llama : add pipeline parallelism support (#6017)

* llama : add pipeline parallelism support for batch processing with multiple CUDA GPUs

ggml-ci

* server : add -ub, --ubatch-size parameter

* fix server embedding test

* llama : fix Mamba inference for pipeline parallelism

Tested to work correctly with both `main` and `parallel` examples.

* llama : limit max batch size to n_batch

* add LLAMA_SCHED_MAX_COPIES to configure the number of input copies for pipeline parallelism
default increase to 4 (from 2)

changing this value may improve performance for some systems, but increases memory usage

* fix hip build

* fix sycl build (disable cpy_tensor_async)

* fix hip build

* llama : limit n_batch and n_ubatch to n_ctx during context creation

* llama : fix norm backend

* batched-bench : sync after decode

* swiftui : sync after decode

* ggml : allow ggml_get_rows to use multiple threads if they are available

* check n_ubatch >= n_tokens with non-casual attention

* llama : do not limit n_batch to n_ctx with non-casual attn

* server : construct batch with size of llama_n_batch

* ggml_backend_cpu_graph_compute : fix return value when alloc fails

* llama : better n_batch and n_ubatch comment

* fix merge

* small fix

* reduce default n_batch to 2048

---------

Co-authored-by: Francis Couture-Harpin <git@compilade.net>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

2 files changed, 14 insertions, 3 deletions

diff --git a/common/common.cpp b/common/common.cpp
index 2f38ac63..73b1b61b 100644
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -483,6 +483,12 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.n_batch = std::stoi(argv[i]);
+        } else if (arg == "-ub" || arg == "--ubatch-size") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_ubatch = std::stoi(argv[i]);
         } else if (arg == "--keep") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -977,7 +983,9 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("                        binary file containing multiple choice tasks.\n");
     printf("  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
     printf("  -c N, --ctx-size N    size of the prompt context (default: %d, 0 = loaded from model)\n", params.n_ctx);
-    printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    printf("  -b N, --batch-size N  logical maximum batch size (default: %d)\n", params.n_batch);
+    printf("  -ub N, --ubatch-size N\n");
+    printf("                        physical maximum batch size (default: %d)\n", params.n_ubatch);
     printf("  --samplers            samplers that will be used for generation in the order, separated by \';\'\n");
     printf("                        (default: %s)\n", sampler_type_names.c_str());
     printf("  --sampling-seq        simplified sequence for samplers that will be used (default: %s)\n", sampler_type_chars.c_str());
@@ -1287,8 +1295,9 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     auto cparams = llama_context_default_params();
 
     cparams.n_ctx             = params.n_ctx;
-    cparams.n_batch           = params.n_batch;
     cparams.n_seq_max         = params.n_parallel;
+    cparams.n_batch           = params.n_batch;
+    cparams.n_ubatch          = params.n_ubatch;
     cparams.n_threads         = params.n_threads;
     cparams.n_threads_batch   = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
     cparams.seed              = params.seed;
@@ -1379,6 +1388,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
         std::vector<llama_token> tmp = { llama_token_bos(model), llama_token_eos(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);
     }
 
diff --git a/common/common.h b/common/common.h
index f8d82b87..0f178b9e 100644
--- a/common/common.h
+++ b/common/common.h
@@ -51,7 +51,8 @@ struct gpt_params {
     int32_t n_threads_batch_draft = -1;
     int32_t n_predict             = -1;    // new tokens to predict
     int32_t n_ctx                 = 512;   // context size
-    int32_t n_batch               = 512;   // batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_batch               = 2048;  // logical batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_ubatch              = 512;   // physical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                = 0;     // number of tokens to keep from initial prompt
     int32_t n_draft               = 5;     // number of tokens to draft during speculative decoding
     int32_t n_chunks              = -1;    // max number of chunks to process (-1 = unlimited)