint i2 = (idx - i3_offset) / (p.ne11*p.ne10);
+    const uint i2_offset = i2*p.ne11*p.ne10;
+    const uint i1 = (idx - i3_offset - i2_offset) / p.ne10;
+    const uint i0 = idx - i3_offset - i2_offset - i1*p.ne10;
+
+    const uint src0_idx = i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0*p.nb00;
+    const uint dst_idx = i3*p.nb13 + i2*p.nb12 + i1*p.nb11 + i0*p.nb10;
+
+    const bool is_src0 = i0 < p.ne00 && i1 < p.ne01 && i2 < p.ne02 && i3 < p.ne03;
+
+    data_d[p.d_offset + dst_idx] = D_TYPE(is_src0 ? data_a[src0_idx] : 0.0f);
+}
diff --git a/ggml/src/vulkan-shaders/relu.comp b/ggml/src/vulkan-shaders/relu.comp
index 7e5baa5b..52a19b62 100644
--- a/ggml/src/vulkan-shaders/relu.comp
+++ b/ggml/src/vulkan-shaders/relu.comp
@@ -11,7 +11,7 @@ layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
 layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 
 void main() {
-    const uint i = gl_GlobalInvocationID.x;
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
 
     if (i >= p.KX) {
         return;
diff --git a/ggml/src/vulkan-shaders/rms_norm.comp b/ggml/src/vulkan-shaders/rms_norm.comp
index cfd08d34..b554400b 100644
--- a/ggml/src/vulkan-shaders/rms_norm.comp
+++ b/ggml/src/vulkan-shaders/rms_norm.comp
@@ -14,7 +14,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 shared FLOAT_TYPE sum[BLOCK_SIZE];
 
 void main() {
-    const uint row = gl_WorkGroupID.x;
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint tid = gl_LocalInvocationID.x;
 
     sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp
diff --git a/ggml/src/vulkan-shaders/scale.comp b/ggml/src/vulkan-shaders/scale.comp
index 510cb723..5cd2f668 100644
--- a/ggml/src/vulkan-shaders/scale.comp
+++ b/ggml/src/vulkan-shaders/scale.comp
@@ -4,9 +4,11 @@
 #include "generic_unary_head.comp"
 
 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
         return;
     }
 
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) * FLOAT_TYPE(p.param1));
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) * FLOAT_TYPE(p.param1));
 }
diff --git a/ggml/src/vulkan-shaders/silu.comp b/ggml/src/vulkan-shaders/silu.comp
index 15920f06..4d36f88e 100644
--- a/ggml/src/vulkan-shaders/silu.comp
+++ b/ggml/src/vulkan-shaders/silu.comp
@@ -11,7 +11,7 @@ layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
 layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 
 void main() {
-    const uint i = gl_GlobalInvocationID.x;
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
 
     if (i >= p.KX) {
         return;
diff --git a/ggml/src/vulkan-shaders/soft_max.comp b/ggml/src/vulkan-shaders/soft_max.comp
index 1b8419c7..0bd51eca 100644
--- a/ggml/src/vulkan-shaders/soft_max.comp
+++ b/ggml/src/vulkan-shaders/soft_max.comp
@@ -28,7 +28,7 @@ shared FLOAT_TYPE vals[BLOCK_SIZE];
 
 void main() {
     const uint tid = gl_LocalInvocationID.x;
-    const uint rowx = gl_WorkGroupID.x;
+    const uint rowx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint rowy = rowx % p.KY;
 
     float slope = 1.0f;
diff --git a/ggml/src/vulkan-shaders/square.comp b/ggml/src/vulkan-shaders/square.comp
index 8dd19333..1fa118c9 100644
--- a/ggml/src/vulkan-shaders/square.comp
+++ b/ggml/src/vulkan-shaders/square.comp
@@ -4,10 +4,12 @@
 #include "generic_unary_head.comp"
 
 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
         return;
     }
 
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]);
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(val * val);
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(val * val);
 }
diff --git a/ggml/src/vulkan-shaders/sum_rows.comp b/ggml/src/vulkan-shaders/sum_rows.comp
index ce2f1e2f..961e5ffa 100644
--- a/ggml/src/vulkan-shaders/sum_rows.comp
+++ b/ggml/src/vulkan-shaders/sum_rows.comp
@@ -14,7 +14,7 @@ layout (constant_id = 0) const uint BLOCK_SIZE = 32;
 shared FLOAT_TYPE tmp[BLOCK_SIZE];
 
 void main() {
-    const uint row = gl_WorkGroupID.x;
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint col = gl_LocalInvocationID.x;
 
     tmp[col] = FLOAT_TYPE(0.0f);
diff --git a/ggml/src/vulkan-shaders/tanh.comp b/ggml/src/vulkan-shaders/tanh.comp
new file mode 100644
index 00000000..74630dc7
--- /dev/null
+++ b/ggml/src/vulkan-shaders/tanh.comp
@@ -0,0 +1,21 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    data_d[i] = D_TYPE(tanh(data_a[i]));
+}
diff --git a/ggml/src/vulkan-shaders/timestep_embedding.comp b/ggml/src/vulkan-shaders/timestep_embedding.comp
new file mode 100644
index 00000000..79e065a9
--- /dev/null
+++ b/ggml/src/vulkan-shaders/timestep_embedding.comp
@@ -0,0 +1,41 @@
+#version 450
+
+#extension GL_EXT_shader_16bit_storage : require
+
+layout (push_constant) uniform parameter
+{
+    uint nb1;
+    uint dim;
+    uint max_period;
+} p;
+
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+#define BLOCK_SIZE 256
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_WorkGroupID.y;
+    const uint j = gl_GlobalInvocationID.x;
+    const uint d_offset = i * p.nb1;
+
+    if (p.dim % 2 != 0 && j == ((p.dim + 1) / 2)) {
+        data_d[d_offset + p.dim] = 0.f;
+    }
+
+    const uint half_dim = p.dim / 2;
+    if (j >= half_dim) {
+        return;
+    }
+
+    const float timestep = float(data_a[i]);
+    const float freq = float(exp(-log(p.max_period) * j / half_dim));
+    const float arg = timestep * freq;
+    data_d[d_offset + j] = D_TYPE(cos(arg));
+    data_d[d_offset + j + half_dim] = D_TYPE(sin(arg));
+}
diff --git a/ggml/src/vulkan-shaders/types.comp b/ggml/src/vulkan-shaders/types.comp
index d24c172c..21dce72f 100644
--- a/ggml/src/vulkan-shaders/types.comp
+++ b/ggml/src/vulkan-shaders/types.comp
@@ -6,7 +6,7 @@
 #define QUANT_K 1
 #define QUANT_R 1
 
-#ifndef LOAD_VEC_A
+#if !defined(LOAD_VEC_A) || LOAD_VEC_A == 1
 #define A_TYPE float
 #elif LOAD_VEC_A == 4
 #define A_TYPE vec4
@@ -19,7 +19,7 @@
 #define QUANT_K 1
 #define QUANT_R 1
 
-#ifndef LOAD_VEC_A
+#if !defined(LOAD_VEC_A) || LOAD_VEC_A == 1
 #define A_TYPE float16_t
 #elif LOAD_VEC_A == 4
 #define A_TYPE f16vec4
diff --git a/ggml/src/vulkan-shaders/upscale.comp b/ggml/src/vulkan-shaders/upscale.comp
new file mode 100644
index 00000000..511a086e
--- /dev/null
+++ b/ggml/src/vulkan-shaders/upscale.comp
@@ -0,0 +1,36 @@
+#version 450
+
+layout (push_constant) uniform parameter
+{
+    uint ne; uint d_offset;
+    uint nb00; uint nb01; uint nb02; uint nb03;
+    uint ne10; uint ne11; uint ne12; uint ne13;
+    float sf0; float sf1; float sf2; float sf3;
+} p;
+
+#include "types.comp"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint i10 = idx % p.ne10;
+    const uint i11 = (idx / p.ne10) % p.ne11;
+    const uint i12 = (idx / (p.ne10 * p.ne11)) % p.ne12;
+    const uint i13 = (idx / (p.ne10 * p.ne11 * p.ne12)) % p.ne13;
+
+    const uint i00 = uint(i10 / p.sf0);
+    const uint i01 = uint(i11 / p.sf1);
+    const uint i02 = uint(i12 / p.sf2);
+    const uint i03 = uint(i13 / p.sf3);
+
+    data_d[p.d_offset + idx] = D_TYPE(data_a[i03 * p.nb03 + i02 * p.nb02 + i01 * p.nb01 + i00 * p.nb00]);
+}
diff --git a/ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
index c5be3754..a792e203 100644
--- a/ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -22,6 +22,7 @@
 #ifdef _WIN32
     #include <windows.h>
     #include <direct.h> // For _mkdir on Windows
+    #include <algorithm> // For std::replace on w64devkit
 #else
     #include <unistd.h>
     #include <sys/wait.h>
@@ -179,11 +180,7 @@ bool string_ends_with(const std::string& str, const std::string& suffix) {
     return std::equal(suffix.rbegin(), suffix.rend(), str.rbegin());
 }
 
-#ifdef _WIN32
-    static const char path_separator = '\\';
-#else
-    static const char path_separator = '/';
-#endif
+static const char path_separator = '/';
 
 std::string join_paths(const std::string& path1, const std::string& path2) {
     return path1 + path_separator + path2;
@@ -198,7 +195,11 @@ void string_to_spv(const std::string& _name, const std::string& in_fname, const
     std::string out_fname = join_paths(output_dir, name + ".spv");
     std::string in_path = join_paths(input_dir, in_fname);
 
-    std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", in_path, "-o", out_fname};
+    #ifdef _WIN32
+        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", "\"" + in_path + "\"", "-o", "\"" + out_fname + "\""};
+    #else
+        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", in_path, "-o",  out_fname};
+    #endif
     for (const auto& define : defines) {
         cmd.push_back("-D" + define.first + "=" + define.second);
     }
@@ -269,9 +270,12 @@ void matmul_shaders(std::vector<std::future<void>>& tasks, bool fp16, bool matmu
 
     for (const auto& tname : type_names) {
         std::string data_a_key = "DATA_A_" + to_uppercase(tname);
+        // For unaligned, load one at a time for f32/f16, or two at a time for quants
+        std::string load_vec_a_unaligned = (tname == "f32" || tname == "f16") ? "1" : "2";
+        // For aligned matmul loads
         std::string load_vec_a = (tname == "f32" || tname == "f16") ? load_vec : "2";
         tasks.push_back(std::async(std::launch::async, [=] {
-            string_to_spv(shader_name + "_" + tname + "_f32", "mul_mm.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16);
+            string_to_spv(shader_name + "_" + tname + "_f32", "mul_mm.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16);
         }));
         tasks.push_back(std::async(std::launch::async, [=] {
             string_to_spv(shader_name + "_" + tname + "_f32_aligned", "mul_mm.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}}), fp16);
@@ -341,6 +345,9 @@ void process_shaders(std::vector<std::future<void>>& tasks) {
         string_to_spv("norm_f32", "norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     }));
     tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("group_norm_f32", "group_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+    }));
+    tasks.push_back(std::async(std::launch::async, [=] {
         string_to_spv("rms_norm_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     }));
 
@@ -357,6 +364,9 @@ void process_shaders(std::vector<std::future<void>>& tasks) {
     tasks.push_back(std::async(std::launch::async, [] {
         string_to_spv("add_f32", "add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
     }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("add_f16_f32_f16", "add.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"FLOAT_TYPE", "float"}});
+    }));
 
     tasks.push_back(std::async(std::launch::async, [] {
         string_to_spv("split_k_reduce", "mul_mat_split_k_reduce.comp", {});
@@ -383,14 +393,41 @@ void process_shaders(std::vector<std::future<void>>& tasks) {
     }));
 
     tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("pad_f32", "pad.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("concat_f32", "concat.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("concat_f16", "concat.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("concat_i32", "concat.comp", {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("upscale_f32", "upscale.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
         string_to_spv("gelu_f32", "gelu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     }));
     tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("gelu_quick_f32", "gelu_quick.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
         string_to_spv("silu_f32", "silu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     }));
     tasks.push_back(std::async(std::launch::async, [] {
         string_to_spv("relu_f32", "relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("leaky_relu_f32", "leaky_relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("tanh_f32", "tanh.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
 
     tasks.push_back(std::async(std::launch::async, [] {
         string_to_spv("diag_mask_inf_f32", "diag_mask_inf.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
@@ -424,6 +461,17 @@ void process_shaders(std::vector<std::future<void>>& tasks) {
     tasks.push_back(std::async(std::launch::async, [=] {
         string_to_spv("sum_rows_f32", "sum_rows.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     }));
+
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("im2col_f32", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+    }));
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("im2col_f32_f16", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}}));
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("timestep_embedding_f32", "timestep_embedding.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+    }));
 }
 
 void write_output_files() {
@@ -435,10 +483,16 @@ void write_output_files() {
 
     for (const auto& pair : shader_fnames) {
         const std::string& name = pair.first;
-        const std::string& path = pair.second;
+        #ifdef _WIN32
+            std::string path = pair.second;
+            std::replace(path.begin(), path.end(), '/', '\\' );
+        #else
+            const std::string& path = pair.second;
+        #endif
+
         FILE* spv = fopen(path.c_str(), "rb");
         if (!spv) {
-            std::cerr << "Error opening SPIR-V file: " << path << "\n";
+            std::cerr << "Error opening SPIR-V file: " << path << " (" << strerror(errno) << ")\n";
             continue;
         }
 
@@ -450,7 +504,7 @@ void write_output_files() {
         size_t read_size = fread(data.data(), 1, size, spv);
         fclose(spv);
         if (read_size != size) {
-            std::cerr << "Error reading SPIR-V file: " << path << "\n";
+            std::cerr << "Error reading SPIR-V file: " << path << " (" << strerror(errno) << ")\n";
             continue;
         }
 
diff --git a/gguf-py/examples/writer.py b/gguf-py/examples/writer.py
index f39eed1a..731873a7 100755
--- a/gguf-py/examples/writer.py
+++ b/gguf-py/examples/writer.py
@@ -15,7 +15,6 @@ def writer_example() -> None:
     # Example usage with a file
     gguf_writer = GGUFWriter("example.gguf", "llama")
 
-    gguf_writer.add_architecture()
     gguf_writer.add_block_count(12)
     gguf_writer.add_uint32("answer", 42)  # Write a 32-bit integer
     gguf_writer.add_float32("answer_in_float", 42.0)  # Write a 32-bit float
diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py
index e343c2ef..1bea66aa 100644
--- a/gguf-py/gguf/constants.py
+++ b/gguf-py/gguf/constants.py
@@ -161,6 +161,7 @@ class Keys:
         SUFFIX_ID            = "tokenizer.ggml.suffix_token_id"
         MIDDLE_ID            = "tokenizer.ggml.middle_token_id"
         EOT_ID               = "tokenizer.ggml.eot_token_id"
+        EOM_ID               = "tokenizer.ggml.eom_token_id"
 
     class Adapter:
         TYPE       = "adapter.type"
@@ -216,6 +217,7 @@ class MODEL_ARCH(IntEnum):
     CHATGLM      = auto()
     BITNET       = auto()
     T5           = auto()
+    T5ENCODER    = auto()
     JAIS         = auto()
 
 
@@ -343,6 +345,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.CHATGLM:        "chatglm",
     MODEL_ARCH.BITNET:         "bitnet",
     MODEL_ARCH.T5:             "t5",
+    MODEL_ARCH.T5ENCODER:      "t5encoder",
     MODEL_ARCH.JAIS:           "jais",
 }
 
@@ -1035,6 +1038,21 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.ENC_FFN_UP,
         MODEL_TENSOR.ENC_OUTPUT_NORM,
     ],
+    MODEL_ARCH.T5ENCODER: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ENC_ATTN_NORM,
+        MODEL_TENSOR.ENC_ATTN_Q,
+        MODEL_TENSOR.ENC_ATTN_K,
+        MODEL_TENSOR.ENC_ATTN_V,
+        MODEL_TENSOR.ENC_ATTN_OUT,
+        MODEL_TENSOR.ENC_ATTN_REL_B,
+        MODEL_TENSOR.ENC_FFN_NORM,
+        MODEL_TENSOR.ENC_FFN_GATE,
+        MODEL_TENSOR.ENC_FFN_DOWN,
+        MODEL_TENSOR.ENC_FFN_UP,
+        MODEL_TENSOR.ENC_OUTPUT_NORM,
+    ],
     MODEL_ARCH.JAIS: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -1145,6 +1163,19 @@ class GGMLQuantizationType(IntEnum):
     F64     = 28
     IQ1_M   = 29
     BF16    = 30
+    Q4_0_4_4 = 31
+    Q4_0_4_8 = 32
+    Q4_0_8_8 = 33
+    IQ1_BN  = 34,
+    IQ2_BN  = 35,
+    Q8_K64  = 36,
+    IQ2_K   = 37,
+    IQ3_K   = 38,
+    IQ4_K   = 39,
+    IQ5_K   = 40,
+    IQ6_K   = 41,
+    IQ2_TN  = 42,
+
 
 
 # TODO: add GGMLFileType from ggml_ftype in ggml.h
@@ -1157,7 +1188,7 @@ class LlamaFileType(IntEnum):
     MOSTLY_F16           = 1   # except 1d tensors
     MOSTLY_Q4_0          = 2   # except 1d tensors
     MOSTLY_Q4_1          = 3   # except 1d tensors
-    MOSTLY_Q4_1_SOME_F16 = 4   # tok_embeddings.weight and output.weight are F16
+    # MOSTLY_Q4_1_SOME_F16 = 4   # tok_embeddings.weight and output.weight are F16
     # MOSTLY_Q4_2        = 5   # support has been removed
     # MOSTLY_Q4_3        = 6   # support has been removed
     MOSTLY_Q8_0          = 7   # except 1d tensors
@@ -1186,6 +1217,18 @@ class LlamaFileType(IntEnum):
     MOSTLY_IQ4_XS        = 30  # except 1d tensors
     MOSTLY_IQ1_M         = 31  # except 1d tensors
     MOSTLY_BF16          = 32  # except 1d tensors
+    MOSTLY_Q4_0_4_4      = 33  # except 1d tensors
+    MOSTLY_Q4_0_4_8      = 34  # except 1d tensors
+    MOSTLY_Q4_0_8_8      = 35  # except 1d tensors
+    MOSTLY_IQ1_BN        = 36, # except 1d tensors
+    MOSTLY_IQ2_BN        = 37, # except 1d tensors
+    MOSTLY_IQ2_K         = 38, # except 1d tensors
+    MOSTLY_IQ3_K         = 39, # except 1d tensors
+    MOSTLY_IQ4_K         = 40, # except 1d tensors
+    MOSTLY_IQ5_K         = 41, # except 1d tensors
+    MOSTLY_IQ6_K         = 42, # except 1d tensors
+    MOSTLY_IQ2_TN        = 43, # except 1d tensors
+
 
     GUESSED              = 1024  # not specified in the model file
 
@@ -1259,6 +1302,9 @@ GGML_QUANT_SIZES: dict[GGMLQuantizationType, tuple[int, int]] = {
     GGMLQuantizationType.F64:     (1, 8),
     GGMLQuantizationType.IQ1_M:   (256, QK_K // 8 + QK_K // 16  + QK_K // 32),
     GGMLQuantizationType.BF16:    (1, 2),
+    GGMLQuantizationType.Q4_0_4_4:(32, 2 + 16),
+    GGMLQuantizationType.Q4_0_4_8:(32, 2 + 16),
+    GGMLQuantizationType.Q4_0_8_8:(32, 2 + 16),
 }
 
 
@@ -1327,3 +1373,4 @@ KEY_TOKENIZER_PRIFIX_ID  = Keys.Tokenizer.PREFIX_ID
 KEY_TOKENIZER_SUFFIX_ID  = Keys.Tokenizer.SUFFIX_ID
 KEY_TOKENIZER_MIDDLE_ID  = Keys.Tokenizer.MIDDLE_ID
 KEY_TOKENIZER_EOT_ID     = Keys.Tokenizer.EOT_ID
+KEY_TOKENIZER_EOM_ID     = Keys.Tokenizer.EOM_ID
diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py
index ba6f53cd..76385a82 100644
--- a/gguf-py/gguf/gguf_writer.py
+++ b/gguf-py/gguf/gguf_writer.py
@@ -312,6 +312,8 @@ class GGUFWriter:
         self.add_key_value(key, val, GGUFValueType.STRING)
 
     def add_array(self, key: str, val: Sequence[Any]) -> None:
+        if len(val) == 0:
+            return
         self.add_key_value(key, val, GGUFValueType.ARRAY)
 
     @staticmethod
@@ -826,6 +828,9 @@ class GGUFWriter:
     def add_eot_token_id(self, id: int) -> None:
         self.add_uint32(Keys.Tokenizer.EOT_ID, id)
 
+    def add_eom_token_id(self, id: int) -> None:
+        self.add_uint32(Keys.Tokenizer.EOM_ID, id)
+
     def _pack(self, fmt: str, value: Any, skip_pack_prefix: bool = False) -> bytes:
         pack_prefix = ''
         if not skip_pack_prefix:
@@ -845,7 +850,14 @@ class GGUFWriter:
             encoded_val = val.encode("utf-8") if isinstance(val, str) else val
             kv_data += self._pack("Q", len(encoded_val))
             kv_data += encoded_val
-        elif vtype == GGUFValueType.ARRAY and isinstance(val, Sequence) and val:
+        elif vtype == GGUFValueType.ARRAY:
+
+            if not isinstance(val, Sequence):
+                raise ValueError("Invalid GGUF metadata array, expecting sequence")
+
+            if len(val) == 0:
+                raise ValueError("Invalid GGUF metadata array. Empty array")
+
             if isinstance(val, bytes):
                 ltype = GGUFValueType.UINT8
             else:
diff --git a/gguf-py/gguf/lazy.py b/gguf-py/gguf/lazy.py
index ac98d9a9..8d4fece2 100644
--- a/gguf-py/gguf/lazy.py
+++ b/gguf-py/gguf/lazy.py
@@ -191,6 +191,8 @@ class LazyBase(ABC, metaclass=LazyMeta):
 class LazyNumpyTensor(LazyBase):
     _tensor_type = np.ndarray
 
+    shape: tuple[int, ...]  # Makes the type checker happy in quants.py
+
     @classmethod
     def meta_with_dtype_and_shape(cls, dtype: DTypeLike, shape: tuple[int, ...]) -> np.ndarray[Any, Any]:
         # The initial idea was to use np.nan as the fill value,
diff --git a/gguf-py/gguf/metadata.py b/gguf-py/gguf/metadata.py
index 15189f71..db318542 100644
--- a/gguf-py/gguf/metadata.py
+++ b/gguf-py/gguf/metadata.py
@@ -174,7 +174,7 @@ class Metadata:
             org_component, model_full_name_component = None, model_id
 
         # Check if we erroneously matched against './' or '../' etc...
-        if org_component is not None and org_component[0] == '.':
+        if org_component is not None and len(org_component) > 0 and org_component[0] == '.':
             org_component = None
 
         name_parts: list[str] = model_full_name_component.split('-')
@@ -284,20 +284,67 @@ class Metadata:
         ########################
         if model_card is not None:
 
-            if "model_name" in model_card and metadata.name is None:
-                # Not part of huggingface model card standard but notice some model creator using it
-                # such as TheBloke in 'TheBloke/Mistral-7B-Instruct-v0.2-GGUF'
-                metadata.name = model_card.get("model_name")
+            def use_model_card_metadata(metadata_key: str, model_card_key: str):
+                if model_card_key in model_card and getattr(metadata, metadata_key, None) is None:
+                    setattr(metadata, metadata_key, model_card.get(model_card_key))
 
-            if "model_creator" in model_card and metadata.author is None:
-                # Not part of huggingface model card standard but notice some model creator using it
-                # such as TheBloke in 'TheBloke/Mistral-7B-Instruct-v0.2-GGUF'
-                metadata.author = model_card.get("model_creator")
+            def use_array_model_card_metadata(metadata_key: str, model_card_key: str):
+                # Note: Will append rather than replace if already exist
+                tags_value = model_card.get(model_card_key, None)
+                if tags_value is None:
+                    return
 
-            if "model_type" in model_card and metadata.basename is None:
-                # Not part of huggingface model card standard but notice some model creator using it
-                # such as TheBloke in 'TheBloke/Mistral-7B-Instruct-v0.2-GGUF'
-                metadata.basename = model_card.get("model_type")
+                current_value = getattr(metadata, metadata_key, None)
+                if current_value is None:
+                    current_value = []
+
+                if isinstance(tags_value, str):
+                    current_value.append(tags_value)
+                elif isinstance(tags_value, list):
+                    current_value.extend(tags_value)
+
+                setattr(metadata, metadata_key, current_value)
+
+            # LLAMA.cpp's direct internal convention
+            # (Definitely not part of hugging face formal/informal standard)
+            #########################################
+            use_model_card_metadata("name", "name")
+            use_model_card_metadata("author", "author")
+            use_model_card_metadata("version", "version")
+            use_model_card_metadata("organization", "organization")
+            use_model_card_metadata("description", "description")
+            use_model_card_metadata("finetune", "finetune")
+            use_model_card_metadata("basename", "basename")
+            use_model_card_metadata("size_label", "size_label")
+            use_model_card_metadata("source_url", "url")
+            use_model_card_metadata("source_doi", "doi")
+            use_model_card_metadata("source_uuid", "uuid")
+            use_model_card_metadata("source_repo_url", "repo_url")
+
+            # LLAMA.cpp's huggingface style convention
+            # (Definitely not part of hugging face formal/informal standard... but with model_ appended to match their style)
+            ###########################################
+            use_model_card_metadata("name", "model_name")
+            use_model_card_metadata("author", "model_author")
+            use_model_card_metadata("version", "model_version")
+            use_model_card_metadata("organization", "model_organization")
+            use_model_card_metadata("description", "model_description")
+            use_model_card_metadata("finetune", "model_finetune")
+            use_model_card_metadata("basename", "model_basename")
+            use_model_card_metadata("size_label", "model_size_label")
+            use_model_card_metadata("source_url", "model_url")
+            use_model_card_metadata("source_doi", "model_doi")
+            use_model_card_metadata("source_uuid", "model_uuid")
+            use_model_card_metadata("source_repo_url", "model_repo_url")
+
+            # Hugging Face Direct Convention
+            #################################
+
+            # Not part of huggingface model card standard but notice some model creator using it
+            # such as TheBloke in 'TheBloke/Mistral-7B-Instruct-v0.2-GGUF'
+            use_model_card_metadata("name", "model_name")
+            use_model_card_metadata("author", "model_creator")
+            use_model_card_metadata("basename", "model_type")
 
             if "base_model" in model_card:
                 # This represents the parent models that this is based on
@@ -329,58 +376,18 @@ class Metadata:
                         base_model["repo_url"] = f"https://huggingface.co/{org_component}/{model_full_name_component}"
                     metadata.base_models.append(base_model)
 
-            if "license" in model_card and metadata.license is None:
-                metadata.license = model_card.get("license")
-
-            if "license_name" in model_card and metadata.license_name is None:
-                metadata.license_name = model_card.get("license_name")
-
-            if "license_link" in model_card and metadata.license_link is None:
-                metadata.license_link = model_card.get("license_link")
-
-            tags_value = model_card.get("tags", None)
-            if tags_value is not None:
-
-                if metadata.tags is None:
-                    metadata.tags = []
-
-                if isinstance(tags_value, str):
-                    metadata.tags.append(tags_value)
-                elif isinstance(tags_value, list):
-                    metadata.tags.extend(tags_value)
-
-            pipeline_tags_value = model_card.get("pipeline_tag", None)
-            if pipeline_tags_value is not None:
-
-                if metadata.tags is None:
-                    metadata.tags = []
-
-                if isinstance(pipeline_tags_value, str):
-                    metadata.tags.append(pipeline_tags_value)
-                elif isinstance(pipeline_tags_value, list):
-                    metadata.tags.extend(pipeline_tags_value)
-
-            language_value = model_card.get("languages", model_card.get("language", None))
-            if language_value is not None:
-
-                if metadata.languages is None:
-                    metadata.languages = []
-
-                if isinstance(language_value, str):
-                    metadata.languages.append(language_value)
-                elif isinstance(language_value, list):
-                    metadata.languages.extend(language_value)
+            use_model_card_metadata("license", "license")
+            use_model_card_metadata("license_name", "license_name")
+            use_model_card_metadata("license_link", "license_link")
 
-            dataset_value = model_card.get("datasets", model_card.get("dataset", None))
-            if dataset_value is not None:
+            use_array_model_card_metadata("tags", "tags")
+            use_array_model_card_metadata("tags", "pipeline_tag")
 
-                if metadata.datasets is None:
-                    metadata.datasets = []
+            use_array_model_card_metadata("languages", "languages")
+            use_array_model_card_metadata("languages", "language")
 
-                if isinstance(dataset_value, str):
-                    metadata.datasets.append(dataset_value)
-                elif isinstance(dataset_value, list):
-                    metadata.datasets.extend(dataset_value)
+            use_array_model_card_metadata("datasets", "datasets")
+            use_array_model_card_metadata("datasets", "dataset")
 
         # Hugging Face Parameter Heuristics
         ####################################
diff --git a/gguf-py/gguf/quants.py b/gguf-py/gguf/quants.py
index 16e0a9aa..ff589b85 100644
--- a/gguf-py/gguf/quants.py
+++ b/gguf-py/gguf/quants.py
@@ -1,42 +1,32 @@
 from __future__ import annotations
-from typing import Callable, Sequence
+from abc import ABC, abstractmethod
+from typing import Any, Callable, Sequence
+from math import log2, ceil
 
 from numpy.typing import DTypeLike
 
-from .constants import GGML_QUANT_SIZES, GGMLQuantizationType
+from .constants import GGML_QUANT_SIZES, GGMLQuantizationType, QK_K
 from .lazy import LazyNumpyTensor
 
 import numpy as np
 
 
-def quant_shape_to_byte_shape(shape: Sequence[int], quant_type: GGMLQuantizationType):
+def quant_shape_to_byte_shape(shape: Sequence[int], quant_type: GGMLQuantizationType) -> tuple[int, ...]:
     block_size, type_size = GGML_QUANT_SIZES[quant_type]
     if shape[-1] % block_size != 0:
         raise ValueError(f"Quantized tensor row size ({shape[-1]}) is not a multiple of {quant_type.name} block size ({block_size})")
     return (*shape[:-1], shape[-1] // block_size * type_size)
 
 
-def quant_shape_from_byte_shape(shape: Sequence[int], quant_type: GGMLQuantizationType):
+def quant_shape_from_byte_shape(shape: Sequence[int], quant_type: GGMLQuantizationType) -> tuple[int, ...]:
     block_size, type_size = GGML_QUANT_SIZES[quant_type]
     if shape[-1] % type_size != 0:
         raise ValueError(f"Quantized tensor bytes per row ({shape[-1]}) is not a multiple of {quant_type.name} type size ({type_size})")
     return (*shape[:-1], shape[-1] // type_size * block_size)
 
 
-# same as ggml_compute_fp32_to_bf16 in ggml-impl.h
-def __compute_fp32_to_bf16(n: np.ndarray) -> np.ndarray:
-    n = n.astype(np.float32, copy=False).view(np.int32)
-    # force nan to quiet
-    n = np.where((n & 0x7fffffff) > 0x7f800000, (n & 0xffff0000) | (64 << 16), n)
-    # flush subnormals to zero
-    n = np.where((n & 0x7f800000) == 0, n & 0x80000000, n)
-    # round to nearest even
-    n = (n + (0x7fff + ((n >> 16) & 1))) >> 16
-    return n.astype(np.int16)
-
-
 # This is faster than np.vectorize and np.apply_along_axis because it works on more than one row at a time
-def __apply_over_grouped_rows(func: Callable[[np.ndarray], np.ndarray], arr: np.ndarray, otype: DTypeLike, oshape: tuple[int, ...]) -> np.ndarray:
+def _apply_over_grouped_rows(func: Callable[[np.ndarray], np.ndarray], arr: np.ndarray, otype: DTypeLike, oshape: tuple[int, ...]) -> np.ndarray:
     rows = arr.reshape((-1, arr.shape[-1]))
     osize = 1
     for dim in oshape:
@@ -48,76 +38,1151 @@ def __apply_over_grouped_rows(func: Callable[[np.ndarray], np.ndarray], arr: np.
     return out.reshape(oshape)
 
 
-def __quantize_bf16_array(n: np.ndarray) -> np.ndarray:
-    return __apply_over_grouped_rows(__compute_fp32_to_bf16, arr=n, otype=np.int16, oshape=n.shape)
+# round away from zero
+# ref: https://stackoverflow.com/a/59143326/22827863
+def np_roundf(n: np.ndarray) -> np.ndarray:
+    a = abs(n)
+    floored = np.floor(a)
+    b = floored + np.floor(2 * (a - floored))
+    return np.sign(n) * b
+
+
+class QuantError(Exception): ...
 
