happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Refactor signal handlers

This commit is contained in:
Ed Addario 2025-10-13 14:54:32 +01:00
parent ca282302b5
commit b1b58e67df
No known key found for this signature in database
GPG Key ID: E7875815A3230993
1 changed files with 16 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

34
src/llama-quant.cpp
View File

@ -632,6 +632,22 @@ static std::unordered_map<std::string, ggml_type> target_bpw_type(
const llama_model_quantize_params * params, const llama_model_quantize_params * params,
int nthread int nthread
) { ) {
// RAII guard for signal handlers
bpw_stop.store(false, std::memory_order_relaxed);
struct signal_scope_guard {
using handler_t = void (*)(int);
handler_t prev_int = SIG_DFL;
handler_t prev_term = SIG_DFL;
signal_scope_guard() {
prev_int = std::signal(SIGINT, signal_handler);
prev_term = std::signal(SIGTERM, signal_handler);
}
~signal_scope_guard() {
std::signal(SIGINT, prev_int);
std::signal(SIGTERM, prev_term);
}
} _signal_guard;
struct candidate_types { struct candidate_types {
ggml_type type; ggml_type type;
float bpw; float bpw;
@ -724,22 +740,6 @@ static std::unordered_map<std::string, ggml_type> target_bpw_type(
return is_quantizable(ggml_get_name(t), model.arch, params); return is_quantizable(ggml_get_name(t), model.arch, params);
}; };
auto install_signal_handlers = [] {
static std::once_flag once;
std::call_once(once, [] {
std::signal(SIGINT, signal_handler);
std::signal(SIGTERM, signal_handler);
});
};
auto uninstall_signal_handlers = [] {
static std::once_flag once;
std::call_once(once, [] {
std::signal(SIGINT, SIG_DFL);
std::signal(SIGTERM, SIG_DFL);
});
};
// Saved state per tensor // Saved state per tensor
struct saved_info { struct saved_info {
std::vector<candidate_types> candidate; std::vector<candidate_types> candidate;
@ -1121,7 +1121,6 @@ static std::unordered_map<std::string, ggml_type> target_bpw_type(
return lambdas; return lambdas;
}; };
install_signal_handlers();
auto bpw_data = load_bpw_state(); auto bpw_data = load_bpw_state();
// Significantly reduce compute time by parallelising tensor processing - courtesy of https://github.com/ddh0 // Significantly reduce compute time by parallelising tensor processing - courtesy of https://github.com/ddh0
@ -1700,7 +1699,6 @@ static std::unordered_map<std::string, ggml_type> target_bpw_type(
} }
delete_bpw_state(); // we're done, clear any checkpoint delete_bpw_state(); // we're done, clear any checkpoint
uninstall_signal_handlers();
return emit_overrides(); return emit_overrides();
} }