llama : ggml-backend integration

common/common.cpp

@@ -543,9 +543,8 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
             params.n_gpu_layers = std::stoi(argv[i]);
-#else
+#ifndef LLAMA_SUPPORTS_GPU_OFFLOAD
             fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n");
             fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
 #endif
@@ -554,9 +553,8 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
             params.n_gpu_layers_draft = std::stoi(argv[i]);
-#else
+#ifndef LLAMA_SUPPORTS_GPU_OFFLOAD
             fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers-draft option will be ignored\n");
             fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
 #endif
@@ -565,40 +563,53 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-#ifdef GGML_USE_CUBLAS
             params.main_gpu = std::stoi(argv[i]);
-#else
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.\n");
-#endif
+#ifndef GGML_USE_CUBLAS
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting the main GPU has no effect.\n");
+#endif // GGML_USE_CUBLAS
+        } else if (arg == "--split-mode" || arg == "-sm") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            std::string arg_next = argv[i];
+            if (arg_next == "none") {
+                params.split_mode = LLAMA_SPLIT_NONE;
+            } else if (arg_next == "layer") {
+                params.split_mode = LLAMA_SPLIT_LAYER;
+            } else if (arg_next == "row") {
+                params.split_mode = LLAMA_SPLIT_ROW;
+            } else {
+                invalid_param = true;
+                break;
+            }
+#ifndef GGML_USE_CUBLAS
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting the split mode has no effect.\n");
+#endif // GGML_USE_CUBLAS
         } else if (arg == "--tensor-split" || arg == "-ts") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-#ifdef GGML_USE_CUBLAS
             std::string arg_next = argv[i];
 
             // split string by , and /
             const std::regex regex{R"([,/]+)"};
             std::sregex_token_iterator it{arg_next.begin(), arg_next.end(), regex, -1};
             std::vector<std::string> split_arg{it, {}};
-            GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES);
-
+            if (split_arg.size() >= LLAMA_MAX_DEVICES) {
+                invalid_param = true;
+                break;
+            }
             for (size_t i = 0; i < LLAMA_MAX_DEVICES; ++i) {
                 if (i < split_arg.size()) {
                     params.tensor_split[i] = std::stof(split_arg[i]);
                 } else {
                     params.tensor_split[i] = 0.0f;
                 }
             }
-#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 == "--no-mul-mat-q" || arg == "-nommq") {
-#ifdef GGML_USE_CUBLAS
-            params.mul_mat_q = false;
-#else
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Disabling mul_mat_q kernels has no effect.\n");
+#ifndef GGML_USE_CUBLAS
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting a tensor split has no effect.\n");
 #endif // GGML_USE_CUBLAS
         } else if (arg == "--no-mmap") {
             params.use_mmap = false;
@@ -909,14 +920,15 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("                        number of layers to store in VRAM\n");
     printf("  -ngld N, --n-gpu-layers-draft N\n");
     printf("                        number of layers to store in VRAM for the draft model\n");
+    printf("  -sm SPLIT_MODE, --split-mode SPLIT_MODE\n");
+    printf("                        how to split the model across multiple GPUs, one of:\n");
+    printf("                          - none: use one GPU only\n");
+    printf("                          - layer (default): split layers and KV across GPUs\n");
+    printf("                          - row: split rows across GPUs\n");
     printf("  -ts SPLIT --tensor-split SPLIT\n");
-    printf("                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    printf("  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
-#ifdef GGML_USE_CUBLAS
-    printf("  -nommq, --no-mul-mat-q\n");
-    printf("                        use " GGML_CUBLAS_NAME " instead of custom mul_mat_q " GGML_CUDA_NAME " kernels.\n");
-    printf("                        Not recommended since this is both slower and uses more VRAM.\n");
-#endif // GGML_USE_CUBLAS
+    printf("                        fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1\n");
+    printf("  -mg i, --main-gpu i   the GPU to use for the model (with split-mode = none),\n");
+    printf("                        or for intermediate results and KV (with split-mode = row) (default: %d)\n", params.main_gpu);
 #endif
     printf("  -gan N, --grp-attn-n N\n");
     printf("                        group-attention factor (default: %d)\n", params.grp_attn_n);
@@ -1033,6 +1045,7 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
         mparams.n_gpu_layers = params.n_gpu_layers;
     }
     mparams.main_gpu        = params.main_gpu;
+    mparams.split_mode      = params.split_mode;
     mparams.tensor_split    = params.tensor_split;
     mparams.use_mmap        = params.use_mmap;
     mparams.use_mlock       = params.use_mlock;

common/common.h

@@ -59,6 +59,7 @@ struct gpt_params {
     float   p_split                         = 0.1f;  // speculative decoding split probability
     int32_t n_gpu_layers                    = -1;    // number of layers to store in VRAM (-1 - use default)
     int32_t n_gpu_layers_draft              = -1;    // number of layers to store in VRAM for the draft model (-1 - use default)
+    llama_split_mode split_mode             = LLAMA_SPLIT_LAYER; // how to split the model across GPUs
     int32_t main_gpu                        = 0;     // the GPU that is used for scratch and small tensors
     float   tensor_split[LLAMA_MAX_DEVICES] = {0};   // how split tensors should be distributed across GPUs
     int32_t n_beams                         = 0;     // if non-zero then use beam search of given width.

