Implement automatic 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 estimate 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 - determine automatically)
     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);

ggml-sycl.cpp

@@ -16022,6 +16022,17 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
 
+GGML_CALL void ggml_backend_sycl_get_free_device_memory(int device, size_t *free) try {
+    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_get_free_device_memory\n");
+    size_t total;
+    ggml_backend_sycl_get_device_memory(device, free, &total);
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
 ////////////////////////////////////////////////////////////////////////////////
 
 // backend interface

ggml-sycl.h

@@ -33,6 +33,7 @@ GGML_API GGML_CALL void   ggml_sycl_get_gpu_list(int *id_list, int max_len);
 GGML_API GGML_CALL void   ggml_sycl_get_device_description(int device, char *description, size_t description_size);
 GGML_API GGML_CALL int   ggml_backend_sycl_get_device_count();
 GGML_API GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
+GGML_API GGML_CALL void ggml_backend_sycl_get_free_device_memory(int device, size_t *free);
 GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id);
 
 // TODO: these are temporary

ggml-vulkan.cpp

@@ -5781,6 +5781,21 @@ GGML_CALL void ggml_backend_vk_get_device_memory(int device, size_t * free, size
     }
 }
 
+GGML_CALL void ggml_backend_vk_get_free_device_memory(int device, size_t * free) {
+    GGML_ASSERT(device < (int) vk_instance.device_indices.size());
+
+    vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]];
+
+    vk::PhysicalDeviceMemoryProperties memprops = vkdev.getMemoryProperties();
+
+    for (const vk::MemoryHeap& heap : memprops.memoryHeaps) {
+        if (heap.flags & vk::MemoryHeapFlagBits::eDeviceLocal) {
+            *free = heap.size;
+            break;
+        }
+    }
+}
+
 // backend registry
 GGML_CALL static ggml_backend_t ggml_backend_reg_vk_init(const char * params, void * user_data) {
     ggml_backend_t vk_backend = ggml_backend_vk_init((int) (intptr_t) user_data);

ggml-vulkan.h

@@ -19,6 +19,7 @@ GGML_API GGML_CALL bool ggml_backend_is_vk(ggml_backend_t backend);
 GGML_API GGML_CALL int  ggml_backend_vk_get_device_count(void);
 GGML_API GGML_CALL void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
 GGML_API GGML_CALL void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
+GGML_API GGML_CALL void ggml_backend_vk_get_free_device_memory(int device, size_t * free);
 
 GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU

llama.cpp

@@ -1650,6 +1650,28 @@ 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_free_device_memory(device, &total, &free);
+    return free;
+#elif defined(GGML_USE_VULKAN)
+    size_t total;
+    size_t free;
+    ggml_backend_vk_get_free_device_memory(device, &total, &free);
+    return free;
+#else
+    return 1;
+    GGML_UNUSED(device);
+#endif
+}
+
 //
 // globals
 //
@@ -3254,6 +3276,17 @@ struct llama_model_loader {
         return cur;
     }
 
+    size_t estimate_tensor_bytes(const std::string & name, const std::vector<int64_t> & ne, bool required = true) const {
+        const struct ggml_tensor * cur = check_tensor_dims(name, ne, required);
+
+        if (cur == NULL) {
+            return 0;
+        }
+
+        int64_t nbytes = ggml_nbytes(cur);
+        return nbytes;
+    }
+
     struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::vector<int64_t> & ne, bool required = true) {
         const struct ggml_tensor * cur = check_tensor_dims(name, ne, required);
 