 
-__quantize_bf16_lazy = LazyNumpyTensor._wrap_fn(__quantize_bf16_array, meta_noop=np.int16)
+_type_traits: dict[GGMLQuantizationType, type[__Quant]] = {}
 
 
-def quantize_bf16(n: np.ndarray):
-    if type(n) is LazyNumpyTensor:
-        return __quantize_bf16_lazy(n)
+def quantize(data: np.ndarray, qtype: GGMLQuantizationType) -> np.ndarray:
+    if qtype == GGMLQuantizationType.F32:
+        return data.astype(np.float32, copy=False)
+    elif qtype == GGMLQuantizationType.F16:
+        return data.astype(np.float16, copy=False)
+    elif (q := _type_traits.get(qtype)) is not None:
+        return q.quantize(data)
     else:
-        return __quantize_bf16_array(n)
+        raise NotImplementedError(f"Quantization for {qtype.name} is not yet implemented")
 
 
-__q8_block_size, __q8_type_size = GGML_QUANT_SIZES[GGMLQuantizationType.Q8_0]
+def dequantize(data: np.ndarray, qtype: GGMLQuantizationType) -> np.ndarray:
+    if qtype == GGMLQuantizationType.F32:
+        return data.view(np.float32)
+    elif qtype == GGMLQuantizationType.F16:
+        return data.view(np.float16).astype(np.float32)
+    elif (q := _type_traits.get(qtype)) is not None:
+        return q.dequantize(data)
+    else:
+        raise NotImplementedError(f"Dequantization for {qtype.name} is not yet implemented")
 
 
-def can_quantize_to_q8_0(n: np.ndarray) -> bool:
-    return n.shape[-1] % __q8_block_size == 0
+class __Quant(ABC):
+    qtype: GGMLQuantizationType
+    block_size: int
+    type_size: int
 
+    grid: np.ndarray[Any, np.dtype[np.float32]] | None = None
+    grid_shape: tuple[int, int] = (0, 0)
+    grid_map: tuple[int | float, ...] = ()
+    grid_hex: bytes | None = None
 
-# round away from zero
-# ref: https://stackoverflow.com/a/59143326/22827863
-def np_roundf(n: np.ndarray) -> np.ndarray:
-    a = abs(n)
-    floored = np.floor(a)
-    b = floored + np.floor(2 * (a - floored))
-    return np.sign(n) * b
+    def __init__(self):
+        return TypeError("Quant conversion classes can't have instances")
 
+    def __init_subclass__(cls, qtype: GGMLQuantizationType) -> None:
+        cls.qtype = qtype
+        cls.block_size, cls.type_size = GGML_QUANT_SIZES[qtype]
+        cls.__quantize_lazy = LazyNumpyTensor._wrap_fn(
+            cls.__quantize_array,
+            meta_noop=(np.uint8, cls.__shape_to_bytes)
+        )
+        cls.__dequantize_lazy = LazyNumpyTensor._wrap_fn(
+            cls.__dequantize_array,
+            meta_noop=(np.float32, cls.__shape_from_bytes)
+        )
+        assert qtype not in _type_traits
+        _type_traits[qtype] = cls
 
-def __quantize_q8_0_shape_change(s: tuple[int, ...]) -> tuple[int, ...]:
-    return (*s[:-1], s[-1] // __q8_block_size * __q8_type_size)
+    @classmethod
+    def init_grid(cls):
+        if cls.grid is not None or cls.grid_hex is None:
+            return
 
+        bits_per_elem = ceil(log2(len(cls.grid_map)))
+        assert bits_per_elem != 0, cls.qtype.name
+        elems_per_byte = 8 // bits_per_elem
 
-# Implementation of Q8_0 with bit-exact same results as reference implementation in ggml-quants.c
-def __quantize_q8_0_rows(n: np.ndarray) -> np.ndarray:
-    shape = n.shape
-    assert shape[-1] % __q8_block_size == 0
+        grid = np.frombuffer(cls.grid_hex, dtype=np.uint8)
+        # decode hexadecimal chars from grid
+        grid = grid.reshape((-1, 2))
+        grid = (np.where(grid > 0x40, grid + 9, grid) & 0x0F) << np.array([4, 0], dtype=np.uint8).reshape((1, 2))
+        grid = grid[..., 0] | grid[..., 1]
+        # unpack the grid values
+        grid = grid.reshape((-1, 1)) >> np.array([i for i in range(0, 8, 8 // elems_per_byte)], dtype=np.uint8).reshape((1, elems_per_byte))
+        grid = (grid & ((1 << bits_per_elem) - 1)).reshape((-1, 1))
+        grid_map = np.array(cls.grid_map, dtype=np.float32).reshape((1, -1))
+        grid = np.take_along_axis(grid_map, grid, axis=-1)
+        cls.grid = grid.reshape((1, 1, *cls.grid_shape))
 
-    n_blocks = n.size // __q8_block_size
+    @classmethod
+    @abstractmethod
+    def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        raise NotImplementedError
 
-    blocks = n.reshape((n_blocks, __q8_block_size)).astype(np.float32, copy=False)
+    @classmethod
+    @abstractmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        raise NotImplementedError
 
-    d = abs(blocks).max(axis=1, keepdims=True) / 127
-    with np.errstate(divide="ignore"):
-        id = np.where(d == 0, 0, 1 / d)
-    qs = np_roundf(blocks * id)
+    @classmethod
+    def quantize_rows(cls, rows: np.ndarray) -> np.ndarray:
+        rows = rows.astype(np.float32, copy=False)
+        shape = rows.shape
+        n_blocks = rows.size // cls.block_size
+        blocks = rows.reshape((n_blocks, cls.block_size))
+        blocks = cls.quantize_blocks(blocks)
+        assert blocks.dtype == np.uint8
+        assert blocks.shape[-1] == cls.type_size
+        return blocks.reshape(cls.__shape_to_bytes(shape))
 
-    # (n_blocks, 2)
-    d = d.astype(np.float16).view(np.uint8)
-    # (n_blocks, block_size)
-    qs = qs.astype(np.int8).view(np.uint8)
+    @classmethod
+    def dequantize_rows(cls, rows: np.ndarray) -> np.ndarray:
+        rows = rows.view(np.uint8)
+        shape = rows.shape
+        n_blocks = rows.size // cls.type_size
+        blocks = rows.reshape((n_blocks, cls.type_size))
+        blocks = cls.dequantize_blocks(blocks)
+        assert blocks.dtype == np.float32
+        assert blocks.shape[-1] == cls.block_size
+        return blocks.reshape(cls.__shape_from_bytes(shape))
 
-    assert d.shape[1] + qs.shape[1] == __q8_type_size
+    @classmethod
+    def __shape_to_bytes(cls, shape: Sequence[int]):
+        return quant_shape_to_byte_shape(shape, cls.qtype)
 
-    return np.concatenate([d, qs], axis=1).reshape(__quantize_q8_0_shape_change(shape))
+    @classmethod
+    def __shape_from_bytes(cls, shape: Sequence[int]):
+        return quant_shape_from_byte_shape(shape, cls.qtype)
 
+    @classmethod
+    def __quantize_array(cls, array: np.ndarray) -> np.ndarray:
+        return _apply_over_grouped_rows(cls.quantize_rows, arr=array, otype=np.uint8, oshape=cls.__shape_to_bytes(array.shape))
 
-def __quantize_q8_0_array(n: np.ndarray) -> np.ndarray:
-    return __apply_over_grouped_rows(__quantize_q8_0_rows, arr=n, otype=np.uint8, oshape=__quantize_q8_0_shape_change(n.shape))
+    @classmethod
+    def __dequantize_array(cls, array: np.ndarray) -> np.ndarray:
+        cls.init_grid()
+        return _apply_over_grouped_rows(cls.dequantize_rows, arr=array, otype=np.float32, oshape=cls.__shape_from_bytes(array.shape))
 
+    @classmethod
+    def __quantize_lazy(cls, lazy_tensor: LazyNumpyTensor, /) -> Any:
+        pass
 
-__quantize_q8_0_lazy = LazyNumpyTensor._wrap_fn(
-    __quantize_q8_0_array,
-    meta_noop=(np.uint8, __quantize_q8_0_shape_change),
-)
+    @classmethod
+    def __dequantize_lazy(cls, lazy_tensor: LazyNumpyTensor, /) -> Any:
+        pass
 
+    @classmethod
+    def can_quantize(cls, tensor: np.ndarray | LazyNumpyTensor) -> bool:
+        return tensor.shape[-1] % cls.block_size == 0
 
