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

Use more descriptive option naming

This commit is contained in:
Ed Addario 2025-11-24 18:26:45 +00:00
parent 7eb77147b9
commit 6616008420
No known key found for this signature in database
GPG Key ID: E7875815A3230993
3 changed files with 13 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
include/llama.h
View File

@ -369,7 +369,7 @@ extern "C" {
float target_bpw; // target bits per weight (bpw)
bool keep_bpw_state; // keep bpw state file
void * bpw_state; // pointer to bpw state file
bool disable_tensor_importance; // treat all tensors equally during quantization
bool no_importance; // allocate target bpw budget equitably across all tensors
} llama_model_quantize_params;
typedef struct llama_logit_bias {

10
src/llama-quant.cpp
View File

@ -1571,7 +1571,7 @@ static std::unordered_map<std::string, ggml_type> target_bpw_type(
// Certain tensors have a higher impact on model quality, so we apply a lower penalty to them
auto is_important = [&](const std::string & tensor_name) -> bool {
bool important = tensor_name == "output.weight";
if (!important && !params->disable_tensor_importance) {
if (!important && !params->no_importance) {
important = tensor_name.find(".attn_v.weight") != std::string::npos ||
tensor_name.find(".time_mix_value.weight") != std::string::npos ||
tensor_name.find(".ffn_down.weight") != std::string::npos ||
@ -2009,10 +2009,10 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
} else {
LLAMA_LOG_INFO("%s: imatrix does not have activations, process may be less accurate\n", __func__);
}
if (params->disable_tensor_importance) {
LLAMA_LOG_INFO("%s: allocating bpw budget to tensors equally\n", __func__);
if (params->no_importance) {
LLAMA_LOG_INFO("%s: distributing bpw budget equitably across all tensors\n", __func__);
} else {
LLAMA_LOG_INFO("%s: allocating more bpw budget to important tensors\n", __func__);
LLAMA_LOG_INFO("%s: assigning more bpw budget to important tensors\n", __func__);
}
LLAMA_LOG_INFO("%s: computing tensor quantization mix to achieve %.4f bpw\n", __func__, params->target_bpw);
@ -2281,7 +2281,7 @@ llama_model_quantize_params llama_model_quantize_default_params() {
/*.target_bpw =*/ -1.0f,
/*.keep_bpw_state =*/ false,
/*.bpw_state =*/ nullptr,
/*.disable_tensor_importance =*/ false
/*.no_importance =*/ false
};
return result;

14
tools/quantize/quantize.cpp
View File

@ -117,9 +117,9 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
[[noreturn]]
static void usage(const char * executable) {
printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] [--target-bpw n]\n", executable);
printf(" [--keep-bpw-state] [--bpw-state filename] [--output-tensor-type] [--token-embedding-type] [--tensor-type] [--prune-layers] [--keep-split] [--override-kv]\n");
printf(" model-f32.gguf [model-quant.gguf] type [nthreads]\n\n");
printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights]\n", executable);
printf(" [--target-bpw n] [--no-importance] [--keep-bpw-state] [--bpw-state filename] [--output-tensor-type] [--token-embedding-type] [--tensor-type]\n");
printf(" [--prune-layers] [--keep-split] [--override-kv] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n");
printf(" --allow-requantize: allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
printf(" --leave-output-tensor: will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
printf(" --pure: disable k-quant mixtures and quantize all tensors to the same type\n");
@ -134,8 +134,8 @@ static void usage(const char * executable) {
printf(" Advanced option to remove all tensors from the given layers\n");
printf(" --target-bpw: target bits per weight (bpw). Must be a positive number between 0.0 and 8.0\n");
printf(" Advanced option to automatically select quantization types to achieve a total bits per weight (bpw) target\n");
printf(" --disable-tensor-importance: treat all tensors equally during bpw quantization\n");
printf(" Advanced option to disable allocating more bpw budget to important tensors. It may increase quality for some models\n");
printf(" --no-importance: distribute bpw budget equitably across all tensors\n");
printf(" Advanced option to disable assigning more bpw budget to important tensors. It may increase quality for some models\n");
printf(" --keep-bpw-state: save the bpw computations to <architecture>-<model hash>.bpw_state\n");
printf(" --bpw-state: file name to use instead of default\n");
printf(" --keep-split: will generate quantized model in the same shards as input\n");
@ -562,8 +562,8 @@ int main(int argc, char ** argv) {
if (arg_idx == argc-1 || !parse_target_bpw(argv[++arg_idx], target_bpw)) {
usage(argv[0]);
}
} else if (strcmp(argv[arg_idx], "--disable-tensor-importance") == 0) {
params.disable_tensor_importance = true;
} else if (strcmp(argv[arg_idx], "--no-importance") == 0) {
params.no_importance = true;
} else if (strcmp(argv[arg_idx], "--keep-bpw-state") == 0) {
params.keep_bpw_state = true;
} else if (strcmp(argv[arg_idx], "--bpw-state") == 0) {