LLM inference in C/C++
C++ ggml
spec: single device drafter should create meta backend wrapper (#28390)
Aman Gupta415e909d8433
diff --git a/common/arg.cpp b/common/arg.cppindex 015196ca14..74241f9312 100644--- a/common/arg.cpp+++ b/common/arg.cpp@@ -894,6 +894,12 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_contextpostprocess_cpu_params(params.speculative.draft.cpuparams, ¶ms.cpuparams);postprocess_cpu_params(params.speculative.draft.cpuparams_batch, ¶ms.cpuparams_batch);+ // default the mmproj device to the global device selection if not set explicitly with -mmdev+ if (params.mmproj_use_gpu && params.mmproj_device == nullptr && !params.devices.empty()) {+ params.mmproj_device = params.devices.front();+ params.mmproj_use_gpu = params.mmproj_device != nullptr;+ }+if (params.prompt_cache_all && (params.interactive || params.interactive_first)) {throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");}@@ -2610,7 +2616,7 @@ common_params_context common_params_parser_init(common_params & params, llama_exadd_opt(common_arg(// note: "-mmdev" must sort after "--rpc" in the preset map, else RPC devices are not registered yet{"-mmdev", "--mmproj-device"}, "DEVICE",- "device to use for multimodal projector (none = don't offload, default: auto)\n"+ "device to use for multimodal projector (none = don't offload, default: follows --device)\n""use --list-devices to see a list of available devices",[](common_params & params, const std::string & value) {if (value == "none") {@@ -4229,7 +4235,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_DRAFT_BACKEND_SAMPLING"));add_opt(common_arg({"--spec-draft-device", "-devd", "--device-draft"}, "<dev1,dev2,..>",- "comma-separated list of devices to use for offloading the draft model (none = don't offload)\n"+ "comma-separated list of devices to use for offloading the draft model (none = don't offload, default: follows --device)\n""use --list-devices to see a list of available devices",[](common_params & params, const std::string & value) {params.speculative.draft.devices = parse_device_list(value);diff --git a/common/speculative.cpp b/common/speculative.cppindex 851a47b9a5..2db381d580 100644--- a/common/speculative.cpp+++ b/common/speculative.cpp@@ -2467,11 +2467,22 @@ common_params common_base_params_to_speculative(const common_params & params) {result.pooling_type = LLAMA_POOLING_TYPE_UNSPECIFIED;if (has_draft) {- result.devices = params_spec.devices;+ // default to global devices value+ if (!params_spec.devices.empty()) {+ result.devices = params_spec.devices;+ }result.model = params_spec.mparams;result.n_gpu_layers = params_spec.n_gpu_layers;result.tensor_buft_overrides = params_spec.tensor_buft_overrides;+ // a draft pinned to a single device doesn't need the meta wrapper an inherited -sm tensor would give it+ // (the device list is null-terminated, so a single device means size 2)+ const size_t n_devs = std::count_if(params_spec.devices.begin(), params_spec.devices.end(),+ [](ggml_backend_dev_t d) { return d != nullptr; });+ if (n_devs == 1) {+ result.split_mode = LLAMA_SPLIT_MODE_LAYER;+ }+if (params_spec.cpuparams.n_threads > 0) {result.cpuparams.n_threads = params_spec.cpuparams.n_threads;result.cpuparams_batch.n_threads = params_spec.cpuparams_batch.n_threads;diff --git a/src/llama-context.cpp b/src/llama-context.cppindex c1ef12f56b..21501574a9 100644--- a/src/llama-context.cpp+++ b/src/llama-context.cpp@@ -3689,6 +3689,9 @@ llama_context * llama_init_from_model(LLAMA_LOG_ERROR("%s: SPLIT_MODE_TENSOR requires flash_attn to be enabled\n", __func__);return nullptr;}+ if (model->get_split_state_ud.n_devices == 1) {+ LLAMA_LOG_WARN("%s: SPLIT_MODE_TENSOR being used for a single device is not recommended\n", __func__);+ }}if ((model->hparams.is_mla() || model->arch == LLM_ARCH_DEEPSEEK4) && params.type_k != params.type_v) {diff --git a/tools/cli/README.md b/tools/cli/README.mdindex b667d341d4..efe653494d 100644--- a/tools/cli/README.md+++ b/tools/cli/README.md@@ -164,7 +164,7 @@| `-mmu, --mmproj-url URL` | URL to a multimodal projector file. see tools/mtmd/README.md<br/>(env: LLAMA_ARG_MMPROJ_URL) || `--mmproj-auto, --no-mmproj, --no-mmproj-auto` | whether to use multimodal projector file (if available), useful when using -hf (default: enabled)<br/>(env: LLAMA_ARG_MMPROJ_AUTO) || `--mmproj-offload, --no-mmproj-offload` | whether to enable GPU offloading for multimodal projector (default: enabled)<br/>(env: LLAMA_ARG_MMPROJ_OFFLOAD) |-| `-mmdev, --mmproj-device DEVICE` | device to use for multimodal projector (none = don't offload, default: auto)<br/>use --list-devices to see a list of available devices<br/>(env: MTMD_BACKEND_DEVICE) |+| `-mmdev, --mmproj-device DEVICE` | device to use for multimodal projector (none = don't offload, default: follows --device)<br/>use --list-devices to see a list of available devices<br/>(env: MTMD_BACKEND_DEVICE) || `--image, --audio, --video FILE` | path to an image, audio, or video file. use with multimodal models, use comma-separated values for multiple files || `--image-min-tokens N` | minimum number of tokens each image can take, only used by vision models with dynamic resolution (default: read from model)<br/>(env: LLAMA_ARG_IMAGE_MIN_TOKENS) || `--image-max-tokens N` | maximum number of tokens each image can take, only used by vision models with dynamic resolution (default: read from model)<br/>(env: LLAMA_ARG_IMAGE_MAX_TOKENS) |@@ -207,7 +207,7 @@| `--spec-draft-p-split, --draft-p-split P` | speculative decoding split probability (default: 0.10)<br/>(env: LLAMA_ARG_SPEC_DRAFT_P_SPLIT) || `--spec-draft-p-min, --draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.00)<br/>(env: LLAMA_ARG_SPEC_DRAFT_P_MIN) || `--spec-draft-backend-sampling, --no-spec-draft-backend-sampling` | offload draft sampling to the backend (default: enabled)<br/>(env: LLAMA_ARG_SPEC_DRAFT_BACKEND_SAMPLING) |-| `--spec-draft-device, -devd, --device-draft <dev1,dev2,..>` | comma-separated list of devices to use for offloading the draft model (none = don't offload)<br/>use --list-devices to see a list of available devices |+| `--spec-draft-device, -devd, --device-draft <dev1,dev2,..>` | comma-separated list of devices to use for offloading the draft model (none = don't offload, default: follows --device)<br/>use --list-devices to see a list of available devices || `--spec-draft-ngl, -ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | max. number of draft model layers to store in VRAM, either an exact number, 'auto', or 'all' (default: auto)<br/>(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) || `--spec-draft-model, -md, --model-draft FNAME` | draft model for speculative decoding (default: unused)<br/>(env: LLAMA_ARG_SPEC_DRAFT_MODEL) || `--spec-type none,draft-simple,draft-eagle3,draft-mtp,draft-dflash,draft-dspark,ngram-simple,ngram-map-k,ngram-map-k4v,ngram-mod,ngram-cache` | comma-separated list of types of speculative decoding to use (default: none)<br/><br/>(env: LLAMA_ARG_SPEC_TYPE) |diff --git a/tools/server/README.md b/tools/server/README.mdindex 952d31e753..71ebb95434 100644--- a/tools/server/README.md+++ b/tools/server/README.md@@ -182,7 +182,7 @@ For the full list of features, please refer to [server's changelog](https://gith| `-mmu, --mmproj-url URL` | URL to a multimodal projector file. see tools/mtmd/README.md<br/>(env: LLAMA_ARG_MMPROJ_URL) || `--mmproj-auto, --no-mmproj, --no-mmproj-auto` | whether to use multimodal projector file (if available), useful when using -hf (default: enabled)<br/>(env: LLAMA_ARG_MMPROJ_AUTO) || `--mmproj-offload, --no-mmproj-offload` | whether to enable GPU offloading for multimodal projector (default: enabled)<br/>(env: LLAMA_ARG_MMPROJ_OFFLOAD) |-| `-mmdev, --mmproj-device DEVICE` | device to use for multimodal projector (none = don't offload, default: auto)<br/>use --list-devices to see a list of available devices<br/>(env: MTMD_BACKEND_DEVICE) |+| `-mmdev, --mmproj-device DEVICE` | device to use for multimodal projector (none = don't offload, default: follows --device)<br/>use --list-devices to see a list of available devices<br/>(env: MTMD_BACKEND_DEVICE) || `--image-min-tokens N` | minimum number of tokens each image can take, only used by vision models with dynamic resolution (default: read from model)<br/>(env: LLAMA_ARG_IMAGE_MIN_TOKENS) || `--image-max-tokens N` | maximum number of tokens each image can take, only used by vision models with dynamic resolution (default: read from model)<br/>(env: LLAMA_ARG_IMAGE_MAX_TOKENS) || `--mtmd-batch-max-tokens N` | maximum number of image tokens per batch when encoding images (default: 1024)<br/>(env: LLAMA_ARG_MTMD_BATCH_MAX_TOKENS) |@@ -268,7 +268,7 @@ For the full list of features, please refer to [server's changelog](https://gith| `--spec-draft-p-split, --draft-p-split P` | speculative decoding split probability (default: 0.10)<br/>(env: LLAMA_ARG_SPEC_DRAFT_P_SPLIT) || `--spec-draft-p-min, --draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.00)<br/>(env: LLAMA_ARG_SPEC_DRAFT_P_MIN) || `--spec-draft-backend-sampling, --no-spec-draft-backend-sampling` | offload draft sampling to the backend (default: enabled)<br/>(env: LLAMA_ARG_SPEC_DRAFT_BACKEND_SAMPLING) |-| `--spec-draft-device, -devd, --device-draft <dev1,dev2,..>` | comma-separated list of devices to use for offloading the draft model (none = don't offload)<br/>use --list-devices to see a list of available devices |+| `--spec-draft-device, -devd, --device-draft <dev1,dev2,..>` | comma-separated list of devices to use for offloading the draft model (none = don't offload, default: follows --device)<br/>use --list-devices to see a list of available devices || `--spec-draft-ngl, -ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | max. number of draft model layers to store in VRAM, either an exact number, 'auto', or 'all' (default: auto)<br/>(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) || `--spec-draft-model, -md, --model-draft FNAME` | draft model for speculative decoding (default: unused)<br/>(env: LLAMA_ARG_SPEC_DRAFT_MODEL) || `--spec-type none,draft-simple,draft-eagle3,draft-mtp,draft-dflash,draft-dspark,ngram-simple,ngram-map-k,ngram-map-k4v,ngram-mod,ngram-cache` | comma-separated list of types of speculative decoding to use (default: none)<br/><br/>(env: LLAMA_ARG_SPEC_TYPE) |