-def quantize_q8_0(data: np.ndarray):
-    if type(data) is LazyNumpyTensor:
-        return __quantize_q8_0_lazy(data)
-    else:
-        return __quantize_q8_0_array(data)
+    @classmethod
+    def quantize(cls, tensor: np.ndarray | LazyNumpyTensor) -> np.ndarray:
+        if not cls.can_quantize(tensor):
+            raise QuantError(f"Can't quantize tensor with shape {tensor.shape} to {cls.qtype.name}")
+        if isinstance(tensor, LazyNumpyTensor):
+            return cls.__quantize_lazy(tensor)
+        else:
+            return cls.__quantize_array(tensor)
+
+    @classmethod
+    def dequantize(cls, tensor: np.ndarray | LazyNumpyTensor) -> np.ndarray:
+        if isinstance(tensor, LazyNumpyTensor):
+            return cls.__dequantize_lazy(tensor)
+        else:
+            return cls.__dequantize_array(tensor)
+
+
+class BF16(__Quant, qtype=GGMLQuantizationType.BF16):
+    @classmethod
+    # same as ggml_compute_fp32_to_bf16 in ggml-impl.h
+    def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n = blocks.view(np.uint32)
+        # force nan to quiet
+        n = np.where((n & 0x7fffffff) > 0x7f800000, (n & np.uint32(0xffff0000)) | np.uint32(64 << 16), n)
+        # round to nearest even
+        n = (np.uint64(n) + (0x7fff + ((n >> 16) & 1))) >> 16
+        return n.astype(np.uint16).view(np.uint8)
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        return (blocks.view(np.int16).astype(np.int32) << 16).view(np.float32)
+
+
+class Q4_0(__Quant, qtype=GGMLQuantizationType.Q4_0):
+    @classmethod
+    def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        imax = abs(blocks).argmax(axis=-1, keepdims=True)
+        max = np.take_along_axis(blocks, imax, axis=-1)
+
+        d = max / -8
+        with np.errstate(divide="ignore"):
+            id = np.where(d == 0, 0, 1 / d)
+        # FIXME: Q4_0's reference rounding is cursed and depends on FMA
+        qs = np.trunc((np.float64(blocks) * np.float64(id)) + np.float64(8.5), dtype=np.float32).astype(np.uint8).clip(0, 15)
+
+        qs = qs.reshape((n_blocks, 2, cls.block_size // 2))
+        qs = qs[..., 0, :] | (qs[..., 1, :] << np.uint8(4))
+
+        d = d.astype(np.float16).view(np.uint8)
+
+        return np.concatenate([d, qs], axis=-1)
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, qs = np.hsplit(blocks, [2])
+
+        d = d.view(np.float16).astype(np.float32)
+
+        qs = qs.reshape((n_blocks, -1, 1, cls.block_size // 2)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
+        qs = (qs & np.uint8(0x0F)).reshape((n_blocks, -1)).astype(np.int8) - np.int8(8)
+
+        return (d * qs.astype(np.float32))
+
+
+class Q4_1(__Quant, qtype=GGMLQuantizationType.Q4_1):
+    @classmethod
+    def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        max = blocks.max(axis=-1, keepdims=True)
+        min = blocks.min(axis=-1, keepdims=True)
+
+        d = (max - min) / 15
+        with np.errstate(divide="ignore"):
+            id = np.where(d == 0, 0, 1 / d)
+        qs = np.trunc((blocks - min) * id + np.float32(0.5), dtype=np.float32).astype(np.uint8).clip(0, 15)
+
+        qs = qs.reshape((n_blocks, 2, cls.block_size // 2))
+        qs = qs[..., 0, :] | (qs[..., 1, :] << np.uint8(4))
+
+        d = d.astype(np.float16).view(np.uint8)
+        m = min.astype(np.float16).view(np.uint8)
+
+        return np.concatenate([d, m, qs], axis=-1)
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, rest = np.hsplit(blocks, [2])
+        m, qs = np.hsplit(rest, [2])
+
+        d = d.view(np.float16).astype(np.float32)
+        m = m.view(np.float16).astype(np.float32)
+
+        qs = qs.reshape((n_blocks, -1, 1, cls.block_size // 2)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
+        qs = (qs & np.uint8(0x0F)).reshape((n_blocks, -1)).astype(np.float32)
+
+        return (d * qs) + m
+
+
+class Q5_0(__Quant, qtype=GGMLQuantizationType.Q5_0):
+    @classmethod
+    def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        imax = abs(blocks).argmax(axis=-1, keepdims=True)
+        max = np.take_along_axis(blocks, imax, axis=-1)
+
+        d = max / -16
+        with np.errstate(divide="ignore"):
+            id = np.where(d == 0, 0, 1 / d)
+        # FIXME: Q5_0's reference rounding is cursed and depends on FMA
+        q = np.trunc((np.float64(blocks) * np.float64(id)) + np.float64(16.5), dtype=np.float32).astype(np.uint8).clip(0, 31)
+
+        qs = q.reshape((n_blocks, 2, cls.block_size // 2))
+        qs = (qs[..., 0, :] & np.uint8(0x0F)) | (qs[..., 1, :] << np.uint8(4))
+
+        qh = np.packbits(q.reshape((n_blocks, 1, 32)) >> np.uint8(4), axis=-1, bitorder="little").reshape(n_blocks, 4)
+
+        d = d.astype(np.float16).view(np.uint8)
+
+        return np.concatenate([d, qh, qs], axis=-1)
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, rest = np.hsplit(blocks, [2])
+        qh, qs = np.hsplit(rest, [4])
+
+        d = d.view(np.float16).astype(np.float32)
+        qh = qh.view(np.uint32)
+
+        qh = qh.reshape((n_blocks, 1)) >> np.array([i for i in range(32)], dtype=np.uint32).reshape((1, 32))
+        ql = qs.reshape((n_blocks, -1, 1, cls.block_size // 2)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
+        qh = (qh & np.uint32(0x01)).astype(np.uint8)
+        ql = (ql & np.uint8(0x0F)).reshape((n_blocks, -1))
+
+        qs = (ql | (qh << np.uint8(4))).astype(np.int8) - np.int8(16)
+
+        return (d * qs.astype(np.float32))
+
+
+class Q5_1(__Quant, qtype=GGMLQuantizationType.Q5_1):
+    @classmethod
+    def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        max = blocks.max(axis=-1, keepdims=True)
+        min = blocks.min(axis=-1, keepdims=True)
+
+        d = (max - min) / 31
+        with np.errstate(divide="ignore"):
+            id = np.where(d == 0, 0, 1 / d)
+        q = np.trunc((blocks - min) * id + np.float32(0.5), dtype=np.float32).astype(np.uint8).clip(0, 31)
+
+        qs = q.reshape((n_blocks, 2, cls.block_size // 2))
+        qs = (qs[..., 0, :] & np.uint8(0x0F)) | (qs[..., 1, :] << np.uint8(4))
+
+        qh = np.packbits(q.reshape((n_blocks, 1, 32)) >> np.uint8(4), axis=-1, bitorder="little").reshape(n_blocks, 4)
+
+        d = d.astype(np.float16).view(np.uint8)
+        m = min.astype(np.float16).view(np.uint8)
+
+        return np.concatenate([d, m, qh, qs], axis=-1)
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, rest = np.hsplit(blocks, [2])
+        m, rest = np.hsplit(rest, [2])
+        qh, qs = np.hsplit(rest, [4])
+
+        d = d.view(np.float16).astype(np.float32)
+        m = m.view(np.float16).astype(np.float32)
+        qh = qh.view(np.uint32)
+
+        qh = qh.reshape((n_blocks, 1)) >> np.array([i for i in range(32)], dtype=np.uint32).reshape((1, 32))
+        ql = qs.reshape((n_blocks, -1, 1, cls.block_size // 2)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
+        qh = (qh & np.uint32(0x01)).astype(np.uint8)
+        ql = (ql & np.uint8(0x0F)).reshape((n_blocks, -1))
+
+        qs = (ql | (qh << np.uint8(4))).astype(np.float32)
+
+        return (d * qs) + m
+
+
+class Q8_0(__Quant, qtype=GGMLQuantizationType.Q8_0):
+    @classmethod
+    # Implementation of Q8_0 with bit-exact same results as reference implementation in ggml-quants.c
+    def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+
+        d = abs(blocks).max(axis=1, keepdims=True) / 127
+        with np.errstate(divide="ignore"):
+            id = np.where(d == 0, 0, 1 / d)
+        qs = np_roundf(blocks * id)
+
+        # (n_blocks, 2)
+        d = d.astype(np.float16).view(np.uint8)
+        # (n_blocks, block_size)
+        qs = qs.astype(np.int8).view(np.uint8)
+
+        return np.concatenate([d, qs], axis=1)
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        d, x = np.split(blocks, [2], axis=1)
+        d = d.view(np.float16).astype(np.float32)
+        x = x.view(np.int8).astype(np.float32)
+
+        return (x * d)
+
+
+class Q2_K(__Quant, qtype=GGMLQuantizationType.Q2_K):
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        scales, rest = np.hsplit(blocks, [QK_K // 16])
+        qs, rest = np.hsplit(rest, [QK_K // 4])
+        d, dmin = np.hsplit(rest, [2])
+
+        d = d.view(np.float16).astype(np.float32)
+        dmin = dmin.view(np.float16).astype(np.float32)
+
+        # (n_blocks, 16, 1)
+        dl = (d * (scales & 0xF).astype(np.float32)).reshape((n_blocks, QK_K // 16, 1))
+        ml = (dmin * (scales >> 4).astype(np.float32)).reshape((n_blocks, QK_K // 16, 1))
+
+        shift = np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4, 1))
+
+        qs = (qs.reshape((n_blocks, -1, 1, 32)) >> shift) & np.uint8(3)
+
+        qs = qs.reshape((n_blocks, QK_K // 16, 16)).astype(np.float32)
+
+        qs = dl * qs - ml
+
+        return qs.reshape((n_blocks, -1))
+
+
+class Q3_K(__Quant, qtype=GGMLQuantizationType.Q3_K):
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        hmask, rest = np.hsplit(blocks, [QK_K // 8])
+        qs, rest = np.hsplit(rest, [QK_K // 4])
+        scales, d = np.hsplit(rest, [12])
+
+        d = d.view(np.float16).astype(np.float32)
+
+        # The scales are packed at 6-bit each in this pattern:
+        #  0: IIIIAAAA
+        #  1: JJJJBBBB
+        #  2: KKKKCCCC
+        #  3: LLLLDDDD
+        #  4: MMMMEEEE
+        #  5: NNNNFFFF
+        #  6: OOOOGGGG
+        #  7: PPPPHHHH
+        #  8: MMIIEEAA
+        #  9: NNJJFFBB
+        # 10: OOKKGGCC
+        # 11: PPLLHHDD
+        lscales, hscales = np.hsplit(scales, [8])
+        lscales = lscales.reshape((n_blocks, 1, 8)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 2, 1))
+        lscales = lscales.reshape((n_blocks, 16))
+        hscales = hscales.reshape((n_blocks, 1, 4)) >> np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 4, 1))
+        hscales = hscales.reshape((n_blocks, 16))
+        scales = (lscales & np.uint8(0x0F)) | ((hscales & np.uint8(0x03)) << np.uint8(4))
+        scales = (scales.astype(np.int8) - np.int8(32)).astype(np.float32)
+
+        dl = (d * scales).reshape((n_blocks, 16, 1))
+
+        ql = qs.reshape((n_blocks, -1, 1, 32)) >> np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4, 1))
+        qh = hmask.reshape(n_blocks, -1, 1, 32) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 8, 1))
+        ql = ql.reshape((n_blocks, 16, QK_K // 16)) & np.uint8(3)
+        qh = (qh.reshape((n_blocks, 16, QK_K // 16)) & np.uint8(1))
+        qh = qh ^ np.uint8(1)  # strangely, the offset is zero when the bitmask is 1
+        q = (ql.astype(np.int8) - (qh << np.uint8(2)).astype(np.int8)).astype(np.float32)
+
+        return (dl * q).reshape((n_blocks, QK_K))
+
+
+class Q4_K(__Quant, qtype=GGMLQuantizationType.Q4_K):
+    K_SCALE_SIZE = 12
+
+    @staticmethod
+    def get_scale_min(scales: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
+        n_blocks = scales.shape[0]
+        scales = scales.view(np.uint8)
+        ### Unpacking the following: ###
+        #  0 EEAAAAAA
+        #  1 FFBBBBBB
+        #  2 GGCCCCCC
+        #  3 HHDDDDDD
+        #  4 eeaaaaaa
+        #  5 ffbbbbbb
+        #  6 ggcccccc
+        #  7 hhdddddd
+        #  8 eeeeEEEE
+        #  9 ffffFFFF
+        # 10 ggggGGGG
+        # 11 hhhhHHHH
+        scales = scales.reshape((n_blocks, 3, 4))
+        d, m, m_d = np.split(scales, 3, axis=-2)
+
+        sc = np.concatenate([d & 0x3F, (m_d & 0x0F) | ((d >> 2) & 0x30)], axis=-1)
+        min = np.concatenate([m & 0x3F, (m_d >> 4) | ((m >> 2) & 0x30)], axis=-1)
+
+        return (sc.reshape((n_blocks, 8)), min.reshape((n_blocks, 8)))
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, rest = np.hsplit(blocks, [2])
+        dmin, rest = np.hsplit(rest, [2])
+        scales, qs = np.hsplit(rest, [cls.K_SCALE_SIZE])
+
+        d = d.view(np.float16).astype(np.float32)
+        dmin = dmin.view(np.float16).astype(np.float32)
+
+        sc, m = Q4_K.get_scale_min(scales)
+
+        d = (d * sc.astype(np.float32)).reshape((n_blocks, -1, 1))
+        dm = (dmin * m.astype(np.float32)).reshape((n_blocks, -1, 1))
+
+        qs = qs.reshape((n_blocks, -1, 1, 32)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
+        qs = (qs & np.uint8(0x0F)).reshape((n_blocks, -1, 32)).astype(np.float32)
+
+        return (d * qs - dm).reshape((n_blocks, QK_K))
+
+
+class Q5_K(__Quant, qtype=GGMLQuantizationType.Q5_K):
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, rest = np.hsplit(blocks, [2])
+        dmin, rest = np.hsplit(rest, [2])
+        scales, rest = np.hsplit(rest, [Q4_K.K_SCALE_SIZE])
+        qh, qs = np.hsplit(rest, [QK_K // 8])
+
+        d = d.view(np.float16).astype(np.float32)
+        dmin = dmin.view(np.float16).astype(np.float32)
+
+        sc, m = Q4_K.get_scale_min(scales)
+
+        d = (d * sc.astype(np.float32)).reshape((n_blocks, -1, 1))
+        dm = (dmin * m.astype(np.float32)).reshape((n_blocks, -1, 1))
+
+        ql = qs.reshape((n_blocks, -1, 1, 32)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
+        qh = qh.reshape((n_blocks, -1, 1, 32)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 8, 1))
+        ql = (ql & np.uint8(0x0F)).reshape((n_blocks, -1, 32))
+        qh = (qh & np.uint8(0x01)).reshape((n_blocks, -1, 32))
+        q = (ql | (qh << np.uint8(4))).astype(np.float32)
+
+        return (d * q - dm).reshape((n_blocks, QK_K))
+
+
+class Q6_K(__Quant, qtype=GGMLQuantizationType.Q6_K):
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        ql, rest = np.hsplit(blocks, [QK_K // 2])
+        qh, rest = np.hsplit(rest, [QK_K // 4])
+        scales, d = np.hsplit(rest, [QK_K // 16])
+
+        scales = scales.view(np.int8).astype(np.float32)
+        d = d.view(np.float16).astype(np.float32)
+        d = (d * scales).reshape((n_blocks, QK_K // 16, 1))
+
+        ql = ql.reshape((n_blocks, -1, 1, 64)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
+        ql = (ql & np.uint8(0x0F)).reshape((n_blocks, -1, 32))
+        qh = qh.reshape((n_blocks, -1, 1, 32)) >> np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4, 1))
+        qh = (qh & np.uint8(0x03)).reshape((n_blocks, -1, 32))
+        q = (ql | (qh << np.uint8(4))).astype(np.int8) - np.int8(32)
+        q = q.reshape((n_blocks, QK_K // 16, -1)).astype(np.float32)
+
+        return (d * q).reshape((n_blocks, QK_K))
+
+
+class IQ2_XXS(__Quant, qtype=GGMLQuantizationType.IQ2_XXS):
+    ksigns: bytes = (
+        b"\x00\x81\x82\x03\x84\x05\x06\x87\x88\x09\x0a\x8b\x0c\x8d\x8e\x0f"
+        b"\x90\x11\x12\x93\x14\x95\x96\x17\x18\x99\x9a\x1b\x9c\x1d\x1e\x9f"
+        b"\xa0\x21\x22\xa3\x24\xa5\xa6\x27\x28\xa9\xaa\x2b\xac\x2d\x2e\xaf"
+        b"\x30\xb1\xb2\x33\xb4\x35\x36\xb7\xb8\x39\x3a\xbb\x3c\xbd\xbe\x3f"
+        b"\xc0\x41\x42\xc3\x44\xc5\xc6\x47\x48\xc9\xca\x4b\xcc\x4d\x4e\xcf"
+        b"\x50\xd1\xd2\x53\xd4\x55\x56\xd7\xd8\x59\x5a\xdb\x5c\xdd\xde\x5f"
+        b"\x60\xe1\xe2\x63\xe4\x65\x66\xe7\xe8\x69\x6a\xeb\x6c\xed\xee\x6f"
+        b"\xf0\x71\x72\xf3\x74\xf5\xf6\x77\x78\xf9\xfa\x7b\xfc\x7d\x7e\xff"
+    )
+
+    # iq2xxs_grid, but with each byte of the original packed in 2 bits,
+    # by mapping 0x08 to 0, 0x19 to 1, and 0x2b to 2.
+    grid_shape = (256, 8)
+    grid_map = (0x08, 0x19, 0x2b)
+    grid_hex = (
+        b"00000200050008000a00110014002000220028002a0041004400500058006100"
+        b"6400800082008a00a20001010401100115014001840198010002020222028202"
+        b"010404041004210424044004420448046004810484049004a404000502050805"
+        b"200546056905800591050906100640068406a406000805080808140828084108"
+        b"440850085208880804094009020a140a01100410101021104010601084109010"
+        b"951000110811201150115a118011241245120014081420142514491480141815"
+        b"6215001616160118041810184018811800190519a019511a002002200a204420"
+        b"6120802082202921482100220222012404241024402456240025412564259026"
+        b"082820289428442a014004401040184021402440404048405640604081408440"
+        b"9040004120416141804185410142104248425642684200440844204480449944"
+        b"124524450046014804481048404845480049584961498249454a904a00500850"
+        b"1150195020508050885004514251a4519152905492540a550156545600581158"
+        b"195864584059085a046010604060686000615561186260620064056410651265"
+        b"84654268008002800a8041808280048118814081118201840484108415844084"
+        b"608400854685948509864086608602880489118a0490109024904090a1901691"
+        b"8091459200942294449451958198209902a050a085a009a100a218a450a804a9"
+    )
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, qs = np.hsplit(blocks, [2])
+
+        d = d.view(np.float16).astype(np.float32)
+
+        qs = qs.view(np.uint32).reshape(n_blocks, -1, 2)
+
+        db = d * (np.float32(0.5) + (qs[..., 1] >> 28).astype(np.float32)) * np.float32(0.25)
+        db = db.reshape((n_blocks, -1, 1, 1))
+
+        # get the sign indices and unpack the bits
+        signs = qs[..., 1].reshape((n_blocks, -1, 1)) >> np.array([0, 7, 14, 21], dtype=np.uint32).reshape((1, 1, 4))
+        ksigns = np.frombuffer(cls.ksigns, dtype=np.uint8).reshape((1, 1, 1, 128))
+        signs = (signs & np.uint32(0x7F)).reshape((n_blocks, -1, 4, 1))
+        signs = np.take_along_axis(ksigns, signs, axis=-1)
+        signs = signs.reshape((n_blocks, -1, 4, 1)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 1, 8))
+        signs = signs & np.uint8(0x01)
+        signs = np.where(signs == 0, np.float32(1), np.float32(-1))
+        signs = signs.reshape((n_blocks, -1, 4, 8))
+
+        assert cls.grid is not None
+        grid = np.take_along_axis(cls.grid, qs[..., 0].copy().view(np.uint8).reshape((n_blocks, -1, 1, 1)), axis=-2)
+        grid = grid.reshape((n_blocks, -1, 4, 8))
+
+        return (db * grid * signs).reshape((n_blocks, -1))
+
+
+class IQ2_XS(__Quant, qtype=GGMLQuantizationType.IQ2_XS):
+    # iq2xs_grid, but with each byte of the original packed in 2 bits,
+    # by mapping 0x08 to 0, 0x19 to 1, and 0x2b to 2.
+    grid_shape = (512, 8)
+    grid_map = (0x08, 0x19, 0x2b)
+    grid_hex = (
+        b"00000200050008000a0011001400160019002000220025002800410044004600"
+        b"49005000520055005800610064008000820085008800910094009900a0000101"
+        b"04010601090110011201150118011a0121012401400142014501480151015401"
+        b"6001680181018401900100020202050208021102140220024102440250025502"
+        b"80028a0201040404060409041004120415041804210424044004420445044804"
+        b"5104540456046004810484049004000502050505080511051405200541054405"
+        b"500561058005010604061006260640064206840600080208050808080a081108"
+        b"14082008250841084408500858088008a008aa08010904091009400981098909"
+        b"000a200a280a960aa00a01100410061009101010121015101810211024104010"
+        b"4210451048105110541060106a10811084109010001102110511081111111411"
+        b"2011411144115011801194119611011204120612101240126012001402140514"
+        b"0814111414142014411444144914501464148014011504151015401500161416"
+        b"49160118041810181218401854188618001905196619511aa91a002002200520"
+        b"08200a201120142020204120442050208020a020012104211021402148216521"
+        b"002222228022a82201240424102429244024002541255225992501261a26a626"
+        b"002808280a28202855288828a22868299029082a202a822a882a8a2a01400440"
+        b"0640094010401240154018402140244040404240454048404a40514054406040"
+        b"6540814084409040004102410541084111411441204141414441504180418541"
+        b"a241014204421042124229424042004402440544084411441444194420444144"
+        b"4444504480449444014504451045244540459a4500460a464446504601480448"
+        b"1048404845485448624800491149444950496949044a00500250055008501150"
+        b"145020502850415044505050805001510451105115514051425100524452aa52"
+        b"0154045410542154405460548154a154005508558055885521566856a1560058"
+        b"14584158505899581a5940594259855a0160046010604060546062608660a960"
+        b"006124624a62926200641664106540654565a46501686a682569066a546a626a"
+        b"00800280058008801180148020802a8041804480508080808280a880aa800181"
+        b"0481068110814081518159810082208280828282a082a8820184048410841284"
+        b"158440846084898400854485a58518866a860088088825885a8880888288a888"
+        b"0689228a808a888a968aa88a0190049010904090569084900091229164915692"
+        b"89920094059444945094589429959095929541965198a6984999159a609a00a0"
+        b"02a008a00aa020a02aa0a0a051a159a1a6a100a202a208a22aa280a2a0a240a4"
+        b"95a465a698a60aa820a822a828a8a0a8a8a804a984a986a928aa2aaa91aaaaaa"
+    )
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, rest = np.hsplit(blocks, [2])
+        qs, scales = np.hsplit(rest, [2 * QK_K // 8])
+
+        d = d.view(np.float16).astype(np.float32)
+        qs = qs.view(np.uint16)
+
+        scales = scales.reshape((n_blocks, -1, 1)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2))
+        scales = (scales & 0x0F).reshape((n_blocks, -1))
+        db = d * (np.float32(0.5) + scales) * np.float32(0.25)
+        db = db.reshape((n_blocks, -1, 1, 1))
+
+        # get the sign indices and unpack the bits
+        signs = np.frombuffer(IQ2_XXS.ksigns, dtype=np.uint8).reshape(1, 1, 128)
+        signs = np.take_along_axis(signs, (qs >> 9).reshape((n_blocks, -1, 1)), axis=-1)
+        signs = signs.reshape((n_blocks, -1, 1)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 8))
+        signs = signs & np.uint8(0x01)
+        signs = np.where(signs == 0, np.float32(1), np.float32(-1))
+        signs = signs.reshape((n_blocks, -1, 2, 8))
+
+        assert cls.grid is not None
+        grid = np.take_along_axis(cls.grid, (qs & np.uint16(511)).reshape((n_blocks, -1, 1, 1)), axis=-2)
+        grid = grid.reshape((n_blocks, -1, 2, 8))
+
+        return (db * grid * signs).reshape((n_blocks, -1))
+
+
+class IQ2_S(__Quant, qtype=GGMLQuantizationType.IQ2_S):
+    # iq2s_grid, but with each byte of the original packed in 2 bits,
+    # by mapping 0x08 to 0, 0x19 to 1, and 0x2b to 2.
+    grid_shape = (1024, 8)
+    grid_map = (0x08, 0x19, 0x2b)
+    grid_hex = (
+        b"00000200050008000a0011001400160019002000220025002800410044004600"
+        b"490050005200550058006100640066006900800082008500880091009400a000"
+        b"a500aa0001010401060109011001120115011801210124014001420145014801"
+        b"510154015601590160016501680181018401900192019501a101a40100020202"
+        b"050208021102140220022a02410244024602490250025502800285028a029402"
+        b"a202010404040604090410041204150418042104240426042904400442044504"
+        b"48044a0451045404560459046004620465048104840486048904900495049804"
+        b"a104a40400050205050508050a05110514051605190520052505280541054405"
+        b"46054905500552055505580561056405800582058505880591059405a0050106"
+        b"0406060609061006150640064506480651065406600681068406900600080208"
+        b"050808081108140816081908200825082a084108440846084908500852085508"
+        b"580861086408800885089408aa08010904091009120915091809210940094509"
+        b"480951095409600981099009000a110a140a220a280a2a0a500a990a01100410"
+        b"0610091010101210151018102110241026104010421045104810511054105610"
+        b"59106010621065106810811084108610901095109810a110a410001102110511"
+        b"08110a1111111411161119112011221125112811411144114611491150115211"
+        b"5511581161116411801182118511881191119411011204120912101215122112"
+        b"2412401245125112541281128412901200140214051408141114141416141914"
+        b"2014251428144114441446144914501452145514581461146414801482148514"
+        b"881491149414a014011504150615091510151215151518152115241540154215"
+        b"4515481551155415601581158415901500160516081611161416201641164416"
+        b"50168016aa160118041806180918101815181818211840184218451848185118"
+        b"541860188118841800190219051908191119141920194119441950196919a219"
+        b"041a101a401a561a00200220052008201120142016201920202025202a204120"
+        b"4420502052205520642080208a209420aa200121042110211221152121214021"
+        b"4221452151215421602181218421902100220a22222228222a22442250228822"
+        b"8a22a82201240424062409241024152418242124242440244224452448245124"
+        b"5424602481248424902400250525082511251425202541254425502566258025"
+        b"0126042610264026592600280528112814284128442850288a28aa2801290429"
+        b"102995290a2a222a642a882a8a2a014004400640094010401240154018401a40"
+        b"21402440264040404240454048404a4051405440564059406040624065408140"
+        b"8440904095409840a140a4400041024105410841114114411641194120412241"
+        b"2541414144414641494150415241554158416141644180418241854188419141"
+        b"9441a04101420442104212421542184224424042454248425142544260428142"
+        b"844200440244054408440a441144144416441944204422442544284441444444"
+        b"46444944504452445544584461446444804482448544884491449444a0440145"
+        b"0445064509451045124515451845214524454045424545454845514554456045"
+        b"6a4581458445904500460246054608461146144620464146444650468046a546"
+        b"0148044809481048124815481848214824484048424845484848514854486048"
+        b"84489048004902490549084911491449204941494449504980499649014a044a"
+        b"104a404a00500250055008501150145016501950205022502550285041504450"
+        b"4650495050505250555058506150645080508250855088509150945001510451"
+        b"0651095110511251155118512151245140514251455148515151545160518151"
+        b"8451905100520552085211521452205241524452505269528052015404540654"
+        b"0954105412541554185421542454405442544554485451545454605481548454"
+        b"9054005502550555085511551455205541554455505580550156045610562656"
+        b"405600580258055808581158145820584158445850585a588058015904591059"
+        b"4059005a195a855aa85a01600460066010601260156018602160246040604560"
+        b"4860516054606060846090600061026105610861116114612061416144615061"
+        b"806199610462106240625662a162006405640864116414642064416444645064"
+        b"806401650465106540654a656865926500669466016804681068656898680069"
+        b"2a69426aa16a0080028005800880118014801980208025804180448050805280"
+        b"5580588061808080858091809480018104810981108112811581188121812481"
+        b"408142814581488151815481818184819081a981008205820a82118214824182"
+        b"4482508201840484068409841084128415841884218440844284458448845184"
+        b"5484608481848484908400850285058508851185148520854185448550858085"
+        b"8a85018604861086298640860088058811881488418844885088a28801890489"
+        b"40896589228a588a5a8a828aa28a019004900990109012901590189024904090"
+        b"4290459048905190549060908190849090900091059111911491419144915091"
+        b"5a910192049210924092a6920094029405940894119414942094419444945094"
+        b"8094969401950495109540959895a19500964696649601980498109826984098"
+        b"a998009949995299909a00a005a00aa014a022a02aa041a044a050a0a2a0aaa0"
+        b"40a165a102a20aa222a228a22aa282a288a28aa2a8a201a404a410a440a489a4"
+        b"a4a400a519a551a60aa828a8a2a854a986a908aa0aaa20aa22aa28aa88aaaaaa"
+    )
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, rest = np.hsplit(blocks, [2])
+        qs, rest = np.hsplit(rest, [QK_K // 8])
+        signs, rest = np.hsplit(rest, [QK_K // 8])
+        qh, scales = np.hsplit(rest, [QK_K // 32])
+
+        d = d.view(np.float16).astype(np.float32)
+
+        scales = scales.reshape((n_blocks, -1, 1)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2))
+        scales = (scales & 0x0F).reshape((n_blocks, -1))
+        db = d * (np.float32(0.5) + scales) * np.float32(0.25)
+        db = db.reshape((n_blocks, -1, 1, 1))
+
+        # unpack the sign bits
+        signs = signs.reshape((n_blocks, -1, 1)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 8))
+        signs = signs & np.uint8(0x01)
+        signs = np.where(signs == 0, np.float32(1), np.float32(-1))
+        signs = signs.reshape((n_blocks, -1, 2, 8))
+
+        qh = qh.reshape((n_blocks, -1, 1)) >> np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4))
+        qs = qs.astype(np.uint16) | ((qh & 0x03).astype(np.uint16) << 8).reshape((n_blocks, -1))
+
+        assert cls.grid is not None
+        grid = np.take_along_axis(cls.grid, qs.reshape((n_blocks, -1, 1, 1)), axis=-2)
+        grid = grid.reshape((n_blocks, -1, 2, 8))
+
+        return (db * grid * signs).reshape((n_blocks, -1))
+
+
+class IQ3_XXS(__Quant, qtype=GGMLQuantizationType.IQ3_XXS):
+    grid_shape = (256, 4)
+    grid_map = (0x04, 0x0c, 0x14, 0x1c, 0x24, 0x2c, 0x34, 0x3e)
+    grid_hex = (
+        b"0000020004001100130017002000220031004200730075000101030110011201"
+        b"2101250130013201410154017001000202020402110220022202310233023702"
+        b"5102570275020103070310031203250370031304370444045704730475040105"
+        b"0705320552053506640610071407160743076107011003101010121021102310"
+        b"3010321034104710501000110211111120112211011203121012121221123012"
+        b"7212001302132013311346136613011405145014201524154615711505162217"
+        b"4017002002201120132020202220262031204220012103210521102112212121"
+        b"3021632167217021002202221122172220222222372240225522012310231423"
+        b"7023742335245324032527254125742501270327162745270130103012302130"
+        b"2330503065307230003102312031313144314631013203321032253252327232"
+        b"1133333330344734723400350635223555351436363663363337603704401740"
+        b"3540374053405740744120423742404260426642074345430444514464442545"
+        b"4345704505471047124730471250415070500051065126515551145232527252"
+        b"0253535310542354275472540255315550562457425724604460466064602161"
+        b"6161176264623063366344640565526533660367216703700570077010703270"
+        b"5270267140711272457252720073157333736073217441740075027524753076"
+    )
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, rest = np.hsplit(blocks, [2])
+        qs, scales = np.hsplit(rest, [QK_K // 4])
+
+        d = d.view(np.float16).astype(np.float32)
+        scales = scales.view(np.uint32)
+
+        db = d * (np.float32(0.5) + (scales >> 28).astype(np.float32)) * np.float32(0.5)
+        db = db.reshape((n_blocks, -1, 1, 1))
+
+        # get the sign indices and unpack the bits
+        signs = scales.reshape((n_blocks, -1, 1)) >> np.array([0, 7, 14, 21], dtype=np.uint32).reshape((1, 1, 4))
+        ksigns = np.frombuffer(IQ2_XXS.ksigns, dtype=np.uint8).reshape((1, 1, 1, 128))
+        signs = (signs & np.uint32(0x7F)).reshape((n_blocks, -1, 4, 1))
+        signs = np.take_along_axis(ksigns, signs, axis=-1)
+        signs = signs.reshape((n_blocks, -1, 4, 1)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 1, 8))
+        signs = signs & np.uint8(0x01)
+        signs = np.where(signs == 0, np.float32(1), np.float32(-1))
+        signs = signs.reshape((n_blocks, -1, 4, 8))
+
+        assert cls.grid is not None
+        grid = np.take_along_axis(cls.grid, qs.reshape((n_blocks, -1, 1, 1)), axis=-2)
+        grid = grid.reshape((n_blocks, -1, 4, 8))
+
+        return (db * grid * signs).reshape((n_blocks, -1))
+
+
+class IQ3_S(__Quant, qtype=GGMLQuantizationType.IQ3_S):
+    grid_shape = (512, 4)
+    grid_map = (0x01, 0x03, 0x05, 0x07, 0x09, 0x0b, 0x0d, 0x0f)
+    grid_hex = (
+        b"0000010002000500070010001100120014001600200021002500330040004200"
+        b"4500470051005300600062007100740077000001010102010401100111011501"
+        b"2001230127013101350144016101650172010002010205020702100213021602"
+        b"2102250230023402420245024702510253027002730203031103150320032203"
+        b"3103330336034403500352036703710375030004130417042104240432044004"
+        b"4304510470040205040520052205260533054105450547056605730506061106"
+        b"1306310652067106000702070407200722072607330750075407001001100210"
+        b"0410101011101310151017102010221031103410361054105610611072100011"
+        b"0111031106111011141121113011331141115011521170117611001212121512"
+        b"1712201224123212401243125512601272120113041307131013131321132713"
+        b"3013341341136213701303140514121414143114331442144614501454140115"
+        b"1015131521153015321551152016241627164416461601170317101712172117"
+        b"3517411762177017002001200320052007201020122014201620212023202720"
+        b"3020322041204320452050205220672070207320752000210221102113211721"
+        b"2221252131213421422151210122042207222122232230223722412253225722"
+        b"7122742200230223052311232223242331233323422350236623012407242024"
+        b"2324322435244124722475240425112522253725402553257025002602260726"
+        b"2126552661260527112726273027432750270230113013301530173022303130"
+        b"3330353042304430473051306330713001310331053114312131233140316031"
+        b"7231763100321232203232323432503201331033143321332333273330334133"
+        b"4333473355337333033411341634223431345234603464340135103512352535"
+        b"3235443556357335163641360137033720372237353700400440124020402440"
+        b"2740324041405040704002410741114113412241304135414341514155410142"
+        b"0342104215422142334240425742624270420443114313432043224331433543"
+        b"0044024424443744404471440545074521456245134634466046104715473047"
+        b"4347514702501050145022504050445047505250665074500151035105511251"
+        b"2151325172510052115223523052365253520253075310532753445351536553"
+        b"7353015404542054325446541255265551555355425602570457225711601360"
+        b"1560316033606060006120612761646112623462426255626262706200631463"
+        b"2163406325644364626400650365346560650566406611671367007004700770"
+        b"2070227036704070547062700271117124714371457101720472107216722172"
+        b"3072517202733273357353730174057413742074507422754275027631760077"
+    )
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, rest = np.hsplit(blocks, [2])
+        qs, rest = np.hsplit(rest, [QK_K // 4])
+        qh, rest = np.hsplit(rest, [QK_K // 32])
+        signs, scales = np.hsplit(rest, [QK_K // 8])
+
+        d = d.view(np.float16).astype(np.float32)
+
+        scales = scales.reshape((n_blocks, -1, 1)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2))
+        scales = (scales & 0x0F).reshape((n_blocks, -1))
+        db = d * (1 + 2 * scales)
+        db = db.reshape((n_blocks, -1, 1, 1))
+
+        # unpack the sign bits
+        signs = signs.reshape((n_blocks, -1, 1)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 8))
+        signs = signs & np.uint8(0x01)
+        signs = np.where(signs == 0, np.float32(1), np.float32(-1))
+        signs = signs.reshape((n_blocks, -1, 4, 8))
+
+        qh = qh.reshape((n_blocks, -1, 1)) >> np.array([i for i in range(8)], dtype=np.uint8)
+        qh = (qh & 0x01).astype(np.uint16).reshape((n_blocks, -1))
+        qs = qs.astype(np.uint16) | (qh << 8)
+
+        assert cls.grid is not None
+        grid = np.take_along_axis(cls.grid, qs.reshape((n_blocks, -1, 1, 1)), axis=-2)
+        grid = grid.reshape((n_blocks, -1, 4, 8))
+
+        return (db * grid * signs).reshape((n_blocks, -1))
+
+
+class IQ1_S(__Quant, qtype=GGMLQuantizationType.IQ1_S):
+    # iq1s_grid, with each byte packed into 2 bits
+    # -1, 0, 1 <=> 0, 1, 2
+    grid_shape = (2048, 8)
+    grid_map = (-1, 0, 1)
+    grid_hex = (
+        b"00000200050008000a00110015002000220028002a0045005100540056006500"
+        b"8000820088008a009500a000a200a800aa000401050111011401160119011a01"
+        b"2501410146014901520155015a0161016401660168018501910194019601a501"
+        b"0002020208020a0215022002220228022a024502510259026402690280028202"
+        b"88028a02910295029902a002a202a802aa021104140416042504410449045504"
+        b"5a046404650491049904a5040105040505050605150518051a05290540054505"
+        b"4a0550055105540555055605590560056205650568056a058105910595059805"
+        b"9a05a105a405a505a605a9051406190641064406500652065506580660066106"
+        b"6606690685069106940699060008020808080a0815082008220828082a084508"
+        b"5108560865088008820888088a089508a008a208a808aa080509110914091909"
+        b"2409250941095009510955096109640969099109940996099909a509000a020a"
+        b"080a0a0a150a200a220a280a2a0a450a510a590a610a650a800a820a850a880a"
+        b"8a0a950aa00aa20aa80aaa0a1010111014101910241025104110441050105510"
+        b"58106110641065106910911094109610a110a510011104110611091110111211"
+        b"1511181121112411291145114a11501151115211541155115611591160116511"
+        b"841192119511a111a41111121412161225124012461249125212551258125a12"
+        b"641266128512911294129612a512011406140914141415141814191421142614"
+        b"41144514461448144a1451145414551456145914621465146814841489149014"
+        b"94149514981499149a14a114a414a514a914021505150a151115141515151615"
+        b"191520152215251528152a154115441545154615511552155415551556155915"
+        b"5a1561156415651566156915801582158415851588158a159015911594159515"
+        b"961599159a15a015a215a51501160416051606161516161618161a1621162616"
+        b"401642164416451648164a165116551656165816591661166416651668166916"
+        b"6a1686168a1692169516a416a916111816182518411844184618491850185518"
+        b"58185a1860186118641866186918851891189418a5181019121915191a192119"
+        b"25194219441945194819511954195519561959195a19601965196a1989199119"
+        b"921995199819a119a619a919091a161a241a261a441a461a491a501a521a551a"
+        b"581a611a661a691a851a911a961a9a1a0020022008200a201520202022202520"
+        b"28202a20452051205920612065208020822088208a209520a020a220a520a820"
+        b"aa2005211121142119212521422144214921552158215a216121642165216621"
+        b"8521902196219921a521012208220a22112215222022222228222a2245225122"
+        b"562259226522812288228a2291229522a022a222a822aa220524142416241924"
+        b"252444244524462449245224552458245a2466248524912494249924a124a524"
+        b"0925152521252925402545254825512554255525592562256525682589259025"
+        b"9425952598259a25a125a425a625a92505261026122619262526412649265526"
+        b"6026612669268426862690269a260028022808280a2815282028222828282a28"
+        b"45285128542865288028822888288a28a028a228a828aa280929112914291929"
+        b"2529462949295229552961296429662969298529902996299929a429a529002a"
+        b"022a082a0a2a202a222a282a2a2a452a512a562a592a652a802a822a882a8a2a"
+        b"952aa02aa22aa82aaa2a054011401640254049405240554058405a4061406440"
+        b"664094409940a140a6400041014104410641094112411541164118411a412141"
+        b"26412941454148414a41514154415541564159415a41654168416a4181418441"
+        b"8641904192419541a041a141a241054211421442164225424142524255425a42"
+        b"6442694289429442a5420144154419442944454448444a445144544455445644"
+        b"61446244654468446a44814486448944904492449544a044a144a94401450245"
+        b"05450a4511451445154516451945204525452a45414544454545464549455045"
+        b"5145544555455645584559456145644565456645694582458445854588459145"
+        b"94459545964599459a45a545a845aa450146054609461446154618461a462146"
+        b"2446294640464246454648465046514652465546564659466246654668468146"
+        b"85468a4694469546a146a446a6460548114815481a4825484248494850485548"
+        b"5848614864486648694885489148944896489948a5480149054906490a491049"
+        b"144915491849214924492649404945494a495149524954495549564959496049"
+        b"6249654966496a49864989499249954996499849a149a449a649a949164a444a"
+        b"464a494a554a584a5a4a644a694a944aa54a0150045005500650095012501550"
+        b"1a50215024502950405045504850515054505550565059506550685086508950"
+        b"95509850a050a150a650a9500551085109510a51115114511551165118511951"
+        b"20512551265128512a5141514451455146514951505151515251545155515651"
+        b"585159515a51615164516551665169518251855191519451955196519951a051"
+        b"a551aa5101520652125215521a5221522452425245524a525152545255525652"
+        b"595262526552855290529252955299529a52a452045405541154145415541654"
+        b"185419542154255428542a54415444544554465449544a545054515454545554"
+        b"5654585459545a54615462546454655466546954805488548a54915494549554"
+        b"96549954a154a454a554aa540155025504550555065509551055115512551455"
+        b"1555165519551a55215524552555265529554055415542554455455546554855"
+        b"4955505551555255545555555655585559555a55605561556455655566556855"
+        b"69556a5581558455855589558a559055915594559555965598559955a155a455"
+        b"a555a655a9550056015602560456065608560956115614561556185619562056"
+        b"2156225624562556265628562956415645564656485649564a56505651565256"
+        b"545655565656585659565a566156645665566956825685568656885689568a56"
+        b"915695569a56a256a556a656a856a95604580558065809581058155818582158"
+        b"2a58455848584a58515854585558565858585958605862586458655882588958"
+        b"9058925895589858a158a9580159025905590a59115914591559165919592559"
+        b"41594459455946594959505951595259545955595659585959595a5961596459"
+        b"655966596959815985598959915994599559965998599959a559045a085a155a"
+        b"1a5a205a255a265a295a455a485a495a515a555a565a585a595a625a655a685a"
+        b"6a5a815a8a5a925a955a965a985a9a5aa15a0560146016601960256044605060"
+        b"5560566058605a60616064606660696081609660a56001610461066109611261"
+        b"15612161226126612961456149615161556156615961656166616a6184618a61"
+        b"92619561a161a661a96111621662196240624162466255625662586260628562"
+        b"91629662a56211641264156416641a6421642664296440644264456448644a64"
+        b"516454645564566459645a646064626465648464856489649064926494649564"
+        b"966498649a64a164a464a964056508650a651165156516651965446545654665"
+        b"496550655165546555655665596561656465656566656965866589658a659165"
+        b"9565966599659a65a265a565a665a86502660966156620662666286629664066"
+        b"456648664a66516654665566566658665a666066656668668066826685668a66"
+        b"9466966698669966a066a466a666aa661668196825684168526855685a686168"
+        b"6968856891689868a66801690469106915692169246926692969406941694569"
+        b"4669486951695469556956695969606965696a69826984698a699569a169a469"
+        b"a569a969116a166a186a416a446a496a506a556a586a5a6a646a656a696a866a"
+        b"946a986a9a6aa66a0080028008800a802080228028802a804580508051805480"
+        b"5680598065808080828088808a809580a080a280a880aa800581118114811681"
+        b"1981258141814481498150815281558156815881598164816681698185818981"
+        b"948196819981a5810082028208820a8215822082228228822a82518254825982"
+        b"65828082828288828a829582a082a282a882aa82148419844184448451845584"
+        b"5a846184648469849484998401850985128515851a8526852985408541854585"
+        b"4885518554855585568559855a856585668568856a8581858485868589859085"
+        b"928595859885a68511861686198625864186448649864a865086558659865a86"
+        b"618666866a86858691869a86a4860088028808880a8815882088228828882a88"
+        b"41884588518854885988658869888088828888888a889588a088a288a888aa88"
+        b"05890689118914891689258941894489468949895089528955895a8961896489"
+        b"858996899989a589008a028a088a0a8a158a208a228a288a2a8a458a518a548a"
+        b"568a808a828a888a8a8a958aa08aa28aa88aaa8a059011901690189019902590"
+        b"419046904990559058905a9069906a9085909190949096909990a59001910491"
+        b"069109911091159118911a912191249126912991409145915091519154915591"
+        b"569159916291659184918691929195919891a191a491a691a991059211921492"
+        b"19922592449246924992509252925592589266926992859294929692a9920194"
+        b"04940694109415941894269440944a9451945494559456945894599460946194"
+        b"62946594849486949294949495949894a194a9940095059508950a9510951195"
+        b"14951595169519952195259529952a9541954495459546954995509551955295"
+        b"549555955695589559955a956195649565956695699581958595889591959295"
+        b"94959595969599959a95a095a295a595a895aa95019604961096159619962096"
+        b"2696299645964896499651965296559656965996659668968296849689968a96"
+        b"929694969596a496a696a9960598169819982598419846985098529855985698"
+        b"5a98649865988598919896989998a59804990699099910991299159918991a99"
+        b"209921992499269940994299459948994a995199549955995699599962996599"
+        b"66996a99819984999099929995999a99a199a699059a159a259a449a469a499a"
+        b"509a559a589a619a859a919a949a959a969a00a002a008a00aa015a020a022a0"
+        b"28a02aa045a051a054a056a059a080a082a088a08aa095a0a0a0a2a0a8a0aaa0"
+        b"05a109a111a114a116a119a11aa146a149a151a155a158a15aa161a164a185a1"
+        b"90a192a196a199a102a208a20aa210a219a222a228a22aa245a251a256a259a2"
+        b"65a280a282a288a28aa295a2a0a2a2a2a8a2aaa219a425a441a444a450a454a4"
+        b"55a458a45aa461a465a466a468a469a485a406a509a510a512a515a518a526a5"
+        b"29a542a545a551a554a555a556a559a565a56aa581a584a585a586a589a592a5"
+        b"95a598a505a611a616a61aa621a625a644a646a64aa652a655a656a658a660a6"
+        b"62a686a690a695a696a699a6a1a6a4a6a6a600a802a808a80aa820a822a828a8"
+        b"2aa851a854a856a859a880a882a888a88aa895a8a0a8a2a8a8a8aaa805a914a9"
+        b"19a921a925a941a950a955a95aa961a966a969a990a996a900aa02aa08aa0aaa"
+        b"20aa22aa28aa2aaa51aa54aa56aa80aa82aa88aa8aaa95aaa0aaa2aaa8aaaaaa"
+    )
+
+    delta = np.float32(0.125)
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, rest = np.hsplit(blocks, [2])
+        qs, qh = np.hsplit(rest, [QK_K // 8])
+
+        d = d.view(np.float16).astype(np.float32)
+        qh = qh.view(np.uint16)
+
+        dl = d * (2 * ((qh >> 12) & 7) + 1)
+        dl = dl.reshape((n_blocks, -1, 1, 1))
+        delta = np.where((qh & np.uint16(0x8000)) == 0, cls.delta, -cls.delta)
+        delta = delta.reshape((n_blocks, -1, 1, 1))
+
+        qh = qh.reshape((n_blocks, -1, 1)) >> np.array([0, 3, 6, 9], dtype=np.uint16).reshape((1, 1, 4))
+        qs = qs.astype(np.uint16) | ((qh & 7) << 8).reshape((n_blocks, -1))
+
+        assert cls.grid is not None
+        grid = np.take_along_axis(cls.grid, qs.reshape((n_blocks, -1, 1, 1)), axis=-2)
+        grid = grid.reshape((n_blocks, -1, 4, 8))
+
+        return (dl * (grid + delta)).reshape((n_blocks, -1))
+
+
+class IQ1_M(__Quant, qtype=GGMLQuantizationType.IQ1_M):
+    grid_shape = IQ1_S.grid_shape
+    grid_map = IQ1_S.grid_map
+    grid_hex = IQ1_S.grid_hex
+
+    delta = IQ1_S.delta
+
+    # Okay *this* type is weird. It's the only one which stores the f16 scales in multiple parts.
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        qs, rest = np.hsplit(blocks, [QK_K // 8])
+        qh, scales = np.hsplit(rest, [QK_K // 16])
+
+        # The f16 scale is packed across multiple bytes
+        scales = scales.view(np.uint16)
+        d = (scales.reshape((n_blocks, 4)) & np.uint16(0xF000)) >> np.array([12, 8, 4, 0], dtype=np.uint16).reshape((1, 4))
+        d = d[..., 0] | d[..., 1] | d[..., 2] | d[..., 3]
+        d = d.view(np.float16).astype(np.float32).reshape((n_blocks, 1))
+
+        scales = scales.reshape(n_blocks, -1, 1) >> np.array([0, 3, 6, 9], dtype=np.uint16).reshape((1, 1, 4))
+        scales = (scales & 0x07).reshape((n_blocks, -1))
+        dl = d * (2 * scales + 1)
+        dl = dl.reshape((n_blocks, -1, 2, 1, 1))
+
+        qh = qh.reshape((n_blocks, -1, 1)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2))
+        qs = qs.astype(np.uint16) | ((qh & 0x07).astype(np.uint16) << 8).reshape((n_blocks, -1))
+
+        delta = np.where(qh & 0x08 == 0, cls.delta, -cls.delta)
+        delta = delta.reshape((n_blocks, -1, 2, 2, 1))
+
+        assert cls.grid is not None
+        grid = np.take_along_axis(cls.grid, qs.reshape((n_blocks, -1, 1, 1)), axis=-2)
+        grid = grid.reshape((n_blocks, -1, 2, 2, 8))
+
+        return (dl * (grid + delta)).reshape((n_blocks, -1))
+
+
+class IQ4_NL(__Quant, qtype=GGMLQuantizationType.IQ4_NL):
+    kvalues = (-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113)
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, qs = np.hsplit(blocks, [2])
+
+        d = d.view(np.float16).astype(np.float32)
+
+        qs = qs.reshape((n_blocks, -1, 1, cls.block_size // 2)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
+
+        qs = (qs & np.uint8(0x0F)).reshape((n_blocks, -1, 1))
+
+        kvalues = np.array(cls.kvalues, dtype=np.int8).reshape(1, 1, 16)
+        qs = np.take_along_axis(kvalues, qs, axis=-1).astype(np.float32).reshape((n_blocks, -1))
+
+        return (d * qs)
+
+
+class IQ4_XS(__Quant, qtype=GGMLQuantizationType.IQ4_XS):
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d, rest = np.hsplit(blocks, [2])
+        scales_h, rest = np.hsplit(rest, [2])
+        scales_l, qs = np.hsplit(rest, [QK_K // 64])
+
+        d = d.view(np.float16).astype(np.float32)
+        scales_h = scales_h.view(np.uint16)
+
+        scales_l = scales_l.reshape((n_blocks, -1, 1)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2))
+        scales_h = scales_h.reshape((n_blocks, 1, -1)) >> np.array([2 * i for i in range(QK_K // 32)], dtype=np.uint16).reshape((1, -1, 1))
+        scales_l = scales_l.reshape((n_blocks, -1)) & np.uint8(0x0F)
+        scales_h = scales_h.reshape((n_blocks, -1)).astype(np.uint8) & np.uint8(0x03)
+
+        scales = (scales_l | (scales_h << np.uint8(4))).astype(np.int8) - np.int8(32)
+        dl = (d * scales.astype(np.float32)).reshape((n_blocks, -1, 1))
+
+        qs = qs.reshape((n_blocks, -1, 1, 16)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
+        qs = qs.reshape((n_blocks, -1, 32, 1)) & np.uint8(0x0F)
+
+        kvalues = np.array(IQ4_NL.kvalues, dtype=np.int8).reshape((1, 1, 1, -1))
+        qs = np.take_along_axis(kvalues, qs, axis=-1).astype(np.float32).reshape((n_blocks, -1, 32))
+
+        return (dl * qs).reshape((n_blocks, -1))
diff --git a/gguf-py/tests/test_quants.py b/gguf-py/tests/test_quants.py
new file mode 100755
index 00000000..8b7a85c2
--- /dev/null
+++ b/gguf-py/tests/test_quants.py
@@ -0,0 +1,237 @@
+#!/usr/bin/env python3
+
+# Test gguf.quants so that it exactly matches the C implementation of the (de)quantization
+
+# NOTE: this is kind of a mess, but at least it worked for initially testing the Python implementations.
+
+from __future__ import annotations
+
+import argparse
+from math import prod
+import os
+import sys
+from pathlib import Path
+import ctypes
+import logging
+import numpy as np
+
+# Necessary to load the local gguf package
+if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent / 'gguf-py').exists():
+    sys.path.insert(0, str(Path(__file__).parent.parent))
+
+import gguf
+from gguf.constants import GGMLQuantizationType
+
+
+logger = logging.getLogger("test-quants")
+
+
+c_float_p = ctypes.POINTER(ctypes.c_float)
+
+
+class ggml_init_params(ctypes.Structure):
+    _fields_ = [
+        ("mem_size", ctypes.c_size_t),
+        ("mem_buffer", ctypes.c_void_p),
+        ("no_alloc", ctypes.c_bool),
+    ]
+
+
+class GGMLQuants:
+    libggml: ctypes.CDLL
+
+    def __init__(self, libggml: Path):
+        self.libggml = ctypes.CDLL(str(libggml))
+        self.libggml.ggml_quantize_chunk.restype = ctypes.c_size_t
+        # enum ggml_type   type,
+        #    const float * src,
+        #           void * dst,
+        #        int64_t   start,
+        #        int64_t   nrows,
+        #        int64_t   n_per_row,
+        #    const float * imatrix) {
+        self.libggml.ggml_quantize_chunk.argtypes = (
+            ctypes.c_int,
+            ctypes.POINTER(ctypes.c_float),
+            ctypes.c_void_p,
+            ctypes.c_int64,
+            ctypes.c_int64,
+            ctypes.c_int64,
+            ctypes.POINTER(ctypes.c_float),
+        )
+
+        self.libggml.ggml_quantize_requires_imatrix.restype = ctypes.c_bool
+        self.libggml.ggml_quantize_requires_imatrix.argtypes = (ctypes.c_int,)
+
+        for t in (
+            "q4_0", "q4_1", "q5_0", "q5_1", "q8_0",
+            "q2_K", "q3_K", "q4_K", "q5_K", "q6_K",
+            "iq2_xxs", "iq2_xs", "iq2_s", "iq3_xxs", "iq3_s", "iq1_s", "iq1_m",
+            "iq4_nl", "iq4_xs",
+        ):
+            dequant_func: ctypes._NamedFuncPointer = getattr(self.libggml, "dequantize_row_" + t)
+            dequant_func.restype = None
+            dequant_func.argtypes = (ctypes.c_void_p, ctypes.POINTER(ctypes.c_float), ctypes.c_int64)
+
+        self.libggml.ggml_fp16_to_fp32_row.restype = None
+        self.libggml.ggml_fp16_to_fp32_row.argtypes = (ctypes.POINTER(ctypes.c_uint16), ctypes.POINTER(ctypes.c_float), ctypes.c_int64)
+        self.libggml.ggml_bf16_to_fp32_row.restype = None
+        self.libggml.ggml_bf16_to_fp32_row.argtypes = (ctypes.POINTER(ctypes.c_uint16), ctypes.POINTER(ctypes.c_float), ctypes.c_int64)
+
+        self.libggml.ggml_init.argtypes = (ggml_init_params,)
+
+        self.libggml.ggml_init(ggml_init_params(1 * 1024 * 1024, 0, False))
+
+    def dequantize(self, tensor: np.ndarray, qtype: GGMLQuantizationType) -> np.ndarray:
+        result = np.zeros(gguf.quant_shape_from_byte_shape(tensor.shape, qtype), dtype=np.float32, order="C")
+        if qtype == GGMLQuantizationType.F32:
+            # no-op
+            result = tensor.view(np.float32)
+        elif qtype == GGMLQuantizationType.F16:
+            self.libggml.ggml_fp16_to_fp32_row(tensor.ctypes.data_as(ctypes.POINTER(ctypes.c_uint16)), result.ctypes.data_as(c_float_p), result.size)
+        elif qtype == GGMLQuantizationType.BF16:
+            self.libggml.ggml_bf16_to_fp32_row(tensor.ctypes.data_as(ctypes.POINTER(ctypes.c_uint16)), result.ctypes.data_as(c_float_p), result.size)
+        else:
+            lw_qname = qtype.name.lower()
+            if lw_qname[-1] == "k":
+                lw_qname = lw_qname[:-1] + "K"
+            dequant_func: ctypes._NamedFuncPointer = getattr(self.libggml, "dequantize_row_" + lw_qname)
+            dequant_func(tensor.ctypes.data_as(ctypes.c_void_p), result.ctypes.data_as(c_float_p), result.size)
+        return result
+
+    def quantize(self, data: np.ndarray, qtype: GGMLQuantizationType) -> np.ndarray:
+        result = np.zeros(gguf.quant_shape_to_byte_shape(data.shape, qtype), dtype=np.uint8, order="C")
+        if self.libggml.ggml_quantize_requires_imatrix(qtype.value):
+            # TODO: is a column-wise sum of squares appropriate?
+            qw = np.sum((data * data).reshape((-1, data.shape[-1])), axis=0).ctypes.data_as(c_float_p)
+        else:
+            qw = ctypes.cast(0, c_float_p)
+        result_size = self.libggml.ggml_quantize_chunk(qtype.value, data.ctypes.data_as(c_float_p), result.ctypes.data_as(ctypes.c_void_p), 0, prod(data.shape[:-1]), data.shape[-1], qw)
+        assert result.size == result_size
+        return result
+
+
+def compare_tensors(t1: np.ndarray, t2: np.ndarray, qtype: GGMLQuantizationType) -> bool:
+    same = np.array_equal(t1, t2)
+    if same:
+        return True
+    else:
+        block_size, type_size = gguf.GGML_QUANT_SIZES[qtype]
+        if t1.dtype == np.float32:
+            t1 = t1.reshape((-1, block_size))
+            t2 = t2.reshape((-1, block_size))
+        else:
+            t1 = t1.reshape((-1, type_size))
+            t2 = t2.reshape((-1, type_size))
+        x = t1.view(np.uint8) ^ t2.view(np.uint8)
+        diff_bits = np.count_nonzero(np.unpackbits(x, axis=-1), axis=-1)
+        num_bad_blocks = np.count_nonzero(diff_bits, axis=0)
+        if num_bad_blocks == 0 and t1.shape == t2.shape:
+            logger.debug("Bits are equal, but arrays don't match, likely contains NANs")
+            return True
+        logger.debug(f"{num_bad_blocks} bad blocks ({100 * num_bad_blocks / x.shape[0]:.6f}%)")
+        bad_block_id = np.argmax(diff_bits, axis=0)
+        logger.debug(f"Worst block id: {bad_block_id}")
+        logger.debug(f"Sample bad block ({diff_bits[bad_block_id]} differing bits):\n{t1[bad_block_id]}\nReference:\n{t2[bad_block_id]}")
+
+        sum_diff_bits = np.sum(diff_bits)
+        logger.debug(f"{sum_diff_bits} bits differ ({100 * sum_diff_bits/(x.size * 8):.6f}%)")
+        return False
+
+
+def do_test(libggml_path: Path, quick: bool = False):
+    ggml_quants = GGMLQuants(libggml_path)
+
+    np.set_printoptions(precision=None, threshold=(4 * 256) + 1, formatter={"int": lambda n: "0x%02X" % n})
+
+    r = np.random.randn(8, 1024, 1024).astype(np.float32, copy=False)
+
+    for qtype in (GGMLQuantizationType.F16, *gguf.quants._type_traits.keys()):
+        has_dequantize = False
+        has_quantize = False
+
+        try:
+            gguf.dequantize(np.zeros((gguf.GGML_QUANT_SIZES[qtype][1]), dtype=np.uint8), qtype)
+            has_dequantize = True
+        except (NotImplementedError, AssertionError) as e:
+            if isinstance(e, AssertionError):
+                logger.error(f"Error with {qtype.name}: {e}")
+                raise e
+        try:
+            gguf.quantize(np.zeros((gguf.GGML_QUANT_SIZES[qtype][0]), dtype=np.float32), qtype)
+            has_quantize = True
+        except (NotImplementedError, AssertionError) as e:
+            if isinstance(e, AssertionError):
+                logger.error(f"Error with {qtype.name}: {e}")
+                raise e
+
+        if not has_dequantize and not has_quantize:
+            continue
+
+        logger.info(f"Testing {qtype.name}")
+
+        rc = r.copy(order="C")
+
+        pyq = None
+        ggq = None
+
+        if has_quantize:
+            logger.debug(f"Quantizing to {qtype.name} with Python")
+            pyq = gguf.quants.quantize(rc, qtype)
+
+            logger.debug(f"Quantizing to {qtype.name} with C")
+            ggq = ggml_quants.quantize(rc, qtype)
+
+            if qtype == GGMLQuantizationType.F16:
+                pyq = pyq.view(np.uint8)
+            quant_equal = compare_tensors(pyq, ggq, qtype)
+
+            if not quant_equal:
+                logger.error(f"Quantization to {qtype.name} does not match ❌")
+            else:
+                logger.info(f"Quantization to {qtype.name} matches exactly ✅")
+
+        if has_dequantize:
+            if ggq is None and not quick:
+                logger.debug(f"Quantizing to {qtype.name} with C")
+                ggq = ggml_quants.quantize(rc, qtype)
+
+            if ggq is not None:
+                logger.debug(f"Dequantizing from {qtype.name} with Python")
+                pydq = gguf.quants.dequantize(ggq, qtype)
+                logger.debug(f"Dequantizing from {qtype.name} with C")
+                ggdq = ggml_quants.dequantize(ggq, qtype)
+
+                dequant_equal = compare_tensors(pydq, ggdq, qtype)
+
+                if not dequant_equal:
+                    logger.error(f"Dequantization from {qtype.name} does not match ❌")
+                else:
+                    logger.info(f"Dequantization from {qtype.name} matches exactly ✅")
+
+            rq_shape = gguf.quants.quant_shape_to_byte_shape((8, 1024, 1024 // 2), qtype)
+            rq = np.random.random(rq_shape).astype(np.float16).view(np.uint8)
+
+            logger.debug(f"Dequantizing random f16 data as {qtype.name} with Python")
+            pydq = gguf.quants.dequantize(rq, qtype)
+            logger.debug(f"Dequantizing random f16 data as {qtype.name} with C")
+            ggdq = ggml_quants.dequantize(rq, qtype)
+
+            dequant_equal = compare_tensors(pydq, ggdq, qtype)
+
+            if not dequant_equal:
+                logger.error(f"Dequantization from random f16 data as {qtype.name} does not match ❌")
+            else:
+                logger.info(f"Dequantization from random f16 data as {qtype.name} matches exactly ✅")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Test Python (de)quantization against the reference C implementation")
+    parser.add_argument("--libggml", type=Path, default=Path(__file__).parent.parent.parent / "build" / "ggml" / "src" / "libggml.so", help="The path to libggml.so")
+    parser.add_argument("--quick", action="store_true", help="Don't quantize with C when it's not strictly necessary")
+
+    args = parser.parse_args()
+
+    logging.basicConfig(level=logging.DEBUG)
+
+    do_test(args.libggml, args.quick)
diff --git a/include/llama.h b/include/llama.h
index 9ae88060..a9af4c48 100644
--- a/include/llama.h
+++ b/include/llama.h
@@ -33,17 +33,15 @@
 
