vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Allow quantize to only copy tensors, some other improvements (#2931)

Browse source 
* Allow quantize tool to only copy tensors to allow repackaging models.

* Slightly better logic when requantizing.

* Change help message to go to `stdout`.
This commit is contained in:
Kerfuffle 2023-09-01 08:02:48 -06:00 • committed by GitHub
parent 0d58936686
commit 5d6f19f16b
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
3 changed files with 37 additions and 13 deletions
Download patch file Download diff file Expand all files Collapse all files

25
llama.cpp
View file

@ -4683,6 +4683,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
llm_load_arch(*ml, model);
llm_load_hparams(*ml, model, 0, 0, 0);
if (params->only_copy) {
ftype = model.ftype;
}
const size_t align = GGUF_DEFAULT_ALIGNMENT;
struct gguf_context * ctx_out = gguf_init_empty();
@ -4769,18 +4773,13 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
// quantize only 2D tensors
quantize &= (tensor->n_dims == 2);
quantize &= params->quantize_output_tensor || name != "output.weight";
quantize &= quantized_type != tensor->type;
quantize &= !params->only_copy;
enum ggml_type new_type;
void * new_data;
size_t new_size;
if (!quantize) {
new_type = tensor->type;
new_data = tensor->data;
new_size = ggml_nbytes(tensor);
LLAMA_LOG_INFO("size = %8.3f MB\n", ggml_nbytes(tensor)/1024.0/1024.0);
} else {
if (quantize) {
new_type = quantized_type;
#ifdef GGML_USE_K_QUANTS
// TODO: avoid hardcoded tensor names - use the TN_* constants
@ -4879,7 +4878,16 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
}
}
#endif
// If we've decided to quantize to the same type the tensor is already
// in then there's nothing to do.
quantize = tensor->type != new_type;
}
if (!quantize) {
new_type = tensor->type;
new_data = tensor->data;
new_size = ggml_nbytes(tensor);
LLAMA_LOG_INFO("size = %8.3f MB\n", ggml_nbytes(tensor)/1024.0/1024.0);
} else {
const size_t nelements = ggml_nelements(tensor);
float * f32_data;
@ -5310,6 +5318,7 @@ struct llama_model_quantize_params llama_model_quantize_default_params() {
/*.ftype =*/ LLAMA_FTYPE_MOSTLY_Q5_1,
/*.allow_requantize =*/ false,
/*.quantize_output_tensor =*/ true,
/*.only_copy =*/ false,
};
return result;