Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

spec: single device drafter should create meta backend wrapper (#28390)

Browse Source
This commit is contained in:
Aman Gupta
2026-09-08 20:44:33 +08:00
committed by GitHub
parent 03fa73cb27
commit 415e909d84
5 changed files with 27 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
common/arg.cpp
+8 -2
View File
@@ -894,6 +894,12 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
postprocess_cpu_params(params.speculative.draft.cpuparams, &params.cpuparams);
postprocess_cpu_params(params.speculative.draft.cpuparams_batch, &params.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_ex
add_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);
common/speculative.cpp
+12 -1
View File
@@ -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;