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

Update llama.cpp - adjustements non-FFN layer tensors

Browse source
This commit is contained in:
Nexesenex 2024-03-25 23:08:19 +01:00 • committed by GitHub
parent b3553335a3
commit 066efbb18f
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 5 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

7
llama.cpp
View file

@ -12479,17 +12479,20 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
}
} else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_XS) {
if (name.find("attn_q.weight") != std::string::npos) {
if (qs.model.hparams.n_expert >= 2) new_type = GGML_TYPE_IQ2_S;
if (qs.model.hparams.n_expert >= 8) new_type = GGML_TYPE_IQ2_S;
else if (qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_IQ2_XS;
else if (qs.model.hparams.n_expert >= 2) new_type = GGML_TYPE_IQ2_XXS;
}
if (name.find("attn_k.weight") != std::string::npos) {
if (qs.model.hparams.n_expert >= 8) new_type = GGML_TYPE_IQ4_XS;
else if (qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_IQ3_S;
else if (qs.model.hparams.n_expert >= 2) new_type = GGML_TYPE_IQ3_XXS;
else if (qs.model.hparams.n_gqa() >= 8) new_type = GGML_TYPE_IQ2_S;
else if (qs.model.hparams.n_gqa() >= 4) new_type = GGML_TYPE_IQ2_XS;
else if (qs.model.hparams.n_gqa() >= 2) new_type = GGML_TYPE_IQ2_XXS;
}
else if (name.find("attn_v.weight") != std::string::npos) {
if (qs.model.hparams.n_expert >= 8) new_type = GGML_TYPE_IQ4_XS;
if (qs.model.hparams.n_expert >= 8) new_type = GGML_TYPE_Q5_K;
else if (qs.model.hparams.n_gqa() >= 8 || qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_IQ4_XS;
else if (qs.model.hparams.n_gqa() >= 4 || qs.model.hparams.n_expert >= 2) new_type = GGML_TYPE_IQ3_S;
else if (qs.model.hparams.n_gqa() >= 2) new_type = GGML_TYPE_IQ3_XXS;