 #define LLAMA_DEFAULT_SEED 0xFFFFFFFF
 
-#define LLAMA_MAX_RNG_STATE (64*1024)
-
 #define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla'
 #define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
 #define LLAMA_FILE_MAGIC_GGSQ 0x67677371u // 'ggsq'
 
 #define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
-#define LLAMA_SESSION_VERSION 7
+#define LLAMA_SESSION_VERSION 8
 
 #define LLAMA_STATE_SEQ_MAGIC   LLAMA_FILE_MAGIC_GGSQ
-#define LLAMA_STATE_SEQ_VERSION 1
+#define LLAMA_STATE_SEQ_VERSION 2
 
 #ifdef __cplusplus
 extern "C" {
@@ -355,7 +353,7 @@ extern "C" {
         int32_t nthread;                     // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
         enum llama_ftype ftype;              // quantize to this llama_ftype
         enum ggml_type output_tensor_type;   // output tensor type
-        enum ggml_type token_embedding_type; // itoken embeddings tensor type
+        enum ggml_type token_embedding_type; // token embeddings tensor type
         bool allow_requantize;               // allow quantizing non-f32/f16 tensors
         bool quantize_output_tensor;         // quantize output.weight
         bool only_copy;                      // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
@@ -515,6 +513,9 @@ extern "C" {
     // Returns true if the model contains an encoder that requires llama_encode() call
     LLAMA_API bool llama_model_has_encoder(const struct llama_model * model);
 
+    // Returns true if the model contains a decoder that requires llama_decode() call
+    LLAMA_API bool llama_model_has_decoder(const struct llama_model * model);
+
     // For encoder-decoder models, this function returns id of the token that must be provided
     // to the decoder to start generating output sequence. For other models, it returns -1.
     LLAMA_API llama_token llama_model_decoder_start_token(const struct llama_model * model);
@@ -700,10 +701,11 @@ extern "C" {
     // State / sessions
     //
 
-    // Returns the maximum size in bytes of the state (rng, logits, embedding
-    // and kv_cache) - will often be smaller after compacting tokens
-    LLAMA_API size_t llama_state_get_size(const struct llama_context * ctx);
-    LLAMA_API DEPRECATED(size_t llama_get_state_size(const struct llama_context * ctx),
+    // Returns the *actual* size in bytes of the state
+    // (rng, logits, embedding and kv_cache)
+    // Only use when saving the state, not when restoring it, otherwise the size may be too small.
+    LLAMA_API size_t llama_state_get_size(struct llama_context * ctx);
+    LLAMA_API DEPRECATED(size_t llama_get_state_size(struct llama_context * ctx),
         "use llama_state_get_size instead");
 
     // Copies the state to the specified destination address.
@@ -711,7 +713,8 @@ extern "C" {
     // Returns the number of bytes copied
     LLAMA_API size_t llama_state_get_data(
             struct llama_context * ctx,
-                         uint8_t * dst);
+                         uint8_t * dst,
+                          size_t   size);
     LLAMA_API DEPRECATED(size_t llama_copy_state_data(
             struct llama_context * ctx,
                          uint8_t * dst),
@@ -721,7 +724,8 @@ extern "C" {
     // Returns the number of bytes read
     LLAMA_API size_t llama_state_set_data(
             struct llama_context * ctx,
-                   const uint8_t * src);
+                   const uint8_t * src,
+                          size_t   size);
     LLAMA_API DEPRECATED(size_t llama_set_state_data(
             struct llama_context * ctx,
                    const uint8_t * src),
@@ -763,6 +767,7 @@ extern "C" {
     LLAMA_API size_t llama_state_seq_get_data(
             struct llama_context * ctx,
                          uint8_t * dst,
+                          size_t   size,
                     llama_seq_id   seq_id);
 
     // Copy the sequence data (originally copied with `llama_state_seq_get_data`) into the specified sequence
@@ -772,6 +777,7 @@ extern "C" {
     LLAMA_API size_t llama_state_seq_set_data(
             struct llama_context * ctx,
                    const uint8_t * src,
+                          size_t   size,
                     llama_seq_id   dest_seq_id);
 
     LLAMA_API size_t llama_state_seq_save_file(
diff --git a/scripts/sync-ggml-am.sh b/scripts/sync-ggml-am.sh
index ba3bedf2..b2989256 100755
--- a/scripts/sync-ggml-am.sh
+++ b/scripts/sync-ggml-am.sh
@@ -62,8 +62,10 @@ while read c; do
         src/ggml*.m \
         src/ggml*.metal \
         src/ggml*.cu \
+        src/ggml-cann/* \
         src/ggml-cuda/* \
         src/ggml-sycl/* \
+        src/vulkan-shaders/* \
         include/ggml*.h \
         tests/test-opt.cpp \
         tests/test-grad0.cpp \
@@ -102,9 +104,13 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
     # cmake/FindSIMD.cmake    -> ggml/cmake/FindSIMD.cmake
     #
     # src/ggml.c              -> ggml/src/ggml.c
+    # src/ggml-aarch64.c      -> ggml/src/ggml-aarch64.c
+    # src/ggml-aarch64.h      -> ggml/src/ggml-aarch64.h
     # src/ggml-alloc.c        -> ggml/src/ggml-alloc.c
     # src/ggml-backend-impl.h -> ggml/src/ggml-backend-impl.h
     # src/ggml-backend.c      -> ggml/src/ggml-backend.c
+    # src/ggml-cann/*         -> ggml/src/ggml-cann/
+    # src/ggml-cann.cpp       -> ggml/src/ggml-cann.cpp
     # src/ggml-common.h       -> ggml/src/ggml-common.h
     # src/ggml-cuda/*         -> ggml/src/ggml-cuda/
     # src/ggml-cuda.cu        -> ggml/src/ggml-cuda.cu
@@ -117,11 +123,13 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
     # src/ggml-sycl/*         -> ggml/src/ggml-sycl/
     # src/ggml-sycl.cpp       -> ggml/src/ggml-sycl.cpp
     # src/ggml-vulkan.cpp     -> ggml/src/ggml-vulkan.cpp
+    # src/vulkan-shaders/*    -> ggml/src/vulkan-shaders/
     #
     # include/ggml.h         -> ggml/include/ggml.h
     # include/ggml-alloc.h   -> ggml/include/ggml-alloc.h
     # include/ggml-backend.h -> ggml/include/ggml-backend.h
     # include/ggml-blas.h    -> ggml/include/ggml-blas.h
+    # include/ggml-cann.h    -> ggml/include/ggml-cann.h
     # include/ggml-cuda.h    -> ggml/include/ggml-cuda.h
     # include/ggml-kompute.h -> ggml/include/ggml-kompute.h
     # include/ggml-metal.h   -> ggml/include/ggml-metal.h
@@ -143,9 +151,13 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
         -e 's/([[:space:]]|[ab]\/)src\/CMakeLists.txt/\1ggml\/src\/CMakeLists.txt/g' \
         -e 's/([[:space:]]|[ab]\/)cmake\/FindSIMD.cmake/\1ggml\/cmake\/FindSIMD.cmake/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml\.c/\1ggml\/src\/ggml.c/g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-aarch64\.c/\1ggml\/src\/ggml-aarch64.c/g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-aarch64\.h/\1ggml\/src\/ggml-aarch64.h/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-alloc\.c/\1ggml\/src\/ggml-alloc.c/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-backend-impl\.h/\1ggml\/src\/ggml-backend-impl.h/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-backend\.c/\1ggml\/src\/ggml-backend.c/g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-cann\//\1ggml\/src\/ggml-cann\//g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-cann\.cpp/\1ggml\/src\/ggml-cann.cpp/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-common\.h/\1ggml\/src\/ggml-common.h/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-cuda\//\1ggml\/src\/ggml-cuda\//g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-cuda\.cu/\1ggml\/src\/ggml-cuda.cu/g' \
@@ -158,22 +170,24 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
         -e 's/([[:space:]]|[ab]\/)src\/ggml-sycl\//\1ggml\/src\/ggml-sycl\//g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-sycl\.cpp/\1ggml\/src\/ggml-sycl.cpp/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-vulkan\.cpp/\1ggml\/src\/ggml-vulkan.cpp/g' \
+        -e 's/([[:space:]]|[ab]\/)src\/vulkan-shaders\//\1ggml\/src\/vulkan-shaders\//g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml\.h/\1ggml\/include\/ggml.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-alloc\.h/\1ggml\/include\/ggml-alloc.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-backend\.h/\1ggml\/include\/ggml-backend.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-blas\.h/\1ggml\/include\/ggml-blas.h/g' \
+        -e 's/([[:space:]]|[ab]\/)include\/ggml-cann\.h/\1ggml\/include\/ggml-cann.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-cuda\.h/\1ggml\/include\/ggml-cuda.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-kompute\.h/\1ggml\/include\/ggml-kompute.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-metal\.h/\1ggml\/include\/ggml-metal.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-rpc\.h/\1ggml\/include\/ggml-rpc.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-sycl\.h/\1ggml\/include\/ggml-sycl.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-vulkan\.h/\1ggml\/include\/ggml-vulkan.h/g' \
-        -e 's/([[:space:]]|[ab]\/)examples\/common\.h/examples\/common.h/g' \
-        -e 's/([[:space:]]|[ab]\/)examples\/common\.cpp/examples\/common.cpp/g' \
-        -e 's/([[:space:]]|[ab]\/)examples\/common-ggml\.h/examples\/common-ggml.h/g' \
-        -e 's/([[:space:]]|[ab]\/)examples\/common-ggml\.cpp/examples\/common-ggml.cpp/g' \
-        -e 's/([[:space:]]|[ab]\/)LICENSE/LICENSE/g' \
-        -e 's/([[:space:]]|[ab]\/)scripts\/gen-authors\.sh/scripts\/gen-authors.sh/g' \
+        -e 's/([[:space:]]|[ab]\/)examples\/common\.h/\1examples\/common.h/g' \
+        -e 's/([[:space:]]|[ab]\/)examples\/common\.cpp/\1examples\/common.cpp/g' \
+        -e 's/([[:space:]]|[ab]\/)examples\/common-ggml\.h/\1examples\/common-ggml.h/g' \
+        -e 's/([[:space:]]|[ab]\/)examples\/common-ggml\.cpp/\1examples\/common-ggml.cpp/g' \
+        -e 's/([[:space:]]|[ab]\/)LICENSE/\1LICENSE/g' \
+        -e 's/([[:space:]]|[ab]\/)scripts\/gen-authors\.sh/\1scripts\/gen-authors.sh/g' \
         > ggml-src.patch.tmp
     mv ggml-src.patch.tmp ggml-src.patch
 
diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last
index 80159b70..eef6768b 100644
--- a/scripts/sync-ggml.last
+++ b/scripts/sync-ggml.last
@@ -1 +1 @@
-e3b3846976c94163f2b3dd128cc959782653edbb
+797faa25af14126eb30134d4033139ae3c5428ed
diff --git a/scripts/sync-ggml.sh b/scripts/sync-ggml.sh
index 402446ef..30a62e08 100755
--- a/scripts/sync-ggml.sh
+++ b/scripts/sync-ggml.sh
@@ -5,9 +5,13 @@ cp -rpv ../ggml/src/CMakeLists.txt   ./ggml/src/CMakeLists.txt
 cp -rpv ../ggml/cmake/FindSIMD.cmake ./ggml/cmake/FindSIMD.cmake
 
 cp -rpv ../ggml/src/ggml.c              ./ggml/src/ggml.c
+cp -rpv ../ggml/src/ggml-aarch64.c      ./ggml/src/ggml-aarch64.c
+cp -rpv ../ggml/src/ggml-aarch64.h      ./ggml/src/ggml-aarch64.h
 cp -rpv ../ggml/src/ggml-alloc.c        ./ggml/src/ggml-alloc.c
 cp -rpv ../ggml/src/ggml-backend-impl.h ./ggml/src/ggml-backend-impl.h
 cp -rpv ../ggml/src/ggml-backend.c      ./ggml/src/ggml-backend.c
+cp -rpv ../ggml/src/ggml-cann/*         ./ggml/src/ggml-cann/
+cp -rpv ../ggml/src/ggml-cann.cpp       ./ggml/src/ggml-cann.cpp
 cp -rpv ../ggml/src/ggml-common.h       ./ggml/src/ggml-common.h
 cp -rpv ../ggml/src/ggml-cuda/*         ./ggml/src/ggml-cuda/
 cp -rpv ../ggml/src/ggml-cuda.cu        ./ggml/src/ggml-cuda.cu
@@ -21,11 +25,13 @@ cp -rpv ../ggml/src/ggml-rpc.cpp        ./ggml/src/ggml-rpc.cpp
 cp -rpv ../ggml/src/ggml-sycl/*         ./ggml/src/ggml-sycl/
 cp -rpv ../ggml/src/ggml-sycl.cpp       ./ggml/src/ggml-sycl.cpp
 cp -rpv ../ggml/src/ggml-vulkan.cpp     ./ggml/src/ggml-vulkan.cpp
+cp -rpv ../ggml/src/vulkan-shaders/*    ./ggml/src/vulkan-shaders/
 
 cp -rpv ../ggml/include/ggml.h         ./ggml/include/ggml.h
 cp -rpv ../ggml/include/ggml-alloc.h   ./ggml/include/ggml-alloc.h
 cp -rpv ../ggml/include/ggml-backend.h ./ggml/include/ggml-backend.h
 cp -rpv ../ggml/include/ggml-blas.h    ./ggml/include/ggml-blas.h
+cp -rpv ../ggml/include/ggml-cann.h    ./ggml/include/ggml-cann.h
 cp -rpv ../ggml/include/ggml-cuda.h    ./ggml/include/ggml-cuda.h
 cp -rpv ../ggml/include/ggml-kompute.h ./ggml/include/ggml-kompute.h
 cp -rpv ../ggml/include/ggml-metal.h   ./ggml/include/ggml-metal.h
diff --git a/src/llama-grammar.cpp b/src/llama-grammar.cpp
index bd9322e2..b123d733 100644
--- a/src/llama-grammar.cpp
+++ b/src/llama-grammar.cpp
@@ -221,7 +221,7 @@ static void llama_grammar_advance_stack(
             // end of alternate (LLAMA_GRETYPE_END, LLAMA_GRETYPE_ALT) or middle of char range
             // (LLAMA_GRETYPE_CHAR_ALT, LLAMA_GRETYPE_CHAR_RNG_UPPER); stack should never be left on
             // those
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
     }
 }
 
@@ -517,7 +517,7 @@ void llama_grammar_accept_token_impl(struct llama_grammar * grammar, const struc
                 return;
             }
         }
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 
     const std::string & piece = vocab->cache_token_to_piece.at(token);
diff --git a/src/llama-impl.h b/src/llama-impl.h
index dcc8c1c1..399b134a 100644
--- a/src/llama-impl.h
+++ b/src/llama-impl.h
@@ -24,3 +24,18 @@ void llama_log_callback_default(ggml_log_level level, const char * text, void *
 #define LLAMA_LOG_INFO(...)  llama_log_internal(GGML_LOG_LEVEL_INFO , __VA_ARGS__)
 #define LLAMA_LOG_WARN(...)  llama_log_internal(GGML_LOG_LEVEL_WARN , __VA_ARGS__)
 #define LLAMA_LOG_ERROR(...) llama_log_internal(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
+
+//
+// helpers
+//
+
+static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
+    if (search.empty()) {
+        return; // Avoid infinite loop if 'search' is an empty string
+    }
+    size_t pos = 0;
+    while ((pos = s.find(search, pos)) != std::string::npos) {
+        s.replace(pos, search.length(), replace);
+        pos += replace.length();
+    }
+}
diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
index c482b368..749f8571 100644
--- a/src/llama-vocab.cpp
+++ b/src/llama-vocab.cpp
@@ -16,20 +16,6 @@
 // helpers
 //
 
