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

Avoid to shrink attn.k.weight for IQ3_XS and XXS when GQA or MOE

Browse source
This commit is contained in:
Nexes the Old 2024-08-08 18:50:48 +02:00 • committed by GitHub
parent 59c5d479de
commit 8006b15fd1
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 2 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

4
src/llama.cpp
View file

@ -15414,10 +15414,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
// TODO: explore better strategies
new_type = GGML_TYPE_Q8_0;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) {
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS && (qs.model.hparams.n_gqa() < 2 || qs.model.hparams.n_expert < 2)) {
new_type = GGML_TYPE_IQ3_XXS;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS && (qs.model.hparams.n_gqa() < 2 || qs.model.hparams.n_expert < 2)) {
new_type = GGML_TYPE_IQ2_S;
}
} else if (name.find("attn_q.weight") != std::string::npos) {