upstream-archive/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-07-23 11:10:55 -05:00
Code Issues Packages Projects Releases Wiki Activity

common: refactor model handling (#24980)

Browse Source 
* common: refactor models handling

* remote preset

* cont

* rm skip_download option

* missing header

* fix plan.model_files

* fix --offline case

* move hf_plan to download

* refactor

* rm redundant curr_ex, add comments

* adapt
This commit is contained in:
Xuan-Son Nguyen
2026-06-25 15:17:51 +02:00
committed by GitHub
parent e8ecce53b8
commit 60bc8866b1
9 changed files with 289 additions and 327 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
tools/server/server-models.cpp
View File

@@ -223,8 +223,8 @@ void server_model_meta::update_caps() {
"LLAMA_ARG_HF_REPO_FILE",
});
params.offline = true;
// params.skip_download = true; // TODO: ideally, we should validate the model here, but it takes too much time
common_params_handle_models(params, LLAMA_EXAMPLE_SERVER, {});
common_models_handler handler = common_models_handler_init(params, LLAMA_EXAMPLE_SERVER);
common_models_handler_apply(handler, params); // note: this won't download the model because offline=true
if (params.mmproj.path.empty()) {
multimodal = { false, false };
} else {
@@ -1393,9 +1393,8 @@ struct server_download_state : public common_download_callback {
bool run(common_params & params) {
try {
common_params_handle_models_params p;
p.callback = this;
common_params_handle_models(params, LLAMA_EXAMPLE_SERVER, p);
common_models_handler handler = common_models_handler_init(params, LLAMA_EXAMPLE_SERVER);
common_models_handler_apply(handler, params, this);
is_ok = true;
} catch (const std::exception & e) {
auto model_name = params.model.get_name();
@@ -1768,23 +1767,14 @@ void server_models_routes::init_routes() {
throw std::invalid_argument("model must be a non-empty string");
}
common_params_model model;
common_download_opts opts;
common_params p;
p.model.hf_repo = name;
p.hf_token = params.hf_token;
model.hf_repo = name;
opts.bearer_token = params.hf_token;
// note: we only check main model, no need sidecar here
opts.download_mmproj = false;
opts.download_mtp = false;
// first, only check if the model is valid and can be downloaded
opts.skip_download = true;
// validate by fetching metadata
bool ok = false;
try {
auto validation = common_download_model(model, opts);
ok = !validation.model_path.empty();
} catch (const common_skip_download_exception &) {
// model is valid and will be downloaded
common_models_handler_init(p, LLAMA_EXAMPLE_SERVER);
ok = true;
} catch (...) {
SRV_ERR("unknown error while validating model '%s'\n", name.c_str());

24
tools/server/server.cpp
View File

@@ -89,15 +89,16 @@ int llama_server(int argc, char ** argv) {
llama_backend_init();
llama_numa_init(params.numa);
// note: router mode also accepts -hf remote-preset, so we need to check that first
if (!params.model.hf_repo.empty()) {
try {
common_params_handle_models_params handle_params;
handle_params.preset_only = true;
common_params_handle_models(params, LLAMA_EXAMPLE_SERVER, handle_params);
} catch (const std::exception & e) {
// ignored for now
common_models_handler models_handler;
try {
models_handler = common_models_handler_init(params, LLAMA_EXAMPLE_SERVER);
if (common_models_handler_is_preset_repo(models_handler)) {
// apply the preset and start the server in router mode
common_models_handler_apply(models_handler, params);
}
} catch (const std::exception & e) {
SRV_ERR("failed to fetch model metadata: %s\n", e.what());
return 1;
}
// router server never loads a model and must not touch the GPU
@@ -274,7 +275,12 @@ int llama_server(int argc, char ** argv) {
return child.run_download(params);
} else if (!is_router_server) {
// single-model mode (NOT spawned by router)
common_params_handle_models(params, LLAMA_EXAMPLE_SERVER, {});
try {
common_models_handler_apply(models_handler, params);
} catch (const std::exception & e) {
SRV_ERR("failed to download model: %s\n", e.what());
return 1;
}
}
//