-static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
-    std::string result;
-    for (size_t pos = 0; ; pos += search.length()) {
-        auto new_pos = s.find(search, pos);
-        if (new_pos == std::string::npos) {
-            result += s.substr(pos, s.size() - pos);
-            break;
-        }
-        result += s.substr(pos, new_pos - pos) + replace;
-        pos = new_pos;
-    }
-    s = std::move(result);
-}
-
 LLAMA_ATTRIBUTE_FORMAT(1, 2)
 static std::string format(const char * fmt, ...) {
     va_list ap;
@@ -152,14 +138,14 @@ static uint8_t llama_token_to_byte(const llama_vocab & vocab, llama_token id) {
             return strtol(buf.c_str(), NULL, 16);
         }
         case LLAMA_VOCAB_TYPE_BPE: {
-            GGML_ASSERT(false);
-            return unicode_utf8_to_byte(token_data.text); // TODO: why is this here after GGML_ASSERT?
+            GGML_ABORT("fatal error");
+            //return unicode_utf8_to_byte(token_data.text); // TODO: why is this here after GGML_ASSERT?
         }
         case LLAMA_VOCAB_TYPE_WPM: {
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
         default:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
     }
 }
 
@@ -816,6 +802,9 @@ struct llm_tokenizer_ugm {
      * the best tokenization.
     */
     void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
+        // get current size of output (for reversal later)
+        size_t output_size = output.size();
+
         // normalize the input first
         std::string normalized;
         normalize(text, &normalized);
@@ -895,7 +884,7 @@ struct llm_tokenizer_ugm {
         }
 
         // reverse the output since we added tokens starting from the end of the input
-        std::reverse(output.begin(), output.end());
+        std::reverse(output.begin() + output_size, output.end());
     }
 
 private:
@@ -1396,7 +1385,7 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
                 }
             } break;
         case LLAMA_VOCAB_TYPE_NONE:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
     }
 
     return output;
@@ -1422,7 +1411,7 @@ llama_token llama_byte_to_token_impl(const llama_vocab & vocab, uint8_t ch) {
             return vocab.token_to_id.at(unicode_byte_to_utf8(ch));
         }
         default:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
     }
 }
 
@@ -1444,7 +1433,8 @@ llama_token_attr llama_token_get_attr_impl(const struct llama_vocab & vocab, lla
 bool llama_token_is_eog_impl(const struct llama_vocab & vocab, llama_token token) {
     return token != -1 && (
         token == llama_token_eos_impl(vocab) ||
-        token == llama_token_eot_impl(vocab)
+        token == llama_token_eot_impl(vocab) ||
+        token == llama_token_eom_impl(vocab)
     );
 }
 
@@ -1500,6 +1490,10 @@ llama_token llama_token_eot_impl(const struct llama_vocab & vocab) {
     return vocab.special_eot_id;
 }
 
+llama_token llama_token_eom_impl(const struct llama_vocab & vocab) {
+    return vocab.special_eom_id;
+}
+
 int32_t llama_tokenize_impl(
     const struct llama_vocab & vocab,
                   const char * text,
@@ -1606,7 +1600,7 @@ int32_t llama_token_to_piece_impl(const struct llama_vocab & vocab, llama_token
                 break;
             }
             default:
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
         }
     }
 
diff --git a/src/llama-vocab.h b/src/llama-vocab.h
index 30b565d5..7adfc16d 100644
--- a/src/llama-vocab.h
+++ b/src/llama-vocab.h
@@ -45,6 +45,7 @@ struct llama_vocab {
     id special_suffix_id = -1;
     id special_middle_id = -1;
     id special_eot_id    = -1; // TODO: move above after "eos_id", and here add "file separator" token
+    id special_eom_id    = -1;
 
     // tokenizer flags
     bool tokenizer_add_space_prefix = false;
@@ -101,6 +102,7 @@ llama_token llama_token_prefix_impl(const struct llama_vocab & vocab);
 llama_token llama_token_middle_impl(const struct llama_vocab & vocab);
 llama_token llama_token_suffix_impl(const struct llama_vocab & vocab);
 llama_token llama_token_eot_impl   (const struct llama_vocab & vocab);
+llama_token llama_token_eom_impl   (const struct llama_vocab & vocab);
 
 int32_t llama_tokenize_impl(
         const struct llama_vocab & vocab,
diff --git a/src/llama.cpp b/src/llama.cpp
index 71be05f9..dce58dfe 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -102,7 +102,6 @@
 #endif
 
 // bump if necessary
-#define LLAMA_MAX_NODES   8192
 #define LLAMA_MAX_LAYERS  512
 #define LLAMA_MAX_EXPERTS 160  // DeepSeekV2
 
@@ -123,20 +122,6 @@ static std::string trim(const std::string & str) {
     return str.substr(start, end - start);
 }
 
-static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
-    std::string result;
-    for (size_t pos = 0; ; pos += search.length()) {
-        auto new_pos = s.find(search, pos);
-        if (new_pos == std::string::npos) {
-            result += s.substr(pos, s.size() - pos);
-            break;
-        }
-        result += s.substr(pos, new_pos - pos) + replace;
-        pos = new_pos;
-    }
-    s = std::move(result);
-}
-
 static bool is_float_close(float a, float b, float abs_tol) {
     // Check for non-negative tolerance
     if (abs_tol < 0.0) {
@@ -224,6 +209,7 @@ enum llm_arch {
     LLM_ARCH_CHATGLM,
     LLM_ARCH_BITNET,
     LLM_ARCH_T5,
+    LLM_ARCH_T5ENCODER,
     LLM_ARCH_JAIS,
     LLM_ARCH_UNKNOWN,
 };
@@ -268,6 +254,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_CHATGLM,         "chatglm"      },
     { LLM_ARCH_BITNET,          "bitnet"       },
     { LLM_ARCH_T5,              "t5"           },
+    { LLM_ARCH_T5ENCODER,       "t5encoder"    },
     { LLM_ARCH_JAIS,            "jais"         },
     { LLM_ARCH_UNKNOWN,         "(unknown)"    },
 };
@@ -364,6 +351,7 @@ enum llm_kv {
     LLM_KV_TOKENIZER_SUFFIX_ID,
     LLM_KV_TOKENIZER_MIDDLE_ID,
     LLM_KV_TOKENIZER_EOT_ID,
+    LLM_KV_TOKENIZER_EOM_ID,
 
     LLM_KV_ADAPTER_TYPE,
     LLM_KV_ADAPTER_LORA_ALPHA,
@@ -461,6 +449,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_TOKENIZER_SUFFIX_ID,            "tokenizer.ggml.suffix_token_id"          },
     { LLM_KV_TOKENIZER_MIDDLE_ID,            "tokenizer.ggml.middle_token_id"          },
     { LLM_KV_TOKENIZER_EOT_ID,               "tokenizer.ggml.eot_token_id"             },
+    { LLM_KV_TOKENIZER_EOM_ID,               "tokenizer.ggml.eom_token_id"             },
 
     { LLM_KV_ADAPTER_TYPE,                  "adapter.type"       },
     { LLM_KV_ADAPTER_LORA_ALPHA,            "adapter.lora.alpha" },
@@ -1208,6 +1197,21 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
         },
     },
     {
+        LLM_ARCH_CHATGLM,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+        },
+    },
+    {
         LLM_ARCH_BITNET,
         {
             { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
@@ -1261,6 +1265,24 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
         },
     },
     {
+        LLM_ARCH_T5ENCODER,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,           "token_embd" },
+            { LLM_TENSOR_OUTPUT,               "output" },
+            { LLM_TENSOR_ENC_OUTPUT_NORM,      "enc.output_norm" },
+            { LLM_TENSOR_ENC_ATTN_NORM,        "enc.blk.%d.attn_norm" },
+            { LLM_TENSOR_ENC_ATTN_Q,           "enc.blk.%d.attn_q" },
+            { LLM_TENSOR_ENC_ATTN_K,           "enc.blk.%d.attn_k" },
+            { LLM_TENSOR_ENC_ATTN_V,           "enc.blk.%d.attn_v" },
+            { LLM_TENSOR_ENC_ATTN_OUT,         "enc.blk.%d.attn_o" },
+            { LLM_TENSOR_ENC_ATTN_REL_B,       "enc.blk.%d.attn_rel_b" },
+            { LLM_TENSOR_ENC_FFN_NORM,         "enc.blk.%d.ffn_norm" },
+            { LLM_TENSOR_ENC_FFN_GATE,         "enc.blk.%d.ffn_gate" },
+            { LLM_TENSOR_ENC_FFN_DOWN,         "enc.blk.%d.ffn_down" },
+            { LLM_TENSOR_ENC_FFN_UP,           "enc.blk.%d.ffn_up" },
+        },
+    },
+    {
         LLM_ARCH_JAIS,
         {
             { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
@@ -2245,8 +2267,7 @@ struct llama_hparams {
             return n_head_arr[il];
         }
 
-        GGML_ASSERT(false);
-        return 0;
+        GGML_ABORT("fatal error");
     }
 
     uint32_t n_head_kv(uint32_t il = 0) const {
@@ -2254,8 +2275,7 @@ struct llama_hparams {
             return n_head_kv_arr[il];
         }
 
-        GGML_ASSERT(false);
-        return 0;
+        GGML_ABORT("fatal error");
     }
 
     uint32_t n_ff(uint32_t il = 0) const {
@@ -2263,8 +2283,7 @@ struct llama_hparams {
             return n_ff_arr[il];
         }
 
-        GGML_ASSERT(false);
-        return 0;
+        GGML_ABORT("fatal error");
     }
 
     uint32_t n_gqa(uint32_t il = 0) const {
@@ -2441,6 +2460,7 @@ struct llama_layer {
     // long rope factors
     struct ggml_tensor * rope_long  = nullptr;
     struct ggml_tensor * rope_short = nullptr;
+    struct ggml_tensor * rope_freqs = nullptr;
 
     // bitnet scale
     struct ggml_tensor * wq_scale;
@@ -2922,7 +2942,7 @@ static bool llama_kv_cache_init(
 
     // TODO: find a nicer way to add other recurrent model architectures
     cache.recurrent = model.arch == LLM_ARCH_MAMBA;
-    cache.v_trans   = !cparams.flash_attn;
+    cache.v_trans   = !cache.recurrent && !cparams.flash_attn;
 
     cache.head = 0;
     cache.size = kv_size;
@@ -3556,6 +3576,15 @@ namespace GGUFMeta {
 
 using llama_buf_map = std::unordered_map<uint32_t, ggml_backend_buffer_t>;
 
+// TODO: update when needed or think of some clever automatic way to do this
+static size_t llama_model_max_nodes(const llama_model & /*model*/) {
+    //if (model.arch == LLM_ARCH_LLAMA && model.hparams.n_layer > ??) { // llama-3 405B
+    //    return 32768;
+    //}
+
+    return 8192;
+}
+
 struct llama_model_loader {
     int n_kv      = 0;
     int n_tensors = 0;
@@ -4467,14 +4496,14 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_IQ4_K:    return "IQ4_K - 4.5 bpw";
         case LLAMA_FTYPE_MOSTLY_IQ5_K:    return "IQ5_K - 5.5 bpw";
         case LLAMA_FTYPE_MOSTLY_IQ6_K:    return "IQ6_K - 6.6 bpw";
+        case LLAMA_FTYPE_MOSTLY_IQ1_BN:   return "IQ1_BN - 1.625 bpw Bitnet";
+        case LLAMA_FTYPE_MOSTLY_IQ2_BN:   return "IQ2_BN - 2.00 bpw Bitnet";
+        case LLAMA_FTYPE_MOSTLY_IQ2_TN:   return "IQT_BN - 2.06 bpw TriLM";
         case LLAMA_FTYPE_MOSTLY_IQ3_S:    return "IQ3_S - 3.4375 bpw";
         case LLAMA_FTYPE_MOSTLY_IQ3_M:    return "IQ3_S mix - 3.66 bpw";
         case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: return "Q4_0_4_4";
         case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: return "Q4_0_4_8";
         case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: return "Q4_0_8_8";
-        case LLAMA_FTYPE_MOSTLY_IQ1_BN:   return "IQ1_BN - 1.625 bpw Bitnet";
-        case LLAMA_FTYPE_MOSTLY_IQ2_BN:   return "IQ2_BN - 2.00 bpw Bitnet";
-        case LLAMA_FTYPE_MOSTLY_IQ2_TN:   return "IQT_BN - 2.06 bpw TriLM";
 
         default: return "unknown, may not work";
     }
@@ -4898,6 +4927,22 @@ static void llm_load_hparams(
                     case 40: model.type = e_model::MODEL_14B; break;
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
+
+                // for backward compatibility ; see: https://github.com/ggerganov/llama.cpp/pull/8931
+                if ((hparams.n_layer == 32 || hparams.n_layer == 40) && hparams.n_ctx_train == 4096) {
+                    // default value for Phi-3-mini-4k-instruct and Phi-3-medium-4k-instruct
+                    hparams.n_swa = 2047;
+                } else if (hparams.n_layer == 32 && hparams.n_head_kv(0) == 32 && hparams.n_ctx_train == 131072) {
+                    // default value for Phi-3-mini-128k-instruct
+                    hparams.n_swa = 262144;
+                } else if (hparams.n_layer == 40 && hparams.n_ctx_train == 131072) {
+                    // default value for Phi-3-medium-128k-instruct
+                    hparams.n_swa = 131072;
+                }
+                bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
+                if (!found_swa && hparams.n_swa == 0) {
+                    throw std::runtime_error("invalid value for sliding_window");
+                }
             } break;
         case LLM_ARCH_PLAMO:
             {
@@ -4965,6 +5010,7 @@ static void llm_load_hparams(
                 hparams.attn_soft_cap = true;
 
                 switch (hparams.n_layer) {
+                    case 26: model.type = e_model::MODEL_2B; break;
                     case 42: model.type = e_model::MODEL_9B; break;
                     case 46: model.type = e_model::MODEL_27B; break;
                     default: model.type = e_model::MODEL_UNKNOWN;
@@ -5194,6 +5240,12 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                }
             } break;
+        case LLM_ARCH_T5ENCODER:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, hparams.n_rel_attn_bkts);
+                model.type = e_model::MODEL_UNKNOWN;
+            } break;
         case LLM_ARCH_JAIS:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
@@ -5588,6 +5640,7 @@ static void llm_load_vocab(
             { LLM_KV_TOKENIZER_SUFFIX_ID, vocab.special_suffix_id },
             { LLM_KV_TOKENIZER_MIDDLE_ID, vocab.special_middle_id },
             { LLM_KV_TOKENIZER_EOT_ID,    vocab.special_eot_id    },
+            { LLM_KV_TOKENIZER_EOM_ID,    vocab.special_eom_id    },
         };
 
         for (const auto & it : special_token_types) {
@@ -5640,6 +5693,17 @@ static void llm_load_vocab(
                 }
             }
         }
+
+        // find EOM token: "<|eom_id|>"
+        //
+        // TODO: convert scripts should provide this token through the KV metadata LLAMA_KV_TOKENIZER_EOM_ID
+        //       for now, we apply this workaround to find the EOM token based on its text
+        if (vocab.special_eom_id == -1) {
+            const auto & t = vocab.token_to_id.find("<|eom_id|>");
+            if (t != vocab.token_to_id.end()) {
+                vocab.special_eom_id = t->second;
+            }
+        }
     }
 
     // build special tokens cache
@@ -6062,6 +6126,8 @@ static bool llm_load_tensors(
 
                         layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
 
+                        layer.rope_freqs = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FREQS, "weight"), {n_embd/n_head/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+
                         if (n_expert == 0) {
                             layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
                             layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
@@ -7423,6 +7489,42 @@ static bool llm_load_tensors(
                         layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_FFN_UP,   "weight", i), {n_embd,   n_ff});
                     }
                 } break;
+            case LLM_ARCH_T5ENCODER:
+                {
+                    const auto n_rel_attn_bkts = hparams.n_rel_attn_bkts;
+
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm_enc = ml.create_tensor(ctx_output, tn(LLM_TENSOR_ENC_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        // if output is NULL, init from the input tok embed
+                        if (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm_enc  = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ENC_ATTN_NORM,  "weight", i), {n_embd});
+                        layer.attn_rel_b_enc = ml.create_tensor(ctx_input, tn(LLM_TENSOR_ENC_ATTN_REL_B, "weight", i), {n_head, n_rel_attn_bkts}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        layer.wq_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_Q,   "weight", i), {n_embd, n_embd_k_gqa});
+                        layer.wk_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
+                        layer.wv_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
+                        layer.wo_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_OUT, "weight", i), {n_embd_v_gqa, n_embd});
+
+                        layer.ffn_norm_enc = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ENC_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_gate_enc = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ENC_FFN_GATE, "weight", i), {n_embd,   n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.ffn_down_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                        layer.ffn_up_enc   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                    }
+                } break;
             case LLM_ARCH_JAIS:
                 {
                     model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
@@ -8104,7 +8206,7 @@ static struct ggml_tensor * llm_build_moe_ffn(
                 cb(gate, "ffn_moe_gelu", il);
             } break;
         default:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
     }
 
     ggml_tensor * par = ggml_mul(ctx, up, gate); // [n_ff, n_expert_used, n_tokens]
@@ -8431,7 +8533,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_k_shift() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         GGML_ASSERT(kv_self.size == n_ctx);
 
@@ -8462,7 +8564,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_s_copy() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         GGML_ASSERT(kv_self.recurrent);
 
@@ -8485,7 +8587,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_defrag(const std::vector<uint32_t> & ids) {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         for (uint32_t i = 0; i < ids.size(); ++i) {
             const uint32_t id = ids[i];
@@ -8563,6 +8665,10 @@ struct llm_build_context {
         // choose long/short freq factors based on the context size
         const auto n_ctx_pre_seq = cparams.n_ctx / cparams.n_seq_max;
 
+        if (model.layers[il].rope_freqs != nullptr) {
+            return model.layers[il].rope_freqs;
+        }
+
         if (n_ctx_pre_seq > hparams.n_ctx_orig_yarn) {
             return model.layers[il].rope_long;
         }
@@ -8667,8 +8773,8 @@ struct llm_build_context {
                 } break;
             default:
                 {
-                    GGML_ASSERT(false && "unknown pooling type");
-                } break;
+                    GGML_ABORT("unknown pooling type");
+                }
         }
 
         cb(cur, "result_embd_pooled", -1);
@@ -8726,7 +8832,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_llama() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         // mutable variable, needed during the last layer of the computation to skip unused tokens
         int32_t n_tokens = this->n_tokens;
@@ -8757,6 +8863,9 @@ struct llm_build_context {
 
             // self-attention
             {
+                // rope freq factors for llama3; may return nullptr for llama2 and other models
+                struct ggml_tensor * rope_factors = build_rope_factors(il);
+
                 // compute Q and K and RoPE them
                 struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
                 cb(Qcur, "Qcur", il);
@@ -8780,14 +8889,14 @@ struct llm_build_context {
                 }
 
                 Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors,
                     n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Qcur, "Qcur", il);
 
                 Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
                     n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
@@ -8869,7 +8978,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_baichuan() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -8923,7 +9032,7 @@ struct llm_build_context {
                         Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd/n_head, n_head, n_tokens);
                         break;
                     default:
-                        GGML_ASSERT(false);
+                        GGML_ABORT("fatal error");
                 }
                 cb(Qcur, "Qcur", il);
                 cb(Kcur, "Kcur", il);
@@ -8984,7 +9093,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_xverse() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -9087,7 +9196,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_falcon() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
@@ -9207,7 +9316,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_grok() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         // mutable variable, needed during the last layer of the computation to skip unused tokens
         int32_t n_tokens = this->n_tokens;
@@ -9364,7 +9473,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_dbrx() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         // mutable variable, needed during the last layer of the computation to skip unused tokens
         int32_t n_tokens = this->n_tokens;
@@ -9490,7 +9599,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_starcoder() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
@@ -9594,7 +9703,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_refact() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -9688,7 +9797,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_bert() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
@@ -9882,7 +9991,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_bloom() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
@@ -9983,7 +10092,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_mpt() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
@@ -10273,7 +10382,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_qwen() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -10385,7 +10494,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_qwen2() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -10497,7 +10606,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_qwen2moe() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         // mutable variable, needed during the last layer of the computation to skip unused tokens
         int32_t n_tokens = this->n_tokens;
@@ -10643,7 +10752,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_phi2() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
@@ -10764,7 +10873,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_phi3() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
@@ -10996,7 +11105,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_gpt2() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
@@ -11101,7 +11210,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_codeshell() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
@@ -11212,7 +11321,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_orion() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -11330,7 +11439,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_internlm2() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -11451,7 +11560,7 @@ struct llm_build_context {
     //      https://github.com/ggerganov/llama.cpp/issues/5276#issuecomment-1925774738
     // based on the original build_llama() function
     struct ggml_cgraph * build_minicpm() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -11595,7 +11704,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_gemma() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head_k = hparams.n_embd_head_k;
 
@@ -11703,7 +11812,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_gemma2() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head_k = hparams.n_embd_head_k;
 
@@ -11753,9 +11862,10 @@ struct llm_build_context {
 
                 // ref: https://github.com/google/gemma_pytorch/commit/03e657582d17cb5a8617ebf333c1c16f3694670e
                 switch (model.type) {
+                    case e_model::MODEL_2B:
                     case e_model::MODEL_9B:  Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k)));   break;
                     case e_model::MODEL_27B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd / n_head))); break;
-                    default: GGML_ASSERT(false);
+                    default: GGML_ABORT("fatal error");
                 };
                 cb(Qcur, "Qcur_scaled", il);
 
@@ -11838,7 +11948,7 @@ struct llm_build_context {
 
 
     struct ggml_cgraph * build_starcoder2() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -11957,7 +12067,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_mamba() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t d_model = n_embd;
         const int64_t d_conv  = hparams.ssm_d_conv;
@@ -12106,7 +12216,7 @@ struct llm_build_context {
 
     struct ggml_cgraph * build_command_r() {
 
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -12260,7 +12370,7 @@ struct llm_build_context {
     //   * removed bias
     //   * removed MoE
     struct ggml_cgraph * build_olmo() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         // mutable variable, needed during the last layer of the computation to skip unused tokens
         int32_t n_tokens = this->n_tokens;
@@ -12384,7 +12494,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_openelm() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -12509,7 +12619,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_gptneox() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
@@ -12651,7 +12761,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_arctic() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         // mutable variable, needed during the last layer of the computation to skip unused tokens
         int32_t n_tokens = this->n_tokens;
@@ -12783,7 +12893,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_deepseek2() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         // mutable variable, needed during the last layer of the computation to skip unused tokens
         int32_t n_tokens = this->n_tokens;
@@ -13011,7 +13121,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_bitnet() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -13233,8 +13343,8 @@ struct llm_build_context {
         return gf;
     }
 
-    struct ggml_cgraph * build_t5() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+    struct ggml_cgraph * build_t5_encoder() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         // mutable variable, needed during the last layer of the computation to skip unused tokens
         int32_t n_tokens = this->n_tokens;
@@ -13248,310 +13358,330 @@ struct llm_build_context {
 
         inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
 
-        if (lctx.is_encoding) {
-            struct ggml_tensor * pos_bucket_enc = llm_build_pos_bucket(false);
-
-            // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-            struct ggml_tensor * KQ_mask_enc = build_inp_KQ_mask(false);
+        GGML_ASSERT(lctx.is_encoding);
+        struct ggml_tensor * pos_bucket_enc = llm_build_pos_bucket(false);
 
-            for (int il = 0; il < n_layer; ++il) {
-                struct ggml_tensor * inpSA = inpL;
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask_enc = build_inp_KQ_mask(false);
 
-                // norm
-                cur = llm_build_norm(ctx0, inpL, hparams,
-                        model.layers[il].attn_norm_enc, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "attn_norm", il);
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
 
-                // self-attention
-                {
-                    struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq_enc, cur);
-                    cb(Qcur, "Qcur", il);
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm_enc, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
 
-                    struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk_enc, cur);
-                    cb(Kcur, "Kcur", il);
+            // self-attention
+            {
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq_enc, cur);
+                cb(Qcur, "Qcur", il);
 
-                    struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv_enc, cur);
-                    cb(Vcur, "Vcur", il);
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk_enc, cur);
+                cb(Kcur, "Kcur", il);
 
-                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv_enc, cur);
+                cb(Vcur, "Vcur", il);
 
-                    struct ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
-                    struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
 
-                    struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
-                    cb(kq, "kq", il);
+                struct ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
+                struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
 
-                    struct ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b_enc ? model.layers[il].attn_rel_b_enc : model.layers[0].attn_rel_b_enc;
-                    struct ggml_tensor * pos_bias = llm_build_pos_bias(pos_bucket_enc, attn_rel_b);
-                    struct ggml_tensor * kq_b = ggml_add(ctx0, kq, pos_bias);
-                    cb(kq_b, "kq_b", il);
+                struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
+                cb(kq, "kq", il);
 
-                    kq = ggml_soft_max_ext(ctx0, kq_b, KQ_mask_enc, 1.0f, hparams.f_max_alibi_bias);
-                    cb(kq, "kq_soft_max_ext", il);
+                struct ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b_enc ? model.layers[il].attn_rel_b_enc : model.layers[0].attn_rel_b_enc;
+                struct ggml_tensor * pos_bias = llm_build_pos_bias(pos_bucket_enc, attn_rel_b);
+                struct ggml_tensor * kq_b = ggml_add(ctx0, kq, pos_bias);
+                cb(kq_b, "kq_b", il);
 
-                    struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_tokens)));
-                    cb(v, "v", il);
+                kq = ggml_soft_max_ext(ctx0, kq_b, KQ_mask_enc, 1.0f, hparams.f_max_alibi_bias);
+                cb(kq, "kq_soft_max_ext", il);
 
-                    struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_tokens, n_embd_head, n_head_kv), kq);
-                    cb(kqv, "kqv", il);
+                struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_tokens)));
+                cb(v, "v", il);
 
-                    struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
-                    cb(kqv_merged, "kqv_merged", il);
+                struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_tokens, n_embd_head, n_head_kv), kq);
+                cb(kqv, "kqv", il);
 
-                    cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
-                    cb(cur, "kqv_merged_cont", il);
+                struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
+                cb(kqv_merged, "kqv_merged", il);
 
-                    ggml_build_forward_expand(gf, cur);
+                cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
+                cb(cur, "kqv_merged_cont", il);
 
-                    cur = ggml_mul_mat(ctx0, model.layers[il].wo_enc, cur);
-                    cb(cur, "kqv_out", il);
-                }
+                ggml_build_forward_expand(gf, cur);
 
-                if (il == n_layer - 1) {
-                    // skip computing output for unused tokens
-                    struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                    n_tokens = n_outputs;
-                    cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                    inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-                }
+                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo_enc, cur);
+                cb(cur, "kqv_out", il);
+            }
 
-                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-                cb(ffn_inp, "ffn_inp", il);
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
 
-                // feed-forward network
-                {
-                    cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                            model.layers[il].ffn_norm_enc, NULL,
-                            LLM_NORM_RMS, cb, il);
-                    cb(cur, "ffn_norm", il);
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
 
-                    // T5 uses relu, flan-T5 uses gelu-gated
-                    cur = llm_build_ffn(ctx0, lctx, cur,
-                            model.layers[il].ffn_up_enc,   NULL, NULL,
-                            model.layers[il].ffn_gate_enc, NULL, NULL,
-                            model.layers[il].ffn_down_enc, NULL, NULL,
-                            NULL,
-                            model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU,
-                            model.layers[il].ffn_gate_enc ? LLM_FFN_PAR  : LLM_FFN_SEQ,
-                            cb, il);
-                    cb(cur, "ffn_out", il);
-                }
+            // feed-forward network
+            {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm_enc, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
 
-                cur = ggml_add(ctx0, cur, ffn_inp);
+                // T5 uses relu, flan-T5 uses gelu-gated
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up_enc,   NULL, NULL,
+                        model.layers[il].ffn_gate_enc, NULL, NULL,
+                        model.layers[il].ffn_down_enc, NULL, NULL,
+                        NULL,
+                        model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU,
+                        model.layers[il].ffn_gate_enc ? LLM_FFN_PAR  : LLM_FFN_SEQ,
+                        cb, il);
                 cb(cur, "ffn_out", il);
+            }
 
-                ggml_tensor * layer_dir = lctx.cvec.tensor_for(il);
-                if (layer_dir != nullptr) {
-                    cur = ggml_add(ctx0, cur, layer_dir);
-                }
-                cb(cur, "l_out", il);
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
 
-                // input for next layer
-                inpL = cur;
+            ggml_tensor * layer_dir = lctx.cvec.tensor_for(il);
+            if (layer_dir != nullptr) {
+                cur = ggml_add(ctx0, cur, layer_dir);
             }
+            cb(cur, "l_out", il);
 
-            cur = inpL;
-            cb(cur, "result_embd", -1);
+            // input for next layer
+            inpL = cur;
+        }
 
-            cur = llm_build_norm(ctx0, cur, hparams,
-                    model.output_norm_enc, NULL,
-                    LLM_NORM_RMS, cb, -1);
-            cb(cur, "result_norm", -1);
-        } else {
-            GGML_ASSERT(n_outputs_enc > 0 && "call llama_encode() first");
+        cur = inpL;
+        cb(cur, "result_embd", -1);
 
-            struct ggml_tensor * embd_enc       = llm_build_inp_embd_enc();
-            struct ggml_tensor * pos_bucket_dec = llm_build_pos_bucket(true);
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm_enc, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
 
-            struct ggml_tensor * KQ_mask_dec   = build_inp_KQ_mask();
-            struct ggml_tensor * KQ_mask_cross = llm_build_inp_KQ_mask_cross();
+        ggml_build_forward_expand(gf, cur);
 
-            for (int il = 0; il < n_layer; ++il) {
-                struct ggml_tensor * inpSA = inpL;
+        return gf;
+    }
 
-                // norm
-                cur = llm_build_norm(ctx0, inpL, hparams,
-                        model.layers[il].attn_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "attn_norm", il);
+    struct ggml_cgraph * build_t5_decoder() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
-                // self-attention
-                {
-                    struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
-                    cb(Qcur, "Qcur", il);
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
 
-                    struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
-                    cb(Kcur, "Kcur", il);
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
 
-                    struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
-                    cb(Vcur, "Vcur", il);
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
 
-                    llm_build_kv_store(ctx0, hparams, cparams, kv_self, gf, Kcur, Vcur, n_tokens, kv_head, cb, il);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
 
-                    struct ggml_tensor * k =
-                        ggml_view_3d(ctx0, kv_self.k_l[il],
-                                n_embd_head_k, n_kv, n_head_kv,
-                                ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa),
-                                ggml_row_size(kv_self.k_l[il]->type, n_embd_head_k),
-                                0);
-                    cb(k, "k", il);
+        GGML_ASSERT(!lctx.is_encoding);
+        GGML_ASSERT(n_outputs_enc > 0 && "call llama_encode() first");
 