@@ -4329,6 +4362,90 @@ 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, enum llama_split_mode split_mode) {
+    const auto & hparams = model.hparams;
+
+    // could become negative - use signed size_t
+    ssize_t available_gpu_memory = 0;
+    int n_layer = hparams.n_layer;
+
+    if (split_mode == LLAMA_SPLIT_MODE_LAYER) {
+        // veeery sketchy, there has to be a better way to do this
+        int deice_count = llama_get_device_count();
+        for (int i = 0; i < deice_count; ++i) {
+            available_gpu_memory += llama_get_available_device_memory(i);
+        }
+    } else {
+        available_gpu_memory = llama_get_available_device_memory(main_gpu);
+    }
+
+    // "avoid a scenario where an application ooms because llama.cpp only left 5 MB of VRAM" - https://github.com/ggerganov/llama.cpp/pull/6502#discussion_r1555060962
+    available_gpu_memory -= 50 * MiB;
+
+    // determine tensor sizes
+    size_t total_size = 0;
+    size_t extra_size = 0;
+    std::vector<size_t> layer_sizes(n_layer);
+    for (int i = 0; i < ml.n_tensors; ++i) {
+        const struct ggml_tensor * tensor = (const struct ggml_tensor *) ml.get_tensor_meta(i);
+        std::vector<int64_t> tensor_ne = { tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3] };
+        size_t tensor_size = ml.estimate_tensor_bytes(tensor->name, tensor_ne);
+        //LLAMA_LOG_INFO("%s: %-30s: %12ld\n", __func__, tensor->name, tensor_size);
+        // all layer specific tensor names have the prefix "blk."
+        if (strncmp(tensor->name, "blk.", 4) == 0) {
+            int layer_no;
+            sscanf(tensor->name+4, "%d", &layer_no);
+            layer_sizes[layer_no] += tensor_size;
+        } else {
+            extra_size += tensor_size;
+        }
+        total_size += tensor_size;
+    }
+
+    // TODO: get buffer size dynamically
+    size_t buf_size = 400 * MiB;
+    size_t buf_size_k = 200 * MiB;
+    size_t buf_size_v = 200 * MiB;
+
+    ssize_t buffer_size = buf_size + buf_size_k + buf_size_v;
+
+    // can/will be pretty big for large models
+    size_t n_ctx = hparams.n_ctx_train;
+
+    size_t ctx_size =
+            ggml_tensor_overhead()*(ml.n_tensors + 1) +
+            ggml_tensor_overhead()*hparams.n_expert*n_layer;
+
+    size_t context_size = n_ctx*ctx_size;
+
+    // Calculate the maximum number of layers that can fit into the available GPU memory
+    int max_ngl = 0;
+    ssize_t used_memory = extra_size+buffer_size+context_size;
+    for (int i = 0; i < n_layer; ++i) {
+        LLAMA_LOG_INFO("%s: layer %2d size:        %12ld\n", __func__, i, layer_sizes[i]);
+        used_memory += layer_sizes[i];
+        if (used_memory > available_gpu_memory) {
+            break;
+        }
+        max_ngl++;
+    }
+
+    LLAMA_LOG_INFO("%s: extra size:           %12ld\n", __func__, extra_size);
+    LLAMA_LOG_INFO("%s: ----------------------------------\n", __func__);
+    LLAMA_LOG_INFO("%s: total size:           %12ld\n", __func__, total_size);
+    LLAMA_LOG_INFO("%s: available_gpu_memory: %12ld\n", __func__, available_gpu_memory);
+    LLAMA_LOG_INFO("%s: buffer size:          %12ld\n", __func__, buffer_size);
+    LLAMA_LOG_INFO("%s: total context size:   %12ld\n", __func__, context_size);
+    LLAMA_LOG_INFO("%s: used memory:          %12ld\n", __func__, used_memory);
+    LLAMA_LOG_INFO("%s: ----------------------------------\n", __func__);
+    LLAMA_LOG_INFO("%s: tokens in context:      %ld\n",    __func__, n_ctx);
+    LLAMA_LOG_INFO("%s: per token context size: %ld\n", __func__, ctx_size);
+    LLAMA_LOG_INFO("%s: layer_count:            %d\n",    __func__, n_layer);
+    LLAMA_LOG_INFO("%s: max_ngl:                %d\n",    __func__, max_ngl);
+
+    return max_ngl;
+}
+
 // Returns false if cancelled by progress_callback
 static bool llm_load_tensors(
         llama_model_loader & ml,
@@ -4344,6 +4461,11 @@ 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, split_mode);
+        LLAMA_LOG_INFO("%s: automatically set n_gpu_layers to %d\n", __func__, n_gpu_layers);
+    }
+
     model.split_mode   = split_mode;
     model.main_gpu     = main_gpu;
     model.n_gpu_layers = n_gpu_layers;