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

Start rebase

This commit is contained in:
Colin Kealty 2026-03-09 16:11:36 -04:00
parent 86103e7e06
commit 86273028d1
1 changed files with 16 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
src/llama-quant.h
View File

@ -29,13 +29,23 @@ struct quantize_state_impl {
bool has_imatrix = false;
// used to figure out if a model shares tok_embd with the output weight
bool has_output = false;
// used to figure out if a model has tied embeddings (tok_embd shares weights with output)
bool has_tied_embeddings = false; // assume tied until we see output.weight
quantize_state_impl(const llama_model & model, const llama_model_quantize_params * params)
: model(model)
, params(params)
{}
// tensor type override patterns (compiled once, used twice)
std::vector<std::pair<std::regex, ggml_type>> tensor_type_patterns;
quantize_state_impl(const llama_model & model, const llama_model_quantize_params * params):
model(model), params(params)
{
// compile regex patterns once - they are expensive
if (params->tensor_types) {
const auto & tensor_types = *static_cast<const std::vector<tensor_type_option> *>(params->tensor_types);
for (const auto & [tname, qtype] : tensor_types) {
tensor_type_patterns.emplace_back(std::regex(tname), qtype);
}
}
}
};
ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_type, const ggml_tensor * tensor, llama_ftype ftype);