-                    struct ggml_tensor * v =
-                        ggml_view_3d(ctx0, kv_self.v_l[il],
-                                n_kv, n_embd_head_v, n_head_kv,
-                                ggml_element_size(kv_self.v_l[il])*n_ctx,
-                                ggml_element_size(kv_self.v_l[il])*n_ctx*n_embd_head_v,
-                                0);
-                    cb(v, "v", il);
+        struct ggml_tensor * embd_enc       = llm_build_inp_embd_enc();
+        struct ggml_tensor * pos_bucket_dec = llm_build_pos_bucket(true);
 
-                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+        struct ggml_tensor * KQ_mask_dec   = build_inp_KQ_mask();
+        struct ggml_tensor * KQ_mask_cross = llm_build_inp_KQ_mask_cross();
 
-                    struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
 
-                    struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
-                    cb(kq, "kq", il);
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
 
-                    struct ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b ? model.layers[il].attn_rel_b : model.layers[0].attn_rel_b;
-                    struct ggml_tensor * pos_bias = llm_build_pos_bias(pos_bucket_dec, attn_rel_b);
-                    struct ggml_tensor * kq_b = ggml_add(ctx0, kq, pos_bias);
-                    cb(kq_b, "kq_b", il);
+            // self-attention
+            {
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
 
-                    kq = ggml_soft_max_ext(ctx0, kq_b, KQ_mask_dec, 1.0f, hparams.f_max_alibi_bias);
-                    cb(kq, "kq_soft_max_ext", il);
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
 
-                    struct ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq);
-                    cb(kqv, "kqv", il);
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
 
-                    struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
-                    cb(kqv_merged, "kqv_merged", il);
+                llm_build_kv_store(ctx0, hparams, cparams, kv_self, gf, Kcur, Vcur, n_tokens, kv_head, cb, il);
 
-                    cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
-                    cb(cur, "kqv_merged_cont", il);
+                struct ggml_tensor * k =
+                    ggml_view_3d(ctx0, kv_self.k_l[il],
+                            n_embd_head_k, n_kv, n_head_kv,
+                            ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa),
+                            ggml_row_size(kv_self.k_l[il]->type, n_embd_head_k),
+                            0);
+                cb(k, "k", il);
 
-                    ggml_build_forward_expand(gf, cur);
+                struct ggml_tensor * v =
+                    ggml_view_3d(ctx0, kv_self.v_l[il],
+                            n_kv, n_embd_head_v, n_head_kv,
+                            ggml_element_size(kv_self.v_l[il])*n_ctx,
+                            ggml_element_size(kv_self.v_l[il])*n_ctx*n_embd_head_v,
+                            0);
+                cb(v, "v", il);
 
-                    cur = ggml_mul_mat(ctx0, model.layers[il].wo, cur);
-                    cb(cur, "kqv_out", il);
-                }
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
 
-                cur = ggml_add(ctx0, cur, inpSA);
-                cb(cur, "cross_inp", il);
+                struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
 
-                struct ggml_tensor * inpCA = cur;
+                struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
+                cb(kq, "kq", il);
 
-                // norm
-                cur = llm_build_norm(ctx0, cur, hparams,
-                        model.layers[il].attn_norm_cross, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "attn_norm_cross", il);
+                struct ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b ? model.layers[il].attn_rel_b : model.layers[0].attn_rel_b;
+                struct ggml_tensor * pos_bias = llm_build_pos_bias(pos_bucket_dec, attn_rel_b);
+                struct ggml_tensor * kq_b = ggml_add(ctx0, kq, pos_bias);
+                cb(kq_b, "kq_b", il);
 
-                // cross-attention
-                {
-                    struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq_cross, cur);
-                    cb(Qcur, "Qcur", il);
+                kq = ggml_soft_max_ext(ctx0, kq_b, KQ_mask_dec, 1.0f, hparams.f_max_alibi_bias);
+                cb(kq, "kq_soft_max_ext", il);
 
-                    struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk_cross, embd_enc);
-                    cb(Kcur, "Kcur", il);
+                struct ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq);
+                cb(kqv, "kqv", il);
 
-                    struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv_cross, embd_enc);
-                    cb(Vcur, "Vcur", il);
+                struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
+                cb(kqv_merged, "kqv_merged", il);
 
-                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_outputs_enc);
+                cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
+                cb(cur, "kqv_merged_cont", il);
 
-                    struct ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
-                    struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
+                ggml_build_forward_expand(gf, cur);
 
-                    struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
-                    cb(kq, "kq", il);
+                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo, cur);
+                cb(cur, "kqv_out", il);
+            }
 
-                    kq = ggml_soft_max_ext(ctx0, kq, KQ_mask_cross, 1.0f, hparams.f_max_alibi_bias);
-                    cb(kq, "kq_soft_max_ext", il);
+            cur = ggml_add(ctx0, cur, inpSA);
+            cb(cur, "cross_inp", il);
 
-                    struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_outputs_enc)));
-                    cb(v, "v", il);
+            struct ggml_tensor * inpCA = cur;
 
-                    struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_outputs_enc, n_embd_head, n_head_kv), kq);
-                    cb(kqv, "kqv", il);
+            // norm
+            cur = llm_build_norm(ctx0, cur, hparams,
+                    model.layers[il].attn_norm_cross, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm_cross", il);
 
-                    struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
-                    cb(kqv_merged, "kqv_merged", il);
+            // cross-attention
+            {
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq_cross, cur);
+                cb(Qcur, "Qcur", il);
 
-                    cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
-                    cb(cur, "kqv_merged_cont", il);
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk_cross, embd_enc);
+                cb(Kcur, "Kcur", il);
 
-                    ggml_build_forward_expand(gf, cur);
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv_cross, embd_enc);
+                cb(Vcur, "Vcur", il);
 
-                    cur = ggml_mul_mat(ctx0, model.layers[il].wo_cross, cur);
-                    cb(cur, "kqv_out", il);
-                }
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_outputs_enc);
 
-                if (il == n_layer - 1) {
-                    // skip computing output for unused tokens
-                    struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                    n_tokens = n_outputs;
-                    cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                    inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-                    inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids);
-                }
+                struct ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
+                struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
 
-                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpCA);
-                cb(ffn_inp, "ffn_inp", il);
+                struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
+                cb(kq, "kq", il);
 
-                // feed-forward network
-                {
-                    cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                            model.layers[il].ffn_norm, NULL,
-                            LLM_NORM_RMS, cb, il);
-                    cb(cur, "ffn_norm", il);
+                kq = ggml_soft_max_ext(ctx0, kq, KQ_mask_cross, 1.0f, hparams.f_max_alibi_bias);
+                cb(kq, "kq_soft_max_ext", il);
 
-                    // T5 uses relu, flan-T5 uses gelu-gated
-                    cur = llm_build_ffn(ctx0, lctx, cur,
-                            model.layers[il].ffn_up,   NULL, NULL,
-                            model.layers[il].ffn_gate, NULL, NULL,
-                            model.layers[il].ffn_down, NULL, NULL,
-                            NULL,
-                            model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU,
-                            model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ,
-                            cb, il);
-                    cb(cur, "ffn_out", il);
-                }
+                struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_outputs_enc)));
+                cb(v, "v", il);
 
-                cur = ggml_add(ctx0, cur, ffn_inp);
-                cb(cur, "ffn_out", il);
+                struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_outputs_enc, n_embd_head, n_head_kv), kq);
+                cb(kqv, "kqv", il);
 
-                ggml_tensor * layer_dir = lctx.cvec.tensor_for(il);
-                if (layer_dir != nullptr) {
-                    cur = ggml_add(ctx0, cur, layer_dir);
-                }
-                cb(cur, "l_out", il);
+                struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
+                cb(kqv_merged, "kqv_merged", il);
 
-                // input for next layer
-                inpL = cur;
+                cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
+                cb(cur, "kqv_merged_cont", il);
+
+                ggml_build_forward_expand(gf, cur);
+
+                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo_cross, cur);
+                cb(cur, "kqv_out", il);
             }
 
-            cur = inpL;
-            cb(cur, "result_embd", -1);
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+                inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids);
+            }
 
-            cur = llm_build_norm(ctx0, cur, hparams,
-                    model.output_norm, NULL,
-                    LLM_NORM_RMS, cb, -1);
-            cb(cur, "result_norm", -1);
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpCA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                // T5 uses relu, flan-T5 uses gelu-gated
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU,
+                        model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ,
+                        cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            ggml_tensor * layer_dir = lctx.cvec.tensor_for(il);
+            if (layer_dir != nullptr) {
+                cur = ggml_add(ctx0, cur, layer_dir);
+            }
+            cb(cur, "l_out", il);
 
-            // lm_head
-            cur = ggml_mul_mat(ctx0, model.output, cur);
-            cb(cur, "result_output", -1);
+            // input for next layer
+            inpL = cur;
         }
 
+        cur = inpL;
+        cb(cur, "result_embd", -1);
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
         ggml_build_forward_expand(gf, cur);
 
         return gf;
     }
 
     struct ggml_cgraph * build_jais() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
@@ -13643,7 +13773,7 @@ struct llm_build_context {
     }
 
     struct ggml_cgraph * build_chatglm() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
@@ -13996,14 +14126,22 @@ static struct ggml_cgraph * llama_build_graph(
             } break;
         case LLM_ARCH_T5:
             {
-                result = llm.build_t5();
+                if (lctx.is_encoding) {
+                    result = llm.build_t5_encoder();
+                } else {
+                    result = llm.build_t5_decoder();
+                }
+            } break;
+        case LLM_ARCH_T5ENCODER:
+            {
+                result = llm.build_t5_encoder();
             } break;
         case LLM_ARCH_JAIS:
             {
                 result = llm.build_jais();
             } break;
         default:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
     }
 
     // add on pooling layer
@@ -14715,12 +14853,15 @@ static int llama_decode_internal(
             res  = nullptr;
             embd = nullptr;
         } else if (cparams.embeddings) {
-            res = nullptr; // do not extract logits for embedding case
-            embd = gf->nodes[gf->n_nodes - 1];
-            if (strcmp(embd->name, "result_embd_pooled") != 0) {
-                embd = gf->nodes[gf->n_nodes - 2];
+            res  = nullptr; // do not extract logits for embedding case
+            embd = nullptr;
+            for (int i = gf->n_nodes - 1; i >= 0; --i) {
+                if (strcmp(gf->nodes[i]->name, "result_embd_pooled") == 0) {
+                    embd = gf->nodes[i];
+                    break;
+                }
             }
-            GGML_ASSERT(strcmp(embd->name, "result_embd_pooled") == 0 && "missing embeddings tensor");
+            GGML_ASSERT(embd != nullptr && "missing embeddings tensor");
         } else {
             embd = nullptr; // do not extract embeddings when not needed
             GGML_ASSERT(strcmp(res->name, "result_output") == 0 && "missing result_output tensor");
@@ -14802,8 +14943,8 @@ static int llama_decode_internal(
                     } break;
                 case LLAMA_POOLING_TYPE_UNSPECIFIED:
                     {
-                        GGML_ASSERT(false && "unknown pooling type");
-                    } break;
+                        GGML_ABORT("unknown pooling type");
+                    }
             }
         }
         n_outputs_prev += lctx.n_outputs;
@@ -14927,9 +15068,24 @@ static int llama_encode_internal(
     ggml_cgraph * gf = llama_build_graph(lctx, batch, false);
 
     // the output embeddings after the final encoder normalization
-    struct ggml_tensor * embd = gf->nodes[gf->n_nodes - 1];
+    struct ggml_tensor * embd = nullptr;
 
-    GGML_ASSERT(strcmp(embd->name, "result_norm") == 0);
+    // there are two cases here
+    if (llama_model_has_decoder(&lctx.model)) {
+        // first case is an encoder-decoder T5 model where embeddings are passed to decoder
+        embd = gf->nodes[gf->n_nodes - 1];
+        GGML_ASSERT(strcmp(embd->name, "result_norm") == 0 && "missing result_output tensor");
+    } else {
+        // second case is an encoder-only T5 model
+        if (cparams.embeddings) {
+            // only output embeddings if required
+            embd = gf->nodes[gf->n_nodes - 1];
+            if (strcmp(embd->name, "result_embd_pooled") != 0) {
+                embd = gf->nodes[gf->n_nodes - 2];
+            }
+            GGML_ASSERT(strcmp(embd->name, "result_embd_pooled") == 0 && "missing embeddings tensor");
+        }
+    }
 
     ggml_backend_sched_alloc_graph(lctx.sched, gf);
 
@@ -14942,20 +15098,54 @@ static int llama_encode_internal(
         ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(lctx.sched, embd);
         GGML_ASSERT(backend_embd != nullptr);
 
-        // extract token embeddings
-        GGML_ASSERT(lctx.embd != nullptr);
+        if (llama_model_has_decoder(&lctx.model)) {
+            lctx.embd_enc.resize(n_tokens*n_embd);
+            float * embd_out = lctx.embd_enc.data();
 
-        lctx.embd_enc.resize(n_tokens*n_embd);
-        float * embd_out = lctx.embd_enc.data();
+            ggml_backend_tensor_get_async(backend_embd, embd, embd_out, 0, n_tokens*n_embd*sizeof(float));
 
-        ggml_backend_tensor_get_async(backend_embd, embd, embd_out, 0, n_tokens*n_embd*sizeof(float));
+            // remember the sequence ids used during the encoding - needed for cross attention later
+            lctx.seq_ids_enc.resize(n_tokens);
+            for (uint32_t i = 0; i < n_tokens; i++) {
+                for (int s = 0; s < batch.n_seq_id[i]; s++) {
+                    llama_seq_id seq_id = batch.seq_id[i][s];
+                    lctx.seq_ids_enc[i].insert(seq_id);
+                }
+            }
+        } else {
+            GGML_ASSERT(lctx.embd != nullptr);
 
-        // remember the sequence ids used during the encoding - needed for cross attention later
-        lctx.seq_ids_enc.resize(n_tokens);
-        for (uint32_t i = 0; i < n_tokens; i++) {
-            for (int s = 0; s < batch.n_seq_id[i]; s++) {
-                llama_seq_id seq_id = batch.seq_id[i][s];
-                lctx.seq_ids_enc[i].insert(seq_id);
+            switch (cparams.pooling_type) {
+                case LLAMA_POOLING_TYPE_NONE:
+                    {
+                        // extract token embeddings
+                        GGML_ASSERT(lctx.embd != nullptr);
+                        float * embd_out = lctx.embd;
+
+                        GGML_ASSERT(n_tokens*n_embd <= (int64_t) lctx.embd_size);
+                        ggml_backend_tensor_get_async(backend_embd, embd, embd_out, 0, n_tokens*n_embd*sizeof(float));
+                    } break;
+                case LLAMA_POOLING_TYPE_MEAN:
+                case LLAMA_POOLING_TYPE_CLS:
+                case LLAMA_POOLING_TYPE_LAST:
+                    {
+                        // extract sequence embeddings
+                        auto & embd_seq_out = lctx.embd_seq;
+                        embd_seq_out.clear();
+
+                        for (uint32_t i = 0; i < n_tokens; i++) {
+                            const llama_seq_id seq_id = batch.seq_id[i][0];
+                            if (embd_seq_out.find(seq_id) != embd_seq_out.end()) {
+                                continue;
+                            }
+                            embd_seq_out[seq_id].resize(n_embd);
+                            ggml_backend_tensor_get_async(backend_embd, embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float));
+                        }
+                    } break;
+                case LLAMA_POOLING_TYPE_UNSPECIFIED:
+                    {
+                        GGML_ABORT("unknown pooling type");
+                    }
             }
         }
     }
@@ -14988,9 +15178,9 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) {
     // each move requires 6*n_layer tensors (see build_defrag)
     //   - source view, destination view, copy operation
     //   - x2 for keys and values
-    //const uint32_t max_moves = LLAMA_MAX_NODES/(6*n_layer);
+    //const uint32_t max_moves = llama_model_max_nodes(model)/(6*n_layer);
     // TODO: tmp fix https://github.com/ggerganov/llama.cpp/issues/6685#issuecomment-2057579516
-    const uint32_t max_moves = (LLAMA_MAX_NODES - 2*n_layer)/(6*n_layer);
+    const uint32_t max_moves = (llama_model_max_nodes(lctx.model) - 2*n_layer)/(6*n_layer);
 
     // determine which KV cells to move where
     //
@@ -15194,7 +15384,7 @@ static void llama_kv_cache_update_internal(struct llama_context & lctx) {
     // apply K-shift if needed
     if (lctx.model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE && lctx.kv_self.has_shift) {
         if (lctx.model.arch == LLM_ARCH_DEEPSEEK2) { // not supported due to MLA
-            GGML_ASSERT(false && "Deepseek2 does not support K-shift");
+            GGML_ABORT("Deepseek2 does not support K-shift");
         }
 
         {
@@ -15333,7 +15523,7 @@ static void llama_tensor_dequantize_internal(
         } else if (ggml_is_quantized(tensor->type)) {
             qtype.to_float(tensor->data, f32_output, nelements);
         } else {
-            GGML_ASSERT(false); // unreachable
+            GGML_ABORT("fatal error"); // unreachable
         }
         return;
     }
@@ -15391,7 +15581,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
     const int n_expert = std::max(1, (int)qs.model.hparams.n_expert);
     auto layer_info = [n_expert] (int i_layer, int n_layer, const char * name) {
         if (n_expert > 1) {
-            // Believe it or not, "experts" in the FFN of Mixtral-8x7B are not consecutive, but iccasionally randomly
+            // Believe it or not, "experts" in the FFN of Mixtral-8x7B are not consecutive, but occasionally randomly
             // sprinkled in the model. Hence, simply dividing i_ffn_down by n_expert does not work
             // for getting the current layer as I initially thought, and we need to resort to parsing the
             // tensor name.
@@ -16706,6 +16896,8 @@ struct llama_context * llama_new_context_with_model(
 
     ctx->sampling.rng = std::mt19937(params.seed);
     ctx->logits_all   = params.logits_all;
+    // build worst-case graph for encoder if a model contains encoder
+    ctx->is_encoding  = llama_model_has_encoder(model);
 
     uint32_t kv_size = cparams.n_ctx;
     ggml_type type_k = params.type_k;
@@ -16921,8 +17113,10 @@ struct llama_context * llama_new_context_with_model(
                 }
             }
 
+            const size_t max_nodes = llama_model_max_nodes(*model);
+
             // buffer used to store the computation graph and the tensor meta data
-            ctx->buf_compute_meta.resize(ggml_tensor_overhead()*LLAMA_MAX_NODES + ggml_graph_overhead_custom(LLAMA_MAX_NODES, false));
+            ctx->buf_compute_meta.resize(ggml_tensor_overhead()*max_nodes + ggml_graph_overhead_custom(max_nodes, false));
 
             // enabling pipeline parallelism in the scheduler increases memory usage, so it is only done when necessary
             bool pipeline_parallel =
@@ -16935,7 +17129,7 @@ struct llama_context * llama_new_context_with_model(
             // currently this is only implemented in the CUDA backend
             pipeline_parallel = false;
 #endif
-            ctx->sched = ggml_backend_sched_new(ctx->backends.data(), backend_buft.data(), ctx->backends.size(), LLAMA_MAX_NODES, pipeline_parallel);
+            ctx->sched = ggml_backend_sched_new(ctx->backends.data(), backend_buft.data(), ctx->backends.size(), max_nodes, pipeline_parallel);
 
             if (pipeline_parallel) {
                 LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(ctx->sched));
@@ -17018,6 +17212,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_MAMBA:
         case LLM_ARCH_JINA_BERT_V2:
         case LLM_ARCH_T5:
+        case LLM_ARCH_T5ENCODER:
         case LLM_ARCH_JAIS:
             return LLAMA_ROPE_TYPE_NONE;
 
@@ -17060,8 +17255,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
 
         // all model arches should be listed explicitly here
         case LLM_ARCH_UNKNOWN:
-            GGML_ASSERT(false && "unknown architecture");
-            break;
+            GGML_ABORT("unknown architecture");
     }
 
     return LLAMA_ROPE_TYPE_NONE;
@@ -17166,8 +17360,16 @@ struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const ch
 
 bool llama_model_has_encoder(const struct llama_model * model) {
     switch (model->arch) {
-        case LLM_ARCH_T5: return true;
-        default:          return false;
+        case LLM_ARCH_T5:        return true;
+        case LLM_ARCH_T5ENCODER: return true;
+        default:                 return false;
+    }
+}
+
+bool llama_model_has_decoder(const struct llama_model * model) {
+    switch (model->arch) {
+        case LLM_ARCH_T5ENCODER: return false;
+        default:                 return true;
     }
 }
 
@@ -17443,18 +17645,18 @@ void llama_kv_cache_update(struct llama_context * ctx) {
 }
 
 // deprecated
-size_t llama_get_state_size(const struct llama_context * ctx) {
+size_t llama_get_state_size(struct llama_context * ctx) {
     return llama_state_get_size(ctx);
 }
 
 // deprecated
 size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) {
-    return llama_state_get_data(ctx, dst);
+    return llama_state_get_data(ctx, dst, -1);
 }
 
 // deprecated
 size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src) {
-    return llama_state_set_data(ctx, src);
+    return llama_state_set_data(ctx, src, -1);
 }
 
 // deprecated
@@ -17467,302 +17669,282 @@ bool llama_save_session_file(struct llama_context * ctx, const char * path_sessi
     return llama_state_save_file(ctx, path_session, tokens, n_token_count);
 }
 
-// Returns the *maximum* size of the state
-size_t llama_state_get_size(const struct llama_context * ctx) {
-    const auto & cparams = ctx->cparams;
-    const auto & hparams = ctx->model.hparams;
-
-    // we don't know size of rng until we actually serialize it. so reserve more than enough memory for its serialized state.
-    // for reference, std::mt19937(1337) serializes to 6701 bytes.
-    const size_t s_rng_size        = sizeof(size_t);
-    const size_t s_rng             = LLAMA_MAX_RNG_STATE;
-    const size_t s_n_outputs       = sizeof(size_t);
-    // assume worst case for outputs although only currently set ones are serialized
-    const size_t s_output_pos      = ctx->cparams.n_batch * sizeof(int32_t);
-    const size_t s_logits_size     = sizeof(size_t);
-    const size_t s_logits          = ctx->logits_size ? cparams.n_batch * hparams.n_vocab * sizeof(float) : 0;
-    const size_t s_embedding_size  = sizeof(size_t);
-    const size_t s_embedding       = ctx->embd_size   ? cparams.n_batch * hparams.n_embd  * sizeof(float) : 0;
-    const size_t s_kv_buf_size     = sizeof(size_t);
-    const size_t s_kv_head         = sizeof(uint32_t);
-    const size_t s_kv_size         = sizeof(uint32_t);
-    const size_t s_kv_used         = sizeof(uint32_t);
-    const size_t s_v_trans         = sizeof(uint32_t);
-    const size_t s_kv              = ctx->kv_self.total_size();
-    const size_t s_kv_cell         = sizeof(llama_pos) + sizeof(size_t) + cparams.n_seq_max*sizeof(llama_seq_id);
-    const size_t s_kv_cells        = ctx->kv_self.size * s_kv_cell;
-
-    const size_t s_total = (
-        + s_rng_size
-        + s_rng
-        + s_n_outputs
-        + s_output_pos
-        + s_logits_size
-        + s_logits
-        + s_embedding_size
-        + s_embedding
-        + s_kv_buf_size
-        + s_kv_head
-        + s_kv_size
-        + s_kv_used
-        + s_v_trans
-        + s_kv
-        + s_kv_cells
-    );
-
-    // on session change it is very likely that the state size has changed - so we need to update this function
-    static_assert(LLAMA_SESSION_VERSION == 7, "So you just bumped the session version - good. But did you remember to update llama_state_get_size?");
-
-    return s_total;
-}
-
-// llama_context_data
-struct llama_data_context {
+// TODO: replace all non-fatal assertions with returned errors or exceptions
+struct llama_data_write {
     virtual void write(const void * src, size_t size) = 0;
+    virtual void write_tensor_data(const struct ggml_tensor * tensor, size_t offset, size_t size) = 0;
     virtual size_t get_size_written() = 0;
-    virtual ~llama_data_context() = default;
-};
-
-struct llama_data_buffer_context : llama_data_context {
-    uint8_t * ptr;
-    size_t size_written = 0;
+    virtual ~llama_data_write() = default;
 
-    llama_data_buffer_context(uint8_t * p) : ptr(p) {}
+    void write_string(const std::string & str) {
+        uint32_t str_size = str.size();
 
-    void write(const void * src, size_t size) override {
-        memcpy(ptr, src, size);
-        ptr += size;
-        size_written += size;
+        write(&str_size,  sizeof(str_size));
+        write(str.data(), str_size);
     }
 
-    size_t get_size_written() override {
-        return size_written;
+    void write_model_info(const struct llama_context * ctx) {
+        std::string arch_str = LLM_ARCH_NAMES.at(ctx->model.arch);
+        write_string(arch_str);
+        // TODO: add more model-specific info which should prevent loading the session file if not identical
     }
-};
 
-struct llama_data_file_context : llama_data_context {
-    llama_file * file;
-    size_t size_written = 0;
+    void write_rng(const std::mt19937 & rng) {
+        std::ostringstream rng_ss;
+        rng_ss << rng;
 
-    llama_data_file_context(llama_file * f) : file(f) {}
+        const std::string & rng_str = rng_ss.str();
 
-    void write(const void * src, size_t size) override {
-        file->write_raw(src, size);
-        size_written += size;
+        write_string(rng_str);
     }
 
-    size_t get_size_written() override {
-        return size_written;
-    }
-};
+    void write_output_ids(const struct llama_context * ctx) {
+        const uint32_t n_outputs = ctx->n_outputs;
 
-/** copy state data into either a buffer or file depending on the passed in context
- *
- * file context:
- * llama_file file("/path", "wb");
- * llama_data_file_context data_ctx(&file);
- * llama_state_get_data(ctx, &data_ctx);
- *
- * buffer context:
- * std::vector<uint8_t> buf(max_size, 0);
- * llama_data_buffer_context data_ctx(&buf.data());
- * llama_state_get_data(ctx, &data_ctx);
- *
-*/
-static void llama_state_get_data_internal(struct llama_context * ctx, llama_data_context * data_ctx) {
-    llama_synchronize(ctx);
+        std::vector<int32_t> output_pos;
 
-    // copy rng
-    {
-        std::ostringstream rng_ss;
-        rng_ss << ctx->sampling.rng;
+        const size_t    n_batch = ctx->cparams.n_batch;
+        const auto & output_ids = ctx->output_ids;
 
-        const std::string & rng_str  = rng_ss.str();
-        const size_t        rng_size = rng_str.size();
+        GGML_ASSERT(n_outputs <= ctx->output_size);
 
-        GGML_ASSERT(rng_size <= LLAMA_MAX_RNG_STATE);
+        output_pos.resize(n_outputs);
 
-        data_ctx->write(&rng_size,      sizeof(rng_size));
-        data_ctx->write(rng_str.data(), rng_size);
+        // build a more compact representation of the output ids
+        for (size_t i = 0; i < n_batch; ++i) {
+            // map an output id to a position in the batch
+            int32_t pos = output_ids[i];
+            if (pos >= 0) {
+                GGML_ASSERT((uint32_t) pos < n_outputs);
+                output_pos[pos] = i;
+            }
+        }
+
+        write(&n_outputs, sizeof(n_outputs));
+
+        if (n_outputs) {
+            write(output_pos.data(), n_outputs * sizeof(int32_t));
+        }
     }
 
-    // copy outputs
-    {
-        // Can't use ctx->n_outputs because it's not for the
-        // entire last batch when n_ubatch is smaller than n_batch
-        size_t n_outputs = 0;
+    void write_logits(const struct llama_context * ctx) {
+        const uint64_t logits_size = std::min((uint64_t) ctx->logits_size, (uint64_t) ctx->n_outputs * ctx->model.hparams.n_vocab);
 
-        // copy output ids
-        {
-            std::vector<int32_t> output_pos;
+        write(&logits_size, sizeof(logits_size));
+
+        if (logits_size) {
+            write(ctx->logits, logits_size * sizeof(float));
+        }
+    }
+
+    void write_embeddings(const struct llama_context * ctx) {
+        const uint64_t embeddings_size = std::min((uint64_t) ctx->embd_size, (uint64_t) ctx->n_outputs * ctx->model.hparams.n_embd);
+
+        write(&embeddings_size, sizeof(embeddings_size));
+
+        if (embeddings_size) {
+            write(ctx->embd, embeddings_size * sizeof(float));
+        }
+    }
+
+    void write_kv_cache_meta(const llama_kv_cache & kv_self, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges, llama_seq_id seq_id = -1) {
 
-            const size_t    n_batch = ctx->cparams.n_batch;
-            const auto & output_ids = ctx->output_ids;
+        for (const auto & range : cell_ranges) {
+            for (uint32_t i = range.first; i < range.second; ++i) {
+                const auto & cell = kv_self.cells[i];
+                const llama_pos pos      = cell.pos;
+                const uint32_t  n_seq_id = seq_id == -1 ? cell.seq_id.size() : 0;
 
-            output_pos.resize(ctx->output_size);
+                write(&pos,      sizeof(pos));
+                write(&n_seq_id, sizeof(n_seq_id));
 
-            // build a more compact representation of the output ids
-            for (size_t i = 0; i < n_batch; ++i) {
-                // map an output id to a position in the batch
-                int32_t pos = output_ids[i];
-                if (pos >= 0) {
-                    if ((size_t) pos >= n_outputs) {
-                        n_outputs = pos + 1;
+                if (n_seq_id) {
+                    for (auto seq_id : cell.seq_id) {
+                        write(&seq_id, sizeof(seq_id));
                     }
-                    GGML_ASSERT((size_t) pos < ctx->output_size);
-                    output_pos[pos] = i;
                 }
             }
+        }
+    }
 
-            data_ctx->write(&n_outputs, sizeof(n_outputs));
+    void write_kv_cache_data(const struct llama_context * ctx, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges) {
+        const struct llama_kv_cache & kv_self = ctx->kv_self;
+        const struct llama_hparams & hparams = ctx->model.hparams;
 
-            if (n_outputs) {
-                data_ctx->write(output_pos.data(), n_outputs * sizeof(int32_t));
-            }
-        }
+        const uint32_t v_trans = kv_self.v_trans ? 1 : 0;
+        const uint32_t n_layer = hparams.n_layer;
 
-        // copy logits
-        {
-            const size_t logits_size = std::min(ctx->logits_size, n_outputs * ctx->model.hparams.n_vocab);
+        write(&v_trans, sizeof(v_trans));
+        write(&n_layer, sizeof(n_layer));
 
-            data_ctx->write(&logits_size, sizeof(logits_size));
+        std::vector<uint8_t> tmp_buf;
 
-            if (logits_size) {
-                data_ctx->write(ctx->logits, logits_size * sizeof(float));
-            }
-        }
+        // Iterate and write all the keys first, each row is a cell
+        // Get whole range at a time
+        for (uint32_t il = 0; il < n_layer; ++il) {
+            const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
 
