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

llama-fit-params: keep explicit --ctx-size 0 (#19070)

This commit is contained in:
Johannes Gäßler 2026-01-24 22:13:08 +01:00 committed by GitHub
parent 4e5b83b226
commit e9fd8dcab4
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
4 changed files with 12 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
common/arg.cpp
View File

@ -1231,6 +1231,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
string_format("size of the prompt context (default: %d, 0 = loaded from model)", params.n_ctx), string_format("size of the prompt context (default: %d, 0 = loaded from model)", params.n_ctx),
[](common_params & params, int value) { [](common_params & params, int value) {
params.n_ctx = value; params.n_ctx = value;
if (value == 0) {
// disable context reduction in llama_params_fit if the user explicitly requests the full context size:
params.fit_params_min_ctx = UINT32_MAX;
}
} }
).set_env("LLAMA_ARG_CTX_SIZE")); ).set_env("LLAMA_ARG_CTX_SIZE"));
add_opt(common_arg( add_opt(common_arg(

1
include/llama.h
View File

@ -489,6 +489,7 @@ extern "C" {
// - returns true if the parameters could be successfully modified to fit device memory // - returns true if the parameters could be successfully modified to fit device memory
// - this function is NOT thread safe because it modifies the global llama logger state // - this function is NOT thread safe because it modifies the global llama logger state
// - only parameters that have the same value as in llama_default_model_params are modified // - only parameters that have the same value as in llama_default_model_params are modified
// with the exception of the context size which is modified if and only if equal to 0
LLAMA_API enum llama_params_fit_status llama_params_fit( LLAMA_API enum llama_params_fit_status llama_params_fit(
const char * path_model, const char * path_model,
struct llama_model_params * mparams, struct llama_model_params * mparams,

8
src/llama.cpp
View File

@ -311,8 +311,12 @@ static void llama_params_fit_impl(
__func__, hp_nct, cparams->n_ctx, memory_reduction/MiB); __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
} }
} else { } else {
LLAMA_LOG_INFO("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n", if (n_ctx_min == UINT32_MAX) {
__func__, hp_nct, n_ctx_min); LLAMA_LOG_INFO("%s: user has requested full context size of %" PRIu32 " -> no change\n", __func__, hp_nct);
} else {
LLAMA_LOG_INFO("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
__func__, hp_nct, n_ctx_min);
}
} }
} else { } else {
LLAMA_LOG_INFO("%s: context size set by user to %" PRIu32 " -> no change\n", __func__, cparams->n_ctx); LLAMA_LOG_INFO("%s: context size set by user to %" PRIu32 " -> no change\n", __func__, cparams->n_ctx);

2
tools/fit-params/fit-params.cpp
View File

@ -36,7 +36,7 @@ int main(int argc, char ** argv) {
LOG_INF("%s: printing fitted CLI arguments to stdout...\n", __func__); LOG_INF("%s: printing fitted CLI arguments to stdout...\n", __func__);
common_log_flush(common_log_main()); common_log_flush(common_log_main());
printf("-c %" PRIu32 " -ngl %" PRIu32, cparams.n_ctx, mparams.n_gpu_layers); printf("-c %" PRIu32 " -ngl %" PRIi32, cparams.n_ctx, mparams.n_gpu_layers);
size_t nd = llama_max_devices(); size_t nd = llama_max_devices();
while (nd > 1 && mparams.tensor_split[nd - 1] == 0.0f) { while (nd > 1 && mparams.tensor_split[nd - 1] == 0.0f) {