server: add dedup-cache-models preset option (#27346)

This commit is contained in:
Xuan-Son Nguyen
2026-08-19 11:04:26 +02:00
committed by GitHub
parent 98d1e92c21
commit ee4c505a4f
8 changed files with 132 additions and 2 deletions
+6
View File
@@ -4658,6 +4658,12 @@ void common_params_add_preset_options(std::vector<common_arg> & args) {
[](common_params &, int) { /* unused */ }
).set_env(COMMON_ARG_PRESET_STOP_TIMEOUT).set_preset_only());
args.push_back(common_arg(
{"dedup-cache-models"}, "0|1",
"in server router mode, hide a cached model from the model list when this preset resolves to the same model file",
[](common_params &, const std::string &) { /* unused */ }
).set_env(COMMON_ARG_PRESET_DEDUP_CACHE_MODELS).set_preset_only());
// args.push_back(common_arg(
// {"pin"},
// "in server router mode, do not unload this model if models_max is exceeded",
+3 -2
View File
@@ -11,8 +11,9 @@
#include <memory>
// pseudo-env variable to identify preset-only arguments
#define COMMON_ARG_PRESET_LOAD_ON_STARTUP "__PRESET_LOAD_ON_STARTUP"
#define COMMON_ARG_PRESET_STOP_TIMEOUT "__PRESET_STOP_TIMEOUT"
#define COMMON_ARG_PRESET_LOAD_ON_STARTUP "__PRESET_LOAD_ON_STARTUP"
#define COMMON_ARG_PRESET_STOP_TIMEOUT "__PRESET_STOP_TIMEOUT"
#define COMMON_ARG_PRESET_DEDUP_CACHE_MODELS "__PRESET_DEDUP_CACHE_MODELS"
//
// CLI argument parsing
+20
View File
@@ -989,6 +989,26 @@ std::vector<common_cached_model_info> common_list_cached_models() {
return result;
}
std::string common_download_resolve_path(const std::string & hf_repo_with_tag, const std::string & hf_file) {
auto [repo, tag] = common_download_split_repo_tag(hf_repo_with_tag);
auto files = hf_cache::get_cached_files(repo);
if (files.empty()) {
return "";
}
if (!hf_file.empty()) {
for (const auto & f : files) {
if (f.path == hf_file) {
return f.local_path;
}
}
return "";
}
return find_best_model(files, tag).local_path;
}
bool common_download_remove(const std::string & hf_repo_with_tag) {
namespace fs = std::filesystem;
+4
View File
@@ -85,6 +85,10 @@ std::vector<std::string> common_download_get_all_parts(const std::string & url);
// returns list of cached models
std::vector<common_cached_model_info> common_list_cached_models();
// resolve the local cached file path for a HF repo without network access (hf_file, if given, must match exactly)
// returns an empty string if the model is not present in the cache
std::string common_download_resolve_path(const std::string & hf_repo_with_tag, const std::string & hf_file = "");
// download single file from url to local path
// returns status code or -1 on error
// skip_etag: if true, don't read/write .etag files (for HF cache where filename is the hash)