-        // copy embeddings
-        {
-            const size_t embeddings_size = std::min(ctx->embd_size, n_outputs * ctx->model.hparams.n_embd);
+            // Write key type
+            const int32_t k_type_i = (int32_t)kv_self.k_l[il]->type;
+            write(&k_type_i, sizeof(k_type_i));
 
-            data_ctx->write(&embeddings_size, sizeof(embeddings_size));
+            // Write row size of key
+            const uint64_t k_size_row = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa);
+            write(&k_size_row, sizeof(k_size_row));
 
-            if (embeddings_size) {
-                data_ctx->write(ctx->embd, embeddings_size * sizeof(float));
+            // Read each range of cells of k_size length each into tmp_buf and write out
+            for (const auto & range : cell_ranges) {
+                const size_t range_size = range.second - range.first;
+                const size_t buf_size = range_size * k_size_row;
+                write_tensor_data(kv_self.k_l[il], range.first * k_size_row, buf_size);
             }
         }
-    }
 
-    // copy kv cache
-    {
-        const auto & kv_self = ctx->kv_self;
-        const auto & hparams = ctx->model.hparams;
-
-        const uint32_t n_layer      = hparams.n_layer;
-
-        // NOTE: kv_size and kv_buf_size are mostly used for sanity checks
-        const uint32_t kv_head     = llama_kv_cache_cell_max(kv_self);
-        const uint32_t kv_size     = kv_self.size;
-        const size_t   kv_buf_size = kv_self.total_size() / (kv_size ? kv_size : 1) * kv_head;
-        const uint32_t kv_used     = kv_self.used;
-        const uint32_t v_trans     = kv_self.v_trans ? 1 : 0;
-
-        data_ctx->write(&kv_buf_size, sizeof(kv_buf_size));
-        data_ctx->write(&kv_head,     sizeof(kv_head));
-        data_ctx->write(&kv_size,     sizeof(kv_size));
-        data_ctx->write(&kv_used,     sizeof(kv_used));
-        data_ctx->write(&v_trans,     sizeof(v_trans));
-
-        if (kv_buf_size) {
-            const size_t pre_kv_buf_size = data_ctx->get_size_written();
-
-            std::vector<uint8_t> tmp_buf;
-            for (int il = 0; il < (int) n_layer; ++il) {
-                const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
+        if (!kv_self.v_trans) {
+            for (uint32_t il = 0; il < n_layer; ++il) {
                 const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
 
-                const size_t k_size = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*kv_head);
-
-                tmp_buf.resize(k_size);
-                ggml_backend_tensor_get(kv_self.k_l[il], tmp_buf.data(), 0, tmp_buf.size());
-                data_ctx->write(tmp_buf.data(), tmp_buf.size());
+                // Write value type
+                const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type;
+                write(&v_type_i, sizeof(v_type_i));
 
-                if (kv_self.recurrent || !kv_self.v_trans) {
-                    // v is contiguous for recurrent models
-                    // TODO: use other tensors for state models than k and v
-                    const size_t v_size = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa*kv_head);
+                // Write row size of value
+                const uint64_t v_size_row = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa);
+                write(&v_size_row, sizeof(v_size_row));
 
-                    tmp_buf.resize(v_size);
-                    ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), 0, tmp_buf.size());
-                    data_ctx->write(tmp_buf.data(), tmp_buf.size());
-                    continue;
+                // Read each range of cells of v_size length each into tmp_buf and write out
+                for (const auto & range : cell_ranges) {
+                    const size_t range_size = range.second - range.first;
+                    const size_t buf_size = range_size * v_size_row;
+                    write_tensor_data(kv_self.v_l[il], range.first * v_size_row, buf_size);
                 }
+            }
+        } else {
+            // When v is transposed, we also need the element size and get the element ranges from each row
+            const uint32_t kv_size = kv_self.size;
+            for (uint32_t il = 0; il < n_layer; ++il) {
+                const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+
+                // Write value type
+                const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type;
+                write(&v_type_i, sizeof(v_type_i));
 
-                // v is not contiguous, copy row by row
-                const size_t v_row_size   = ggml_row_size(kv_self.v_l[il]->type, kv_head);
-                const size_t v_row_stride = ggml_row_size(kv_self.v_l[il]->type, kv_size);
+                // Write element size
+                const uint32_t v_size_el = ggml_type_size(kv_self.v_l[il]->type);
+                write(&v_size_el, sizeof(v_size_el));
 
-                tmp_buf.resize(v_row_size);
-                for (int ir = 0; ir < (int) n_embd_v_gqa; ++ir) {
-                    ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), ir*v_row_stride, tmp_buf.size());
-                    data_ctx->write(tmp_buf.data(), tmp_buf.size());
+                // Write GQA embedding size
+                write(&n_embd_v_gqa, sizeof(n_embd_v_gqa));
+
+                // For each row, we get the element values of each cell
+                for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
+                    // Read each range of cells of v_size_el length each into tmp_buf and write out
+                    for (const auto & range : cell_ranges) {
+                        const size_t range_size = range.second - range.first;
+                        const size_t src_offset = (range.first + j * kv_size) * v_size_el;
+                        const size_t buf_size = range_size * v_size_el;
+                        write_tensor_data(kv_self.v_l[il], src_offset, buf_size);
+                    }
                 }
             }
-            GGML_ASSERT(kv_buf_size == data_ctx->get_size_written() - pre_kv_buf_size);
         }
+    }
 
-        for (uint32_t i = 0; i < kv_head; ++i) {
-            const auto & cell = kv_self.cells[i];
-
-            const llama_pos pos         = cell.pos;
-            const size_t    seq_id_size = cell.seq_id.size();
-
-            data_ctx->write(&pos,         sizeof(pos));
-            data_ctx->write(&seq_id_size, sizeof(seq_id_size));
+    void write_kv_cache(const struct llama_context * ctx, llama_seq_id seq_id = -1) {
+        const struct llama_kv_cache & kv_self = ctx->kv_self;
+        std::vector<std::pair<uint32_t, uint32_t>> cell_ranges; // ranges, from inclusive, to exclusive
+        uint32_t cell_count = 0;
 
-            for (auto seq_id : cell.seq_id) {
-                data_ctx->write(&seq_id, sizeof(seq_id));
+        // Count the number of cells with the specified seq_id
+        // Find all the ranges of cells with this seq id (or all, when -1)
+        uint32_t cell_range_begin = kv_self.size;
+        for (uint32_t i = 0; i < kv_self.size; ++i) {
+            const auto & cell = kv_self.cells[i];
+            if ((seq_id == -1 && !cell.is_empty()) || cell.has_seq_id(seq_id)) {
+                ++cell_count;
+                if (cell_range_begin == kv_self.size) {
+                    cell_range_begin = i;
+                }
+            } else {
+                if (cell_range_begin != kv_self.size) {
+                    cell_ranges.emplace_back(cell_range_begin, i);
+                    cell_range_begin = kv_self.size;
+                }
             }
         }
-    }
-}
+        if (cell_range_begin != kv_self.size) {
+            cell_ranges.emplace_back(cell_range_begin, kv_self.size);
+        }
 
-size_t llama_state_get_data(struct llama_context * ctx, uint8_t * dst) {
-    llama_data_buffer_context data_ctx(dst);
-    llama_state_get_data_internal(ctx, &data_ctx);
+        // DEBUG CHECK: Sum of cell counts in ranges should equal the total cell count
+        uint32_t cell_count_check = 0;
+        for (const auto & range : cell_ranges) {
+            cell_count_check += range.second - range.first;
+        }
+        GGML_ASSERT(cell_count == cell_count_check);
 
-    return data_ctx.get_size_written();
-}
+        write(&cell_count, sizeof(cell_count));
 
-// Sets the state reading from the specified source address
-size_t llama_state_set_data(struct llama_context * ctx, const uint8_t * src) {
-    llama_synchronize(ctx);
+        write_kv_cache_meta(kv_self, cell_ranges, seq_id);
+        write_kv_cache_data(ctx, cell_ranges);
+    }
+};
 
-    const uint8_t * inp = src;
+struct llama_data_read {
+    virtual const uint8_t * read(size_t size) = 0;
+    virtual void read_to(void * dst, size_t size) = 0;
+    virtual size_t get_size_read() = 0;
+    virtual ~llama_data_read() = default;
 
-    // set rng
-    {
-        size_t rng_size;
-        memcpy(&rng_size, inp, sizeof(rng_size)); inp += sizeof(rng_size);
+    void read_string(std::string & str) {
+        uint32_t str_size;
+        read_to(&str_size, sizeof(str_size));
+
+        str.assign((const char *) read(str_size), str_size);
+    }
 
-        GGML_ASSERT(rng_size <= LLAMA_MAX_RNG_STATE);
+    // validate model information
+    void read_model_info(const struct llama_context * ctx) {
+        std::string cur_arch_str = LLM_ARCH_NAMES.at(ctx->model.arch);
+        std::string arch_str;
+        read_string(arch_str);
+        if (cur_arch_str != arch_str) {
+            throw std::runtime_error(format("wrong model arch: '%s' instead of '%s'", arch_str.c_str(), cur_arch_str.c_str()));
+        }
+        // TODO: add more info which needs to be identical but which is not verified otherwise
+    }
 
-        std::string rng_str((const char *)inp, rng_size); inp += rng_size;
+    void read_rng(std::mt19937 & rng) {
+        std::string rng_str;
+        read_string(rng_str);
 
         std::istringstream rng_ss(rng_str);
-        rng_ss >> ctx->sampling.rng;
+        rng_ss >> rng;
 
-        GGML_ASSERT(!rng_ss.fail());
+        if (rng_ss.fail()) {
+            throw std::runtime_error("failed to load RNG state");
+        }
     }
 
-    // set output ids
-    {
-        size_t n_outputs;
+    void read_output_ids(struct llama_context * ctx) {
         std::vector<int32_t> output_pos;
 
-        memcpy(&n_outputs, inp, sizeof(n_outputs)); inp += sizeof(n_outputs);
+        uint32_t n_outputs;
+        read_to(&n_outputs, sizeof(n_outputs));
 
-        GGML_ASSERT(n_outputs <= llama_output_reserve(*ctx, n_outputs));
+        if (n_outputs > llama_output_reserve(*ctx, n_outputs)) {
+            throw std::runtime_error("could not reserve outputs");
+        }
 
         if (n_outputs) {
             output_pos.resize(n_outputs);
-            memcpy(output_pos.data(), inp, n_outputs * sizeof(int32_t));
-            inp += n_outputs * sizeof(int32_t);
+            read_to(output_pos.data(), n_outputs * sizeof(int32_t));
 
             for (int32_t i = 0; i < (int32_t) output_pos.size(); ++i) {
                 int32_t id = output_pos[i];
-                GGML_ASSERT((uint32_t) id < ctx->cparams.n_batch);
+                if ((uint32_t) id >= ctx->cparams.n_batch) {
+                    throw std::runtime_error(format("invalid output id, %d does not fit in batch size of %u", id, ctx->cparams.n_batch));
+                }
                 ctx->output_ids[id] = i;
             }
 
@@ -17770,611 +17952,571 @@ size_t llama_state_set_data(struct llama_context * ctx, const uint8_t * src) {
         }
     }
 
-    // set logits
-    {
-        size_t logits_size;
+    void read_logits(struct llama_context * ctx) {
+        uint64_t logits_size;
+        read_to(&logits_size, sizeof(logits_size));
 
-        memcpy(&logits_size, inp, sizeof(logits_size)); inp += sizeof(logits_size);
-
-        GGML_ASSERT(ctx->logits_size >= logits_size);
+        if (ctx->logits_size < logits_size) {
+            throw std::runtime_error("logits buffer too small");
+        }
 
         if (logits_size) {
-            memcpy(ctx->logits, inp, logits_size * sizeof(float));
-            inp += logits_size * sizeof(float);
+            read_to(ctx->logits, logits_size * sizeof(float));
         }
     }
 
-    // set embeddings
-    {
-        size_t embeddings_size;
-
-        memcpy(&embeddings_size, inp, sizeof(embeddings_size)); inp += sizeof(embeddings_size);
+    void read_embeddings(struct llama_context * ctx) {
+        uint64_t embeddings_size;
+        read_to(&embeddings_size, sizeof(embeddings_size));
 
-        GGML_ASSERT(ctx->embd_size >= embeddings_size);
+        if (ctx->embd_size < embeddings_size) {
+            throw std::runtime_error("embeddings buffer too small");
+        }
 
         if (embeddings_size) {
-            memcpy(ctx->embd, inp, embeddings_size * sizeof(float));
-            inp += embeddings_size * sizeof(float);
+            read_to(ctx->embd, embeddings_size * sizeof(float));
         }
     }
 
-    // set kv cache
-    {
-        const auto & kv_self = ctx->kv_self;
-        const auto & hparams = ctx->model.hparams;
+    bool read_kv_cache_meta(struct llama_context * ctx, uint32_t cell_count, llama_seq_id dest_seq_id = -1) {
+        struct llama_kv_cache & kv_self = ctx->kv_self;
 
-        const uint32_t n_layer      = hparams.n_layer;
+        if (dest_seq_id != -1) {
+            // single sequence
 
-        size_t   kv_buf_size;
-        uint32_t kv_head;
-        uint32_t kv_size;
-        uint32_t kv_used;
-        uint32_t v_trans;
+            llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1);
 
-        memcpy(&kv_buf_size, inp, sizeof(kv_buf_size)); inp += sizeof(kv_buf_size);
-        memcpy(&kv_head,     inp, sizeof(kv_head));     inp += sizeof(kv_head);
-        memcpy(&kv_size,     inp, sizeof(kv_size));     inp += sizeof(kv_size);
-        memcpy(&kv_used,     inp, sizeof(kv_used));     inp += sizeof(kv_used);
-        memcpy(&v_trans,     inp, sizeof(v_trans));     inp += sizeof(v_trans);
+            llama_batch batch = llama_batch_init(cell_count, 0, 1);
+            batch.n_tokens = cell_count;
+            for (uint32_t i = 0; i < cell_count; ++i) {
+                llama_pos pos;
+                uint32_t n_seq_id;
 
-        GGML_ASSERT(kv_self.v_trans == (bool) v_trans); // incompatible V transposition
+                read_to(&pos, sizeof(pos));
+                read_to(&n_seq_id, sizeof(n_seq_id));
 
-        if (kv_self.size != kv_size) {
-            // the KV cache needs to be big enough to load all the KV cells from the saved state
-            GGML_ASSERT(kv_self.size >= kv_head);
+                if (n_seq_id != 0) {
+                    LLAMA_LOG_ERROR("%s: invalid seq_id-agnostic kv cell\n", __func__);
+                    return false;
+                }
 
-            LLAMA_LOG_INFO("%s: state contains %d KV cells, was saved with kv_size=%d, but is loaded with kv_size=%d (fine, but different)\n",
-                __func__, kv_head, kv_size, kv_self.size);
-        }
+                batch.pos[i] = pos;
+                batch.n_seq_id[i] = 1;
+                batch.seq_id[i][0] = dest_seq_id;
+            }
+            if (!llama_kv_cache_find_slot(kv_self, batch)) {
+                llama_batch_free(batch);
+                LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__);
+                return false;
+            }
 
-        llama_kv_cache_clear(ctx);
+            // DEBUG CHECK: kv_self.head should be our first cell, kv_self.head + cell_count - 1 should be our last cell (verify seq_id and pos values)
+            // Assume that this is one contiguous block of cells
+            GGML_ASSERT(kv_self.head + cell_count <= kv_self.size);
+            GGML_ASSERT(kv_self.cells[kv_self.head].pos == batch.pos[0]);
+            GGML_ASSERT(kv_self.cells[kv_self.head + cell_count - 1].pos == batch.pos[cell_count - 1]);
+            GGML_ASSERT(kv_self.cells[kv_self.head].has_seq_id(dest_seq_id));
+            GGML_ASSERT(kv_self.cells[kv_self.head + cell_count - 1].has_seq_id(dest_seq_id));
 
-        if (kv_buf_size) {
-            const size_t pre_kv_buf_size = inp - src;
+            // Cleanup
+            llama_batch_free(batch);
+        } else {
+            // whole KV cache restore
 
-            GGML_ASSERT(kv_self.total_size() >= kv_buf_size);
+            if (cell_count > kv_self.size) {
+                LLAMA_LOG_ERROR("%s: not enough cells in kv cache\n", __func__);
+                return false;
+            }
 
-            for (int il = 0; il < (int) n_layer; ++il) {
-                const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
-                const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+            llama_kv_cache_clear(kv_self);
 
-                const size_t k_size = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*kv_head);
+            for (uint32_t i = 0; i < cell_count; ++i) {
+                llama_kv_cell & cell = kv_self.cells[i];
 
-                ggml_backend_tensor_set(kv_self.k_l[il], inp, 0, k_size);
-                inp += k_size;
+                llama_pos pos;
+                uint32_t  n_seq_id;
 
-                if (kv_self.recurrent || !kv_self.v_trans) {
-                    // v is contiguous for recurrent models
-                    // TODO: use other tensors for state models than k and v
-                    const size_t v_size = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa*kv_head);
+                read_to(&pos,      sizeof(pos));
+                read_to(&n_seq_id, sizeof(n_seq_id));
 
-                    ggml_backend_tensor_set(kv_self.v_l[il], inp, 0, v_size);
-                    inp += v_size;
-                    continue;
-                }
+                cell.pos = pos;
+
+                for (uint32_t j = 0; j < n_seq_id; ++j) {
+                    llama_seq_id seq_id;
+                    read_to(&seq_id, sizeof(seq_id));
 
-                // v is not contiguous, copy row by row
-                const size_t v_row_size   = ggml_row_size(kv_self.v_l[il]->type, kv_head);
-                const size_t v_row_stride = ggml_row_size(kv_self.v_l[il]->type, kv_self.size);
+                    if (seq_id < 0 || (uint32_t) seq_id >= llama_n_seq_max(ctx)) {
+                        LLAMA_LOG_ERROR("%s: invalid seq_id, %d is out of range [0, %u)\n", __func__, seq_id, llama_n_seq_max(ctx));
+                        return false;
+                    }
 
-                for (int ir = 0; ir < (int) n_embd_v_gqa; ++ir) {
-                    ggml_backend_tensor_set(kv_self.v_l[il], inp, ir*v_row_stride, v_row_size);
-                    inp += v_row_size;
+                    cell.seq_id.insert(seq_id);
                 }
             }
-            GGML_ASSERT(kv_buf_size == inp - src - pre_kv_buf_size);
+
+            kv_self.head = 0;
+            kv_self.used = cell_count;
         }
 
-        ctx->kv_self.head = kv_head;
-        ctx->kv_self.used = kv_used;
+        return true;
+    }
 
-        for (uint32_t i = 0; i < kv_head; ++i) {
-            llama_pos pos;
-            size_t    seq_id_size;
+    bool read_kv_cache_data(struct llama_context * ctx, uint32_t cell_count) {
+        const struct llama_hparams & hparams = ctx->model.hparams;
+        struct llama_kv_cache & kv_self = ctx->kv_self;
+        uint32_t v_trans;
+        uint32_t n_layer;
+        read_to(&v_trans, sizeof(v_trans));
+        read_to(&n_layer, sizeof(n_layer));
 
-            memcpy(&pos,         inp, sizeof(pos));         inp += sizeof(pos);
-            memcpy(&seq_id_size, inp, sizeof(seq_id_size)); inp += sizeof(seq_id_size);
+        if (n_layer != hparams.n_layer) {
+            LLAMA_LOG_ERROR("%s: mismatched layer count (%u instead of %u)\n", __func__, n_layer, hparams.n_layer);
+            return false;
+        }
+        if (cell_count > kv_self.size) {
+            LLAMA_LOG_ERROR("%s: not enough cells in kv cache to restore state (%u > %u)\n", __func__, cell_count, kv_self.size);
+            return false;
+        }
+        if (kv_self.v_trans != (bool) v_trans) {
+            LLAMA_LOG_ERROR("%s: incompatible V transposition\n", __func__);
+            return false;
+        }
 
-            ctx->kv_self.cells[i].pos = pos;
+        // For each layer, read the keys for each cell, one row is one cell, read as one contiguous block
+        for (uint32_t il = 0; il < n_layer; ++il) {
+            const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
 
-            llama_seq_id seq_id;
+            // Read type of key
+            int32_t k_type_i_ref;
+            read_to(&k_type_i_ref, sizeof(k_type_i_ref));
+            const int32_t k_type_i = (int32_t)kv_self.k_l[il]->type;
+            if (k_type_i != k_type_i_ref) {
+                LLAMA_LOG_ERROR("%s: mismatched key type (%d != %d, layer %d)\n", __func__, k_type_i, k_type_i_ref, il);
+                return false;
+            }
+
+            // Read row size of key
+            uint64_t k_size_row_ref;
+            read_to(&k_size_row_ref, sizeof(k_size_row_ref));
+            const size_t k_size_row = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa);
+            if (k_size_row != k_size_row_ref) {
+                LLAMA_LOG_ERROR("%s: mismatched key row size (%zu != %zu, layer %d)\n", __func__, k_size_row, (size_t) k_size_row_ref, il);
+                return false;
+            }
 
-            for (size_t j = 0; j < seq_id_size; ++j) {
-                memcpy(&seq_id, inp, sizeof(seq_id)); inp += sizeof(seq_id);
-                ctx->kv_self.cells[i].seq_id.insert(seq_id);
+            if (cell_count) {
+                // Read and set the keys for the whole cell range
+                ggml_backend_tensor_set(kv_self.k_l[il], read(cell_count * k_size_row), kv_self.head * k_size_row, cell_count * k_size_row);
             }
         }
-    }
 
-    const size_t nread    = inp - src;
-    const size_t max_size = llama_state_get_size(ctx);
+        if (!kv_self.v_trans) {
+            for (uint32_t il = 0; il < n_layer; ++il) {
+                const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
 
-    GGML_ASSERT(nread <= max_size);
+                // Read type of value
+                int32_t v_type_i_ref;
+                read_to(&v_type_i_ref, sizeof(v_type_i_ref));
+                const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type;
+                if (v_type_i != v_type_i_ref) {
+                    LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il);
+                    return false;
+                }
 
-    return nread;
-}
+                // Read row size of value
+                uint64_t v_size_row_ref;
+                read_to(&v_size_row_ref, sizeof(v_size_row_ref));
+                const size_t v_size_row = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa);
+                if (v_size_row != v_size_row_ref) {
+                    LLAMA_LOG_ERROR("%s: mismatched value row size (%zu != %zu, layer %d)\n", __func__, v_size_row, (size_t) v_size_row_ref, il);
+                    return false;
+                }
 
-static bool llama_state_load_file_internal(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) {
-    llama_file file(path_session, "rb");
+                if (cell_count) {
+                    // Read and set the values for the whole cell range
+                    ggml_backend_tensor_set(kv_self.v_l[il], read(cell_count * v_size_row), kv_self.head * v_size_row, cell_count * v_size_row);
+                }
+            }
+        } else {
+            // For each layer, read the values for each cell (transposed)
+            for (uint32_t il = 0; il < n_layer; ++il) {
+                const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
 
-    // sanity checks
-    {
-        const uint32_t magic   = file.read_u32();
-        const uint32_t version = file.read_u32();
+                // Read type of value
+                int32_t v_type_i_ref;
+                read_to(&v_type_i_ref, sizeof(v_type_i_ref));
+                const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type;
+                if (v_type_i != v_type_i_ref) {
+                    LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il);
+                    return false;
+                }
 
-        if (magic != LLAMA_SESSION_MAGIC || version != LLAMA_SESSION_VERSION) {
-            LLAMA_LOG_ERROR("%s : unknown (magic, version) for session file: %08x, %08x\n", __func__, magic, version);
-            return false;
-        }
+                // Read element size of value
+                uint32_t v_size_el_ref;
+                read_to(&v_size_el_ref, sizeof(v_size_el_ref));
+                const size_t v_size_el = ggml_type_size(kv_self.v_l[il]->type);
+                if (v_size_el != v_size_el_ref) {
+                    LLAMA_LOG_ERROR("%s: mismatched value element size (%zu != %zu, layer %d)\n", __func__, v_size_el, (size_t) v_size_el_ref, il);
+                    return false;
+                }
 
-        llama_hparams session_hparams;
-        file.read_raw(&session_hparams, sizeof(llama_hparams));
+                // Read GQA embedding size
+                uint32_t n_embd_v_gqa_ref;
+                read_to(&n_embd_v_gqa_ref, sizeof(n_embd_v_gqa_ref));
+                if (n_embd_v_gqa != n_embd_v_gqa_ref) {
+                    LLAMA_LOG_ERROR("%s: mismatched GQA embedding size (%u != %u, layer %d)\n", __func__, n_embd_v_gqa, n_embd_v_gqa_ref, il);
+                    return false;
+                }
 
-        if (session_hparams != ctx->model.hparams) {
-            LLAMA_LOG_INFO("%s : model hparams didn't match from session file!\n", __func__);
-            return false;
+                if (cell_count) {
+                    // For each row in the transposed matrix, read the values for the whole cell range
+                    for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
+                        const size_t dst_offset = (kv_self.head + j * kv_self.size) * v_size_el;
+                        ggml_backend_tensor_set(kv_self.v_l[il], read(cell_count * v_size_el), dst_offset, cell_count * v_size_el);
+                    }
+                }
+            }
         }
+        return true;
     }
 
-    // load the prompt
-    {
-        const uint32_t n_token_count = file.read_u32();
+    void read_kv_cache(struct llama_context * ctx, llama_seq_id seq_id = -1) {
+        uint32_t cell_count;
+        read_to(&cell_count, sizeof(cell_count));
 
-        if (n_token_count > n_token_capacity) {
-            LLAMA_LOG_ERROR("%s : token count in session file exceeded capacity! %u > %zu\n", __func__, n_token_count, n_token_capacity);
-            return false;
-        }
+        bool res = read_kv_cache_meta(ctx, cell_count, seq_id) && read_kv_cache_data(ctx, cell_count);
 
-        file.read_raw(tokens_out, sizeof(llama_token) * n_token_count);
-        *n_token_count_out = n_token_count;
+        if (!res) {
+            if (seq_id == -1) {
+                llama_kv_cache_clear(ctx);
+            } else {
+                llama_kv_cache_seq_rm(ctx, seq_id, -1, -1);
+            }
+            throw std::runtime_error("failed to restore kv cache");
+        }
     }
+};
 
-    // restore the context state
-    {
-        const size_t n_state_size_cur = file.size - file.tell();
-        const size_t n_state_size_max = llama_state_get_size(ctx);
-
-        if (n_state_size_cur > n_state_size_max) {
-            LLAMA_LOG_ERROR("%s : the state size in session file is too big! max %zu, got %zu\n", __func__, n_state_size_max, n_state_size_cur);
-            return false;
-        }
+struct llama_data_write_dummy : llama_data_write {
+    size_t size_written = 0;
 
-        std::vector<uint8_t> state_data(n_state_size_max);
-        file.read_raw(state_data.data(), n_state_size_cur);
+    llama_data_write_dummy() {}
 
-        llama_state_set_data(ctx, state_data.data());
+    void write(const void * /* src */, size_t size) override {
+        size_written += size;
     }
 
-    return true;
-}
-
-bool llama_state_load_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) {
-    try {
-        return llama_state_load_file_internal(ctx, path_session, tokens_out, n_token_capacity, n_token_count_out);
-    } catch (const std::exception & err) {
-        LLAMA_LOG_ERROR("error loading session file: %s\n", err.what());
-        return false;
+    void write_tensor_data(const struct ggml_tensor * /* tensor */, size_t /* offset */, size_t size) override {
+        size_written += size;
     }
-}
-
-static bool llama_state_save_file_internal(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) {
-    llama_file file(path_session, "wb");
 
-    file.write_u32(LLAMA_SESSION_MAGIC);
-    file.write_u32(LLAMA_SESSION_VERSION);
-
-    file.write_raw(&ctx->model.hparams, sizeof(llama_hparams));
-
-    // save the prompt
-    file.write_u32((uint32_t) n_token_count);
-    file.write_raw(tokens, sizeof(llama_token) * n_token_count);
-
-    // save the context state using stream saving
-    llama_data_file_context data_ctx(&file);
-    llama_state_get_data_internal(ctx, &data_ctx);
-
-    return true;
-}
-
-bool llama_state_save_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) {
-    try {
-        return llama_state_save_file_internal(ctx, path_session, tokens, n_token_count);
-    } catch (const std::exception & err) {
-        LLAMA_LOG_ERROR("error saving session file: %s\n", err.what());
-        return false;
+    size_t get_size_written() override {
+        return size_written;
     }
-}
-
-size_t llama_state_seq_get_size(struct llama_context* ctx, llama_seq_id seq_id) {
-    // save the size of size_t as a uint32_t for safety check
-    const size_t size_t_size_size = sizeof(uint32_t);
+};
 
-    // other values
-    const size_t s_cell_count_size = sizeof(uint32_t);
-    const size_t s_layer_count_size = sizeof(uint32_t);
-    const size_t n_embd_v_gqa_size = sizeof(uint32_t);
+struct llama_data_write_buffer : llama_data_write {
+    uint8_t * ptr;
+    size_t buf_size = 0;
+    size_t size_written = 0;
 
-    size_t s_cell_count = 0;
-    size_t s_cell_data_size = 0;
-    const auto & kv_self = ctx->kv_self;
-    const auto & hparams = ctx->model.hparams;
+    llama_data_write_buffer(uint8_t * p, size_t len) : ptr(p), buf_size(len) {}
 
-    const uint32_t n_layer = hparams.n_layer;
+    void write(const void * src, size_t size) override {
+        if (size > buf_size) {
+            throw std::runtime_error("unexpectedly reached end of buffer");
+        }
+        memcpy(ptr, src, size);
+        ptr += size;
+        size_written += size;
+        buf_size -= size;
+    }
 
-    for (uint32_t i = 0; i < kv_self.size; ++i) {
-        const auto & cell = kv_self.cells[i];
-        if (cell.seq_id.count(seq_id) > 0) {
-            ++s_cell_count;
-            s_cell_data_size += sizeof(llama_pos);
+    void write_tensor_data(const struct ggml_tensor * tensor, size_t offset, size_t size) override {
+        if (size > buf_size) {
+            throw std::runtime_error("unexpectedly reached end of buffer");
         }
+        ggml_backend_tensor_get(tensor, ptr, offset, size);
+        ptr += size;
+        size_written += size;
+        buf_size -= size;
     }
 
-    for (int il = 0; il < (int)n_layer; ++il) {
-        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
-        const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+    size_t get_size_written() override {
+        return size_written;
+    }
+};
 
