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

Update llama.cpp

Browse source 
- IQ4_XS output for models lesser than 8 experts or GQA 8
- granularity for QKV tensor when existing
- also, drop for Mistral & Yi attn.k.weight from IQ2_XS to IQ2_XXS
This commit is contained in:
Nexesenex 2024-03-29 13:12:27 +01:00 • committed by GitHub
parent 915be46ffc
commit ed4be6bb0d
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 5 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

8
llama.cpp
View file

@ -12773,7 +12773,8 @@ 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_Q5_K; if (qs.model.hparams.n_expert >= 2) new_type = GGML_TYPE_Q5_K;
else new_type = GGML_TYPE_Q4_K; else if (qs.model.hparams.n_gqa() >= 8) new_type = GGML_TYPE_Q4_K;
else new_type = GGML_TYPE_IQ4_XS;
} }
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS || else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M ||
@ -12814,7 +12815,6 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_IQ3_S; 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_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() >= 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 (qs.model.hparams.n_gqa() >= 2) new_type = GGML_TYPE_IQ2_XXS;
} }
else if (name.find("attn_v.weight") != std::string::npos) { else if (name.find("attn_v.weight") != std::string::npos) {
@ -12832,7 +12832,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else new_type = GGML_TYPE_IQ2_XXS; else new_type = GGML_TYPE_IQ2_XXS;
} }
else if (name.find("attn_qkv.weight") != std::string::npos) { else if (name.find("attn_qkv.weight") != std::string::npos) {
if (qs.model.hparams.n_expert >= 2) new_type = GGML_TYPE_Q4_K; if (qs.model.hparams.n_expert >= 8) new_type = GGML_TYPE_Q4_K;
else if (qs.model.hparams.n_gqa() > 4 || qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_IQ3_S;
else if (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) new_type = GGML_TYPE_IQ3_XXS;
else new_type = GGML_TYPE_Q2_K; 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 ||