examples/batched-bench/batched-bench.cpp

@@ -88,7 +88,10 @@ int main(int argc, char ** argv) {
 
     llama_model_params model_params = llama_model_default_params();
 
+    const std::vector<float> t_split (LLAMA_MAX_DEVICES, 0.0f);
+
     model_params.n_gpu_layers = n_gpu_layers;
+    model_params.tensor_split = t_split.data();
 
     llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 

ggml-alloc.c

@@ -229,6 +229,7 @@ void ggml_tallocr_reset(ggml_tallocr_t alloc) {
         alloc->free_blocks[0].size = SIZE_MAX/2; // restrict maximum size of a measure allocator to half size_t max to avoid overflows
     } else {
         alloc->free_blocks[0].size = ggml_backend_buffer_get_size(alloc->buffer) - align_offset;
+        ggml_backend_buffer_reset(alloc->buffer);
     }
 }
 
@@ -779,10 +780,21 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
 
     if (nbytes == 0) {
         // all the tensors in the context are already allocated
+#ifndef NDEBUG
+        fprintf(stderr, "%s: all tensors in the context are already allocated\n", __func__);
+#endif
         return NULL;
     }
 
     ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, nbytes);
+    if (buffer == NULL) {
+        // failed to allocate buffer
+#ifndef NDEBUG
+        fprintf(stderr, "%s: failed to allocate buffer\n", __func__);
+#endif
+        return NULL;
+    }
+
     ggml_tallocr_t tallocr = ggml_tallocr_new_from_buffer(buffer);
 
     for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {

ggml-backend-impl.h

@@ -16,9 +16,10 @@ extern "C" {
     typedef void * ggml_backend_buffer_type_context_t;
 
     struct ggml_backend_buffer_type_i {
+        const char *          (*get_name)        (ggml_backend_buffer_type_t buft);
         ggml_backend_buffer_t (*alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
         size_t                (*get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
-        size_t                (*get_alloc_size)  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
+        size_t                (*get_alloc_size)  (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
         bool                  (*supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
         // check if tensor data is in host memory
         // should be equivalent to supports_backend(buft, ggml_backend_cpu_init())
@@ -34,23 +35,25 @@ extern "C" {
     typedef void * ggml_backend_buffer_context_t;
 
     struct ggml_backend_buffer_i {
-        void   (*free_buffer)    (ggml_backend_buffer_t buffer);
-        //void     (*reset)      (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
-        void * (*get_base)       (ggml_backend_buffer_t buffer);
-        void   (*init_tensor)    (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-        void   (*set_tensor)     (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void   (*get_tensor)     (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        const char * (*get_name)       (ggml_backend_buffer_t buffer);
+        void         (*free_buffer)    (ggml_backend_buffer_t buffer);
+        void *       (*get_base)       (ggml_backend_buffer_t buffer);
+        void         (*init_tensor)    (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+        void         (*set_tensor)     (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void         (*get_tensor)     (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
         // (optional) copy tensor between different buffer-type, allow for single-copy tranfers
-        void   (*cpy_tensor_from)(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void   (*cpy_tensor_to)  (ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void   (*clear)          (ggml_backend_buffer_t buffer, uint8_t value);
+        void         (*cpy_tensor_from)(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst);
+        void         (*cpy_tensor_to)  (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst);
+        void         (*clear)          (ggml_backend_buffer_t buffer, uint8_t value);
+        void         (*reset)          (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
     };
 
     struct ggml_backend_buffer {
         struct ggml_backend_buffer_i  iface;
         ggml_backend_buffer_type_t    buft;
         ggml_backend_buffer_context_t context;
         size_t size;
+        enum ggml_backend_buffer_usage usage;
     };
 
     ggml_backend_buffer_t ggml_backend_buffer_init(
@@ -79,13 +82,13 @@ extern "C" {
         void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
 
         // (optional) asynchroneous tensor copy
-        void (*cpy_tensor_from_async)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void (*cpy_tensor_to_async)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+        void (*cpy_tensor_from_async)(ggml_backend_t backend, const struct ggml_tensor * src, struct ggml_tensor * dst);
+        void (*cpy_tensor_to_async)  (ggml_backend_t backend, const struct ggml_tensor * src, struct ggml_tensor * dst);
 
         void (*synchronize)(ggml_backend_t backend);
 
         // compute graph with a plan
-        ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+        ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
         void                      (*graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
         void                      (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);