-        // types of keys and values
-        s_cell_data_size += sizeof(int32_t) * 2;
-        // k_size_row and v_size_el values of layer
-        s_cell_data_size += sizeof(size_t) * 2;
+struct llama_data_read_buffer : llama_data_read {
+    const uint8_t * ptr;
+    size_t buf_size = 0;
+    size_t size_read = 0;
 
-        // keys
-        const size_t k_size_row = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa);
-        s_cell_data_size += k_size_row * s_cell_count;
+    llama_data_read_buffer(const uint8_t * p, size_t len) : ptr(p), buf_size(len) {}
 
-        // values (transposed)
-        const size_t v_size_el = ggml_type_size(kv_self.v_l[il]->type);
-        s_cell_data_size += v_size_el * s_cell_count * n_embd_v_gqa;
+    const uint8_t * read(size_t size) override {
+        const uint8_t * base_ptr = ptr;
+        if (size > buf_size) {
+            throw std::runtime_error("unexpectedly reached end of buffer");
+        }
+        ptr += size;
+        size_read += size;
+        buf_size -= size;
+        return base_ptr;
     }
 
-    const size_t s_total = (
-        size_t_size_size +
-        s_cell_count_size +
-        s_layer_count_size +
-        n_embd_v_gqa_size +
-        s_cell_data_size
-        );
-
-    return s_total;
-}
-
-static size_t llama_state_seq_get_data_internal(struct llama_context * ctx, llama_data_context & data_ctx, llama_seq_id seq_id) {
-    llama_synchronize(ctx);
+    void read_to(void * dst, size_t size) override {
+        memcpy(dst, read(size), size);
+    }
 
-    const auto & kv_self = ctx->kv_self;
-    GGML_ASSERT(!kv_self.recurrent); // not implemented
+    size_t get_size_read() override {
+        return size_read;
+    }
+};
 
-    // Save the size of size_t as a uint32_t for safety check
-    const uint32_t size_t_size = sizeof(size_t);
-    data_ctx.write(&size_t_size, sizeof(size_t_size));
+struct llama_data_write_file : llama_data_write {
+    llama_file * file;
+    size_t size_written = 0;
+    std::vector<uint8_t> temp_buffer;
 
-    std::vector<std::pair<uint32_t, uint32_t>> cell_ranges; // ranges, from inclusive, to exclusive
-    uint32_t cell_count = 0;
+    llama_data_write_file(llama_file * f) : file(f) {}
 
-    // Count the number of cells with the specified seq_id
-    // Find all the ranges of cells with this seq id
-    {
-        uint32_t cell_range_begin = kv_self.size;
-        for (uint32_t i = 0; i < kv_self.size; ++i) {
-            const auto & cell = kv_self.cells[i];
-            if (cell.has_seq_id(seq_id)) {
-                ++cell_count;
-                if (cell_range_begin == kv_self.size) {
-                    cell_range_begin = i;
-                }
-            }
-            else {
-                if (cell_range_begin != kv_self.size) {
-                    cell_ranges.emplace_back(cell_range_begin, i);
-                    cell_range_begin = kv_self.size;
-                }
-            }
-        }
-        if (cell_range_begin != kv_self.size) {
-            cell_ranges.emplace_back(cell_range_begin, kv_self.size);
-        }
+    void write(const void * src, size_t size) override {
+        file->write_raw(src, size);
+        size_written += size;
+    }
 
-        // DEBUG CHECK: Sum of cell counts in ranges should equal the total cell count
-        uint32_t cell_count_check = 0;
-        for (const auto & range : cell_ranges) {
-            cell_count_check += range.second - range.first;
-        }
-        GGML_ASSERT(cell_count == cell_count_check);
+    void write_tensor_data(const struct ggml_tensor * tensor, size_t offset, size_t size) override {
+        temp_buffer.resize(size);
+        ggml_backend_tensor_get(tensor, temp_buffer.data(), offset, size);
+        write(temp_buffer.data(), temp_buffer.size());
     }
 
-    // Write the cell count
-    data_ctx.write(&cell_count, sizeof(cell_count));
+    size_t get_size_written() override {
+        return size_written;
+    }
+};
 
-    const auto & hparams = ctx->model.hparams;
-    const uint32_t n_layer = hparams.n_layer;
+struct llama_data_read_file : llama_data_read {
+    llama_file * file;
+    size_t size_read = 0;
+    std::vector<uint8_t> temp_buffer;
 
-    // Write the layer count
-    data_ctx.write(&n_layer, sizeof(n_layer));
+    llama_data_read_file(llama_file * f) : file(f) {}
 
-    // Write n_embd_v_gqa (reference value)
-    {
-        const uint32_t n_embd_v_gqa_ref = hparams.n_embd_v_gqa() + hparams.n_embd_k_s();
-        data_ctx.write(&n_embd_v_gqa_ref, sizeof(n_embd_v_gqa_ref));
+    void read_to(void * dst, size_t size) override {
+        file->read_raw(dst, size);
+        size_read += size;
     }
 
-    // Iterate the ranges and write all the pos (this is the token position in the prompt)
-    for (const auto & range : cell_ranges) {
-        for (uint32_t i = range.first; i < range.second; ++i) {
-            const auto & cell = kv_self.cells[i];
-            data_ctx.write(&cell.pos, sizeof(cell.pos));
-        }
+    const uint8_t * read(size_t size) override {
+        temp_buffer.resize(size);
+        read_to(temp_buffer.data(), size);
+        return temp_buffer.data();
     }
 
-    // Iterate and write all the keys first, each row is a cell
-    // Get whole range at a time
-    std::vector<uint8_t> tmp_buf;
-    for (int il = 0; il < (int)n_layer; ++il) {
-        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
+    size_t get_size_read() override {
+        return size_read;
+    }
+};
 
-        // Write key type
-        const int32_t k_type_i = (int32_t)kv_self.k_l[il]->type;
-        data_ctx.write(&k_type_i, sizeof(k_type_i));
+/** copy state data into either a buffer or file depending on the passed in context
+ *
+ * file context:
+ * llama_file file("/path", "wb");
+ * llama_data_write_file data_ctx(&file);
+ * llama_state_get_data_internal(ctx, data_ctx);
+ *
+ * buffer context:
+ * std::vector<uint8_t> buf(max_size, 0);
+ * llama_data_write_buffer data_ctx(buf.data(), max_size);
+ * llama_state_get_data_internal(ctx, data_ctx);
+ *
+*/
+static size_t llama_state_get_data_internal(struct llama_context * ctx, llama_data_write & data_ctx) {
+    llama_synchronize(ctx);
 
-        // Write row size of key
-        const size_t k_size_row = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa);
-        data_ctx.write(&k_size_row, sizeof(k_size_row));
+    data_ctx.write_model_info(ctx);
 
-        // Read each range of cells of k_size length each into tmp_buf and write out
-        for (const auto & range : cell_ranges) {
-            const size_t range_size = range.second - range.first;
-            tmp_buf.resize(range_size * k_size_row);
-            ggml_backend_tensor_get(kv_self.k_l[il], tmp_buf.data(), range.first * k_size_row, range_size * k_size_row);
-            data_ctx.write(tmp_buf.data(), tmp_buf.size());
-        }
-    }
+    data_ctx.write_rng(ctx->sampling.rng);
 
-    // TODO: simplify, reduce copy-paste
-    if (!kv_self.v_trans) {
-        for (int il = 0; il < (int)n_layer; ++il) {
-            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+    // copy outputs
+    data_ctx.write_output_ids(ctx);
+    data_ctx.write_logits(ctx);
+    data_ctx.write_embeddings(ctx);
 
-            // Write value type
-            const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type;
-            data_ctx.write(&v_type_i, sizeof(v_type_i));
+    data_ctx.write_kv_cache(ctx);
 
-            // Write row size of value
-            const size_t v_size_row = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa);
-            data_ctx.write(&v_size_row, sizeof(v_size_row));
+    return data_ctx.get_size_written();
+}
 
-            // Read each range of cells of v_size length each into tmp_buf and write out
-            for (const auto & range : cell_ranges) {
-                const size_t range_size = range.second - range.first;
-                tmp_buf.resize(range_size * v_size_row);
-                ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), range.first * v_size_row, range_size * v_size_row);
-                data_ctx.write(tmp_buf.data(), tmp_buf.size());
-            }
-        }
-    } else {
-        // For the values, they are transposed, so we also need the element size and get the element ranges from each row
-        const uint32_t kv_size = kv_self.size;
-        for (int il = 0; il < (int)n_layer; ++il) {
-            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
-
-            // Write value type
-            const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type;
-            data_ctx.write(&v_type_i, sizeof(v_type_i));
-
-            // Write element size
-            const size_t v_size_el = ggml_type_size(kv_self.v_l[il]->type);
-            data_ctx.write(&v_size_el, sizeof(v_size_el));
-
-            // For each row, we get the element values of each cell
-            for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
-                // Read each range of cells of v_size_el length each into tmp_buf and write out
-                for (const auto & range : cell_ranges) {
-                    const size_t range_size = range.second - range.first;
-                    const size_t src_offset = (range.first + j * kv_size) * v_size_el;
-                    tmp_buf.resize(range_size * v_size_el);
-                    ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), src_offset, tmp_buf.size());
-                    data_ctx.write(tmp_buf.data(), tmp_buf.size());
-                }
-            }
-        }
+size_t llama_state_get_data(struct llama_context * ctx, uint8_t * dst, size_t size) {
+    llama_data_write_buffer data_ctx(dst, size);
+    try {
+        return llama_state_get_data_internal(ctx, data_ctx);
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: error saving state: %s\n", __func__, err.what());
+        return 0;
     }
-
-    return data_ctx.get_size_written();
 }
 
-size_t llama_state_seq_get_data(struct llama_context* ctx, uint8_t* dst, llama_seq_id seq_id) {
-    llama_data_buffer_context data_ctx(dst);
-    return llama_state_seq_get_data_internal(ctx, data_ctx, seq_id);
+// Returns the *actual* size of the state.
+// Intended to be used when saving to state to a buffer.
+size_t llama_state_get_size(struct llama_context * ctx) {
+    llama_data_write_dummy data_ctx;
+    try {
+        return llama_state_get_data_internal(ctx, data_ctx);
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: error getting state size: %s\n", __func__, err.what());
+        return 0;
+    }
 }
 
-size_t llama_state_seq_set_data(struct llama_context * ctx, const uint8_t * src, llama_seq_id dest_seq_id) {
+static size_t llama_state_set_data_internal(struct llama_context * ctx, llama_data_read & data_ctx) {
     llama_synchronize(ctx);
 
-    auto & kv_self = ctx->kv_self;
-    GGML_ASSERT(!kv_self.recurrent); // not implemented
+    data_ctx.read_model_info(ctx);
+
+    // set rng
+    data_ctx.read_rng(ctx->sampling.rng);
 
-    // Wipe the slot
-    llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1);
+    // set outputs
+    data_ctx.read_output_ids(ctx);
+    data_ctx.read_logits(ctx);
+    data_ctx.read_embeddings(ctx);
 
-    const uint8_t * inp = src;
+    data_ctx.read_kv_cache(ctx);
 
-    // Read size of size_t
-    uint32_t size_t_size;
-    memcpy(&size_t_size, inp, sizeof(size_t_size));
-    inp += sizeof(size_t_size);
-    if (size_t_size != sizeof(size_t)) {
-        LLAMA_LOG_ERROR("%s: size_t size mismatch\n", __func__);
+    return data_ctx.get_size_read();
+}
+
+// Sets the state reading from the specified source address
+size_t llama_state_set_data(struct llama_context * ctx, const uint8_t * src, size_t size) {
+    llama_data_read_buffer data_ctx(src, size);
+    try {
+        return llama_state_set_data_internal(ctx, data_ctx);
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: error loading state: %s\n", __func__, err.what());
         return 0;
     }
+}
 
-    // Read the cell count
-    uint32_t cell_count;
-    memcpy(&cell_count, inp, sizeof(cell_count));
-    inp += sizeof(cell_count);
+static bool llama_state_load_file_internal(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) {
+    llama_file file(path_session, "rb");
 
-    // Read the layer count
-    uint32_t n_layer_ref;
-    memcpy(&n_layer_ref, inp, sizeof(n_layer_ref));
-    inp += sizeof(n_layer_ref);
+    // sanity checks
+    {
+        const uint32_t magic   = file.read_u32();
+        const uint32_t version = file.read_u32();
 
-    // Read n_embd_v_gqa
-    uint32_t n_embd_v_gqa_ref;
-    memcpy(&n_embd_v_gqa_ref, inp, sizeof(n_embd_v_gqa_ref));
-    inp += sizeof(n_embd_v_gqa_ref);
+        if (magic != LLAMA_SESSION_MAGIC || version != LLAMA_SESSION_VERSION) {
+            LLAMA_LOG_ERROR("%s: unknown (magic, version) for session file: %08x, %08x\n", __func__, magic, version);
+            return false;
+        }
+    }
 
-    // Sanity check model compatibility
-    const auto & hparams = ctx->model.hparams;
-    const uint32_t n_layer = hparams.n_layer;
+    // load the prompt
+    {
+        const uint32_t n_token_count = file.read_u32();
 
-    if (n_layer != n_layer_ref) {
-        LLAMA_LOG_ERROR("%s: mismatched n_layer (%d != %d)\n", __func__, n_layer, n_layer_ref);
-        return 0;
-    }
+        if (n_token_count > n_token_capacity) {
+            LLAMA_LOG_ERROR("%s: token count in session file exceeded capacity! %u > %zu\n", __func__, n_token_count, n_token_capacity);
+            return false;
+        }
 
-    if (hparams.n_embd_v_gqa() != n_embd_v_gqa_ref) {
-        LLAMA_LOG_ERROR("%s: mismatched n_embd_v_gqa (%d != %d)\n", __func__, hparams.n_embd_v_gqa(), n_embd_v_gqa_ref);
-        return 0;
+        file.read_raw(tokens_out, sizeof(llama_token) * n_token_count);
+        *n_token_count_out = n_token_count;
     }
 
-    // Allocate the new cells for the slot
-    if (cell_count) {
-        llama_batch batch = llama_batch_init(cell_count, 0, 1);
-        batch.n_tokens = cell_count;
-        for (uint32_t i = 0; i < cell_count; ++i) {
-            llama_pos pos;
-            memcpy(&pos, inp, sizeof(pos));
-            inp += sizeof(pos);
+    // restore the context state
+    {
+        const size_t n_state_size_cur = file.size - file.tell();
+
+        llama_data_read_file data_ctx(&file);
+        const size_t n_read = llama_state_set_data_internal(ctx, data_ctx);
 
-            batch.pos[i] = pos;
-            batch.n_seq_id[i] = 1;
-            batch.seq_id[i][0] = dest_seq_id;
-        }
-        if (!llama_kv_cache_find_slot(kv_self, batch)) {
-            llama_batch_free(batch);
-            LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__);
-            return 0;
+        if (n_read != n_state_size_cur) {
+            LLAMA_LOG_ERROR("%s: did not read all of the session file data! size %zu, got %zu\n", __func__, n_state_size_cur, n_read);
+            return false;
         }
+    }
+    return true;
+}
 
-        // DEBUG CHECK: kv_self.head should be our first cell, kv_self.head + cell_count - 1 should be our last cell (verify seq_id and pos values)
-        // Assume that this is one contiguous block of cells
-        GGML_ASSERT(kv_self.head + cell_count <= kv_self.size);
-        GGML_ASSERT(kv_self.cells[kv_self.head].pos == batch.pos[0]);
-        GGML_ASSERT(kv_self.cells[kv_self.head + cell_count - 1].pos == batch.pos[cell_count - 1]);
-        GGML_ASSERT(kv_self.cells[kv_self.head].has_seq_id(dest_seq_id));
-        GGML_ASSERT(kv_self.cells[kv_self.head + cell_count - 1].has_seq_id(dest_seq_id));
-
-        // Cleanup
-        llama_batch_free(batch);
+bool llama_state_load_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) {
+    try {
+        return llama_state_load_file_internal(ctx, path_session, tokens_out, n_token_capacity, n_token_count_out);
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: error loading session file: %s\n", __func__, err.what());
+        return false;
     }
+}
 
-    const uint32_t kv_size = kv_self.size;
-    const uint32_t kv_head = kv_self.head;
-
-    // For each layer, read the keys for each cell, one row is one cell, read as one contiguous blo
-    for (int il = 0; il < (int)n_layer; ++il) {
-        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
-
-        // Read type of key
-        int32_t k_type_i_ref;
-        memcpy(&k_type_i_ref, inp, sizeof(k_type_i_ref));
-        inp += sizeof(k_type_i_ref);
-        const int32_t k_type_i = (int32_t)kv_self.k_l[il]->type;
-        if (k_type_i != k_type_i_ref) {
-            llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1);
-            LLAMA_LOG_ERROR("%s: mismatched key type (%d != %d, layer %d)\n", __func__, k_type_i, k_type_i_ref, il);
-            return 0;
-        }
+static bool llama_state_save_file_internal(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) {
+    llama_file file(path_session, "wb");
 
-        // Read row size of key
-        size_t k_size_row_ref;
-        memcpy(&k_size_row_ref, inp, sizeof(k_size_row_ref));
-        inp += sizeof(k_size_row_ref);
-        const size_t k_size_row = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa);
-        if (k_size_row != k_size_row_ref) {
-            llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1);
-            LLAMA_LOG_ERROR("%s: mismatched key row size (%zu != %zu, layer %d)\n", __func__, k_size_row, k_size_row_ref, il);
-            return 0;
-        }
+    file.write_u32(LLAMA_SESSION_MAGIC);
+    file.write_u32(LLAMA_SESSION_VERSION);
 
-        if (cell_count) {
-            // Read and set the keys for the whole cell range
-            ggml_backend_tensor_set(kv_self.k_l[il], inp, kv_head * k_size_row, cell_count * k_size_row);
-            inp += cell_count * k_size_row;
-        }
+    // save the prompt
+    file.write_u32((uint32_t) n_token_count);
+    file.write_raw(tokens, sizeof(llama_token) * n_token_count);
+
+    // save the context state using stream saving
+    llama_data_write_file data_ctx(&file);
+    llama_state_get_data_internal(ctx, data_ctx);
+
+    return true;
+}
+
+bool llama_state_save_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) {
+    try {
+        return llama_state_save_file_internal(ctx, path_session, tokens, n_token_count);
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: error saving session file: %s\n", __func__, err.what());
+        return false;
     }
+}
 
-    // TODO: simplify, reduce copy-paste
-    if (!kv_self.v_trans) {
-        for (int il = 0; il < (int)n_layer; ++il) {
-            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+static size_t llama_state_seq_get_data_internal(struct llama_context * ctx, llama_data_write & data_ctx, llama_seq_id seq_id) {
+    llama_synchronize(ctx);
 
-            // Read type of value
-            int32_t v_type_i_ref;
-            memcpy(&v_type_i_ref, inp, sizeof(v_type_i_ref));
-            inp += sizeof(v_type_i_ref);
-            const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type;
-            if (v_type_i != v_type_i_ref) {
-                llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1);
-                LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il);
-                return 0;
-            }
+    data_ctx.write_kv_cache(ctx, seq_id);
 
-            // Read row size of value
-            size_t v_size_row_ref;
-            memcpy(&v_size_row_ref, inp, sizeof(v_size_row_ref));
-            inp += sizeof(v_size_row_ref);
-            const size_t v_size_row = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa);
-            if (v_size_row != v_size_row_ref) {
-                llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1);
-                LLAMA_LOG_ERROR("%s: mismatched value row size (%zu != %zu, layer %d)\n", __func__, v_size_row, v_size_row_ref, il);
-                return 0;
-            }
+    return data_ctx.get_size_written();
+}
 
-            if (cell_count) {
-                // Read and set the values for the whole cell range
-                ggml_backend_tensor_set(kv_self.v_l[il], inp, kv_head * v_size_row, cell_count * v_size_row);
-                inp += cell_count * v_size_row;
-            }
-        }
-    } else {
-        // For each layer, read the values for each cell (transposed)
-        for (int il = 0; il < (int)n_layer; ++il) {
-            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
-
-            // Read type of value
-            int32_t v_type_i_ref;
-            memcpy(&v_type_i_ref, inp, sizeof(v_type_i_ref));
-            inp += sizeof(v_type_i_ref);
-            const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type;
-            if (v_type_i != v_type_i_ref) {
-                llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1);
-                LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il);
-                return 0;
-            }
-
-            // Read element size of value
-            size_t v_size_el_ref;
-            memcpy(&v_size_el_ref, inp, sizeof(v_size_el_ref));
-            inp += sizeof(v_size_el_ref);
-            const size_t v_size_el = ggml_type_size(kv_self.v_l[il]->type);
-            if (v_size_el != v_size_el_ref) {
-                llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1);
-                LLAMA_LOG_ERROR("%s: mismatched value element size (%zu != %zu, layer %d)\n", __func__, v_size_el, v_size_el_ref, il);
-                return 0;
-            }
+size_t llama_state_seq_get_size(struct llama_context * ctx, llama_seq_id seq_id) {
+    llama_data_write_dummy data_ctx;
+    return llama_state_seq_get_data_internal(ctx, data_ctx, seq_id);
+}
 
-            if (cell_count) {
-                // For each row in the transposed matrix, read the values for the whole cell range
-                for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
-                    const size_t dst_offset = (kv_head + j * kv_size) * v_size_el;
-                    ggml_backend_tensor_set(kv_self.v_l[il], inp, dst_offset, cell_count * v_size_el);
-                    inp += cell_count * v_size_el;
-                }
-            }
-        }
+size_t llama_state_seq_get_data(struct llama_context * ctx, uint8_t * dst, size_t size, llama_seq_id seq_id) {
+    llama_data_write_buffer data_ctx(dst, size);
+    try {
+        return llama_state_seq_get_data_internal(ctx, data_ctx, seq_id);
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: error saving sequence state: %s\n", __func__, err.what());
+        return 0;
     }
+}
 
-    const size_t nread = inp - src;
+static size_t llama_state_seq_set_data_internal(struct llama_context * ctx, llama_data_read & data_ctx, llama_seq_id dest_seq_id) {
+    llama_synchronize(ctx);
+
+    data_ctx.read_kv_cache(ctx, dest_seq_id);
 
-    return nread;
+    return data_ctx.get_size_read();
+}
+
+size_t llama_state_seq_set_data(struct llama_context * ctx, const uint8_t * src, size_t size, llama_seq_id dest_seq_id) {
+    llama_data_read_buffer data_ctx(src, size);
+    try {
+        return llama_state_seq_set_data_internal(ctx, data_ctx, dest_seq_id);
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: error loading sequence state: %s\n", __func__, err.what());
+        return 0;
+    }
 }
 
 static size_t llama_state_seq_save_file_internal(struct llama_context * ctx, const char * filepath, llama_seq_id seq_id, const llama_token * tokens, size_t n_token_count) {
@@ -18384,11 +18526,11 @@ static size_t llama_state_seq_save_file_internal(struct llama_context * ctx, con
     file.write_u32(LLAMA_STATE_SEQ_VERSION);
 
     // save the prompt
-    file.write_u32((uint32_t)n_token_count);
+    file.write_u32((uint32_t) n_token_count);
     file.write_raw(tokens, sizeof(llama_token) * n_token_count);
 
     // save the context state using stream saving
-    llama_data_file_context data_ctx(&file);
+    llama_data_write_file data_ctx(&file);
     llama_state_seq_get_data_internal(ctx, data_ctx, seq_id);
 
     const size_t res = file.tell();
@@ -18426,9 +18568,8 @@ static size_t llama_state_seq_load_file_internal(struct llama_context * ctx, con
     // restore the context state
     {
         const size_t state_size = file.size - file.tell();
-        std::vector<uint8_t> state_data(state_size);
-        file.read_raw(state_data.data(), state_size);
-        const size_t nread = llama_state_seq_set_data(ctx, state_data.data(), dest_seq_id);
+        llama_data_read_file data_ctx(&file);
+        const size_t nread = llama_state_seq_set_data_internal(ctx, data_ctx, dest_seq_id);
         if (!nread) {
             LLAMA_LOG_ERROR("%s: failed to restore sequence state\n", __func__);
             return 0;
@@ -18444,7 +18585,7 @@ size_t llama_state_seq_save_file(struct llama_context * ctx, const char * filepa
     try {
         return llama_state_seq_save_file_internal(ctx, filepath, seq_id, tokens, n_token_count);
     } catch (const std::exception & err) {
-        LLAMA_LOG_ERROR("error saving sequence state file: %s\n", err.what());
+        LLAMA_LOG_ERROR("%s: error saving sequence state file: %s\n", __func__, err.what());
         return 0;
     }
 }
@@ -18453,7 +18594,7 @@ size_t llama_state_seq_load_file(struct llama_context * ctx, const char * filepa
     try {
         return llama_state_seq_load_file_internal(ctx, filepath, dest_seq_id, tokens_out, n_token_capacity, n_token_count_out);
     } catch (const std::exception & err) {
-        LLAMA_LOG_ERROR("error loading sequence state file: %s\n", err.what());
+        LLAMA_LOG_ERROR("%s: error loading sequence state file: %s\n", __func__, err.what());
         return 0;
     }
 }
@@ -18625,7 +18766,7 @@ float * llama_get_logits_ith(struct llama_context * ctx, int32_t i) {
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: invalid logits id %d, reason: %s\n", __func__, i, err.what());
 #ifndef NDEBUG
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
 #endif
         return nullptr;
     }
@@ -18670,7 +18811,7 @@ float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i) {
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: invalid embeddings id %d, reason: %s\n", __func__, i, err.what());
 #ifndef NDEBUG
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
 #endif
         return nullptr;
     }
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index 2c03c60d..a2182c1b 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -94,7 +94,7 @@ static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float m
         // This is going to create some weird integers though.
         ggml_backend_tensor_set(tensor, data.data(), 0, ggml_nbytes(tensor));
     } else {
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 }
 
@@ -132,7 +132,7 @@ static std::vector<float> tensor_to_float(const ggml_tensor * t) {
                         tt.to_float(&buf[i], vq.data(), bs);
                         tv.insert(tv.end(), vq.begin(), vq.end());
                     } else {
-                        GGML_ASSERT(false);
+                        GGML_ABORT("fatal error");
                     }
                 }
             }
@@ -804,8 +804,7 @@ struct test_cpy : public test_case {
 
     test_cpy(ggml_type type_src = GGML_TYPE_F32, ggml_type type_dst = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {10, 10, 10, 1},
-            std::array<int64_t, 4> permute = {0, 0, 0, 0},
-            bool _dst_use_permute = false)
+            std::array<int64_t, 4> permute = {0, 0, 0, 0})
         : type_src(type_src), type_dst(type_dst), ne(ne), permute(permute),
           _src_use_permute(permute[0] + permute[1] + permute[2] + permute[3] > 0) {}
 
@@ -1435,7 +1434,7 @@ struct test_argsort : public test_case {
                     ggml_backend_tensor_set(t, data.data(), r * t->nb[1], t->ne[0] * sizeof(float));
                 }
             } else {
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
             }
         }
     }
@@ -1512,6 +1511,7 @@ struct test_group_norm : public test_case {
     const ggml_type type;
     const std::array<int64_t, 4> ne;
     const int32_t num_groups;
+    const float eps;
 
     std::string vars() override {
         return VARS_TO_STR3(type, ne, num_groups);
@@ -1519,12 +1519,13 @@ struct test_group_norm : public test_case {
 
     test_group_norm(ggml_type type = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {64, 64, 320, 1},
-            int32_t num_groups = 32)
-        : type(type), ne(ne), num_groups(num_groups) {}
+            int32_t num_groups = 32,
+            float eps = 1e-6f)
+        : type(type), ne(ne), num_groups(num_groups), eps(eps) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_tensor * out = ggml_group_norm(ctx, a, num_groups);
+        ggml_tensor * out = ggml_group_norm(ctx, a, num_groups, eps);
         return out;
     }
 };
@@ -2140,6 +2141,19 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
 
     test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32));
     test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16));
+    // test cases for 1D im2col
+    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false));
+    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false));
+
+    test_cases.emplace_back(new test_conv_transpose_1d());
+    test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 3, 0, 1));
+    test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 2, 0, 1));
+    test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 1, 0, 1));
+    test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {3,2,2,1}, 2, 0, 1));
+    test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {3,2,2,1}, 1, 0, 1));
+    test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {3,1,2,1}, 1, 0, 1));
+    test_cases.emplace_back(new test_conv_transpose_1d({2,1,1,1}, {3,1,1,1}, 1, 0, 1));
+
 
     test_cases.emplace_back(new test_conv_transpose_1d());
     test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 3, 0, 1));
@@ -2269,7 +2283,10 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
 
     for (ggml_type type_a : other_types) {
         for (ggml_type type_b : {GGML_TYPE_F32}) {
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, { 1,  1}, {1, 1}));
+            if (ggml_blck_size(type_a) != 256) {
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, ggml_blck_size(type_a), {1,  1}, {1, 1}));
+            }
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {1,  1}, {1, 1}));
         }
     }
 
@@ -2462,7 +2479,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
         return true;
     }
 
-    GGML_ASSERT(false);
+    GGML_ABORT("fatal error");
     return false;
 }
 
diff --git a/tests/test-sampling.cpp b/tests/test-sampling.cpp
index 6374958f..de858bd3 100644
--- a/tests/test-sampling.cpp
+++ b/tests/test-sampling.cpp
@@ -166,12 +166,12 @@ static void test_sampler_queue(
     for (auto s : samplers_sequence) {
         switch (s){
             case 'k': llama_sample_top_k    (nullptr, &candidates_p, top_k, 1); break;
-            case 'f': GGML_ASSERT(false && "tail_free test not implemented");   break;
-            case 'y': GGML_ASSERT(false && "typical test not implemented");     break;
+            case 'f': GGML_ABORT("tail_free test not implemented");   break;
+            case 'y': GGML_ABORT("typical test not implemented");     break;
             case 'p': llama_sample_top_p    (nullptr, &candidates_p, top_p, 1); break;
             case 'm': llama_sample_min_p    (nullptr, &candidates_p, min_p, 1); break;
-            case 't': GGML_ASSERT(false && "temperature test not implemented"); break;
-            default : GGML_ASSERT(false && "Unknown sampler");                  break;
+            case 't': GGML_ABORT("temperature test not implemented"); break;
+            default : GGML_ABORT("Unknown sampler");                  break;
         }
 
         llama_sample_softmax(nullptr, &candidates_p); // make sure tokens are sorted for tests
@@ -222,7 +222,7 @@ static void test_sampler_queue(
             GGML_ASSERT(candidates_p.data[0].id == max_token_id);
             GGML_ASSERT(candidates_p.data[expected_size-1].id == min_token_id);
         } else {
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
     }