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

Update llama.cpp - Non-FFN layer-tensors strategy

Browse source
This commit is contained in:
Nexesenex 2024-03-25 21:04:01 +01:00 • committed by GitHub
parent 1c4da5ddac
commit ddc7701588
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 30 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

31
llama.cpp
View file

@ -12468,7 +12468,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
} }
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_XS) { else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_XS) {
if (qs.model.hparams.n_expert >= 2) new_type = GGML_TYPE_Q2_K; if (qs.model.hparams.n_expert >= 2) new_type = GGML_TYPE_Q2_K;
new_type = GGML_TYPE_IQ2_S; else new_type = GGML_TYPE_IQ2_S;
} }
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) { else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) {
new_type = GGML_TYPE_IQ3_S; new_type = GGML_TYPE_IQ3_S;
@ -12477,6 +12477,35 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
new_type = GGML_TYPE_IQ3_S; new_type = GGML_TYPE_IQ3_S;
} }
} }
} 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 (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() >= 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;
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;
else new_type = GGML_TYPE_Q2_K;
++qs.i_attention_wv;
}
else if (name.find("attn_output.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 new_type = GGML_TYPE_IQ2_XXS;
}
else if (name.find("attn_qkv.weight") != std::string::npos) {
if (qs.model.hparams.n_expert >= 2) new_type = GGML_TYPE_IQ4_XS;
else new_type = GGML_TYPE_Q2_K;
}
} else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S ||
ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) { ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) {
if (name.find("attn_v.weight") != std::string::npos) { if (name.find("attn_v.weight") != std::string::npos) {