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

Correct IQ3_M

Browse source
This commit is contained in:
Nexesenex 2024-08-18 17:44:11 +02:00
parent ddb13732c4
commit 503048a197
1 changed files with 4 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

9
src/llama.cpp
View file

@ -16045,8 +16045,8 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) {
if (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)
new_type = use_some_bits(qs.i_attention_wk, qs.n_attention_wk) ? GGML_TYPE_Q5_K : GGML_TYPE_IQ4_XS;
else new_type = use_some_bits(qs.i_attention_wk, qs.n_attention_wk) ? GGML_TYPE_IQ4_XS : GGML_TYPE_IQ3_S;
new_type = use_few_bits(qs.i_attention_wk, qs.n_attention_wk) ? GGML_TYPE_Q5_K : GGML_TYPE_IQ4_XS;
else new_type = use_few_bits(qs.i_attention_wk, qs.n_attention_wk) ? GGML_TYPE_IQ4_XS : GGML_TYPE_IQ3_S;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XL) {
if (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)
@ -16123,9 +16123,8 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
: arch != LLM_ARCH_FALCON || use_more_bits(i_layer, n_layer) ? GGML_TYPE_Q4_K
: GGML_TYPE_Q3_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M && (use_some_bits(i_layer, n_layer) ||
(qs.model.hparams.n_expert >= 4 && use_more_bits(i_layer, n_layer)))) {
new_type = GGML_TYPE_Q4_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) {
new_type = use_few_bits(i_layer, n_layer) ? GGML_TYPE_IQ4_XS : GGML_TYPE_IQ3_S;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XL) {
new_type = use_some_bits(i_layer, n_layer) ? GGML_TYPE_IQ4_XS : GGML_TYPE_IQ3_S;