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

harmonize formatting of tensor type conditions

Browse source
This commit is contained in:
Nexesenex 2024-08-20 11:59:41 +02:00
parent ce86019770
commit dbadcdd5cf
1 changed files with 3 additions and 6 deletions
Download patch file Download diff file Expand all files Collapse all files

9
src/llama.cpp
View file

@ -16275,8 +16275,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q4_K;
}
++qs.i_attention_wo;
}
else if (name.find("attn_qkv.weight") != std::string::npos) {
} else if (name.find("attn_qkv.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q2_K || ftype == LLAMA_FTYPE_MOSTLY_Q2_K_L) {
new_type = GGML_TYPE_Q3_K;
}
@ -16300,8 +16299,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS) new_type = GGML_TYPE_Q4_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ4_XSR || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXXL) new_type = GGML_TYPE_Q5_K;
++qs.i_attention_wv;
}
else if (name.find("ffn_gate") != std::string::npos) {
} else if (name.find("ffn_gate") != std::string::npos) {
auto info = layer_info(qs.i_ffn_gate, qs.n_ffn_gate, name.c_str());
int i_layer = info.first, n_layer = info.second;
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_L && (difquant_half_tensors(i_layer, n_layer))) new_type = GGML_TYPE_Q3_K;
@ -16319,8 +16317,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXL && (difquant_half_tensors(i_layer, n_layer))) new_type = GGML_TYPE_IQ4_XS;
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXXL && (difquant_six_eights_tensors(i_layer, n_layer))) new_type = GGML_TYPE_IQ4_XS;
++qs.i_ffn_gate;
}
else if (name.find("ffn_up") != std::string::npos) {
} else if (name.find("ffn_up") != std::string::npos) {
auto info = layer_info(qs.i_ffn_up, qs.n_ffn_up, name.c_str());
int i_layer = info.first, n_layer = info.second;
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_L && (difquant_half_tensors(i_layer, n_layer))) new_type = GGML_TYPE_Q3_K;