implement automatic max ngl detection

common/common.cpp

@@ -836,7 +836,12 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
             invalid_param = true;
             return true;
         }
-        params.n_gpu_layers = std::stoi(argv[i]);
+        std::string argValue = argv[i];
+        if (argValue == "auto" || argValue == "a") {
+            params.n_gpu_layers = -2;
+        } else {
+            params.n_gpu_layers = std::stoi(argValue);
+        }
         if (!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");
@@ -1407,6 +1412,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     if (llama_supports_gpu_offload()) {
         printf("  -ngl N, --n-gpu-layers N\n");
         printf("                        number of layers to store in VRAM\n");
+        printf("                        set to 'auto' or 'a' to determine max automatically based on VRAM size\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");
@@ -2480,7 +2486,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "model: %s # default: models/7B/ggml-model.bin\n", params.model.c_str());
     fprintf(stream, "model_draft: %s # default:\n", params.model_draft.c_str());
     fprintf(stream, "multiline_input: %s # default: false\n", params.multiline_input ? "true" : "false");
-    fprintf(stream, "n_gpu_layers: %d # default: -1\n", params.n_gpu_layers);
+    fprintf(stream, "n_gpu_layers: %d # default: -1, auto: -2\n", params.n_gpu_layers);
     fprintf(stream, "n_predict: %d # default: -1 (unlimited)\n", params.n_predict);
     fprintf(stream, "n_probs: %d # only used by server binary, default: 0\n", sparams.n_probs);
     fprintf(stream, "no_mmap: %s # default: false\n", !params.use_mmap ? "true" : "false");

common/common.h

@@ -62,7 +62,7 @@ struct gpt_params {
     int32_t n_parallel            = 1;     // number of parallel sequences to decode
     int32_t n_sequences           = 1;     // number of sequences to decode
     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          = -1;    // number of layers to store in VRAM (-1 - use default, -2 - automatically determine)
     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_MODE_LAYER; // how to split the model across GPUs
     int32_t main_gpu              = 0;     // the GPU that is used for scratch and small tensors

ggml-cuda.cu

@@ -2612,6 +2612,11 @@ GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, si
     CUDA_CHECK(cudaMemGetInfo(free, total));
 }
 
+GGML_CALL void ggml_backend_cuda_get_free_device_memory(int device, size_t * free) {
+    size_t total;
+    ggml_backend_cuda_get_device_memory(device, free, &total);
+}
+
 GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size) {
     if (getenv("GGML_CUDA_REGISTER_HOST") == nullptr) {
         return false;

ggml-cuda.h

@@ -34,6 +34,7 @@ GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type
 GGML_API GGML_CALL int  ggml_backend_cuda_get_device_count(void);
 GGML_API GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
 GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
+GGML_API GGML_CALL void ggml_backend_cuda_get_free_device_memory(int device, size_t * free);
 
 GGML_API GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
 GGML_API GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer);

llama.cpp

@@ -1648,6 +1648,28 @@ static size_t llama_get_device_memory(int device) {
 #endif
 }
 
+// TODO: implement for other backends to return free memory
+static size_t llama_get_available_device_memory(int device) {
+#if defined(GGML_USE_CUDA)
+    size_t free;
+    ggml_backend_cuda_get_free_device_memory(device, &free);
+    return free;
+#elif defined(GGML_USE_SYCL)
+    size_t total;
+    size_t free;
+    ggml_backend_sycl_get_device_memory(device, &total, &free);
+    return free;
+#elif defined(GGML_USE_VULKAN)
+    size_t total;
+    size_t free;
+    ggml_backend_vk_get_device_memory(device, &total, &free);
+    return free;
+#else
+    return 1;
+    GGML_UNUSED(device);
+#endif
+}
+
 //
 // globals
 //
@@ -4327,6 +4349,32 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
     if (vocab.linefeed_id    != -1) { LLAMA_LOG_INFO( "%s: LF token         = %d '%s'\n", __func__, vocab.linefeed_id,    vocab.id_to_token[vocab.linefeed_id].text.c_str() );    }
 }
 
+static int llm_determine_max_ngl(const llama_model_loader & ml, const llama_model & model, const int main_gpu) {
+    const auto & hparams = model.hparams;
+
+    size_t available_gpu_memory = llama_get_available_device_memory(main_gpu);
+
+    // TODO: This is a rough, pretty inaccurate estimate, should implement using existing layer size and not guesstimating
+    size_t model_size = ml.n_bytes;
+    int32_t model_layers = hparams.n_layer;
+    size_t memory_per_layer = model_size / model_layers;
+
+    // TODO: get buffer size dynamically
+    int32_t buf_size = 400 * MiB;
+    int32_t buf_size_k = 200 * MiB;
+    int32_t buf_size_v = 200 * MiB;
+
+    int32_t total_buf_size = buf_size + buf_size_k + buf_size_v;
+
+    available_gpu_memory = available_gpu_memory - hparams.n_ctx_train; // context size
+    available_gpu_memory = available_gpu_memory - total_buf_size; // buffer size
+
+    // Calculate the maximum number of layers that can fit into the GPU memory
+    int32_t max_ngl = std::floor(static_cast<float>(available_gpu_memory) / memory_per_layer);
+
+    return max_ngl;
+}
+
 // Returns false if cancelled by progress_callback
 static bool llm_load_tensors(
         llama_model_loader & ml,
@@ -4342,6 +4390,11 @@ static bool llm_load_tensors(
 
     auto & hparams = model.hparams;
 
+    if (n_gpu_layers == -2) {
+        n_gpu_layers = llm_determine_max_ngl(ml, model, main_gpu);
+        LLAMA_LOG_INFO("%s: automatically set n_gpu_layers = %d\n", __func__, n_gpu_layers);
+    }
+
     model.split_mode   = split_mode;
     model.main_gpu     = main_gpu;
     model.n_gpu_layers = n_gpu_layers;