CUDA full GPU acceleration, KV cache in VRAM (#1827)

* Fixed CUDA RoPE

* ggml_cuda_mul_mat_vec_p021

* ggml_cuda_scale

* ggml_cuda_diag_mask_inf

* ggml_is_permuted

* ggml_cuda_cpy

* flatten rows for ggml_cuda_op

* Added a --low-vram option

* Fixed Windows performance

* Fixed LLAMA_CUDA_DMMV_Y > 1 for WizardLM

1 files changed, 8 insertions, 0 deletions

diff --git a/examples/common.cpp b/examples/common.cpp
index df69f273..dc69e537 100644
--- a/examples/common.cpp
+++ b/examples/common.cpp
@@ -332,6 +332,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
 #else
       fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
 #endif // GGML_USE_CUBLAS
+        } else if (arg == "--low-vram" || arg == "-lv") {
+#ifdef GGML_USE_CUBLAS
+            params.low_vram = true;
+#else
+      fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
+#endif // GGML_USE_CUBLAS
         } else if (arg == "--no-mmap") {
             params.use_mmap = false;
         } else if (arg == "--mtest") {
@@ -479,6 +485,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  -ts SPLIT --tensor-split SPLIT\n");
     fprintf(stderr, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
     fprintf(stderr, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n" );
+    fprintf(stderr, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n" );
 #endif
     fprintf(stderr, "  --mtest               compute maximum memory usage\n");
     fprintf(stderr, "  --export              export the computation graph to 'llama.ggml'\n");
@@ -528,6 +535,7 @@ struct llama_context * llama_init_from_gpt_params(const gpt_params & params) {
     lparams.n_gpu_layers = params.n_gpu_layers;
     lparams.main_gpu     = params.main_gpu;
     memcpy(lparams.tensor_split, params.tensor_split, LLAMA_MAX_DEVICES*sizeof(float));
+    lparams.low_vram     = params.low_vram;
     lparams.seed         = params.seed;
     lparams.f16_kv       = params.memory_f16;
     lparams.use_mmap     = params.use_mmap;