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

Further refactor attn_k

Browse source 
With attn_k set for all quants bellow 3bpw except Q2_K_S.
This commit is contained in:
Nexesenex 2024-08-10 16:33:55 +02:00
parent 8f1b99fee8
commit aa4eb594ef
1 changed files with 32 additions and 10 deletions
Download patch file Download diff file Expand all files Collapse all files

42
src/llama.cpp
View file

@ -15391,23 +15391,45 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
++qs.i_attention_wv;
} else if (name.find("attn_k.weight") != std::string::npos) {
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_IQ1_M) {
if (qs.model.hparams.n_expert >= 8) new_type = GGML_TYPE_Q5_K;
}
else if (qs.model.hparams.n_expert >= 8) {
if (qs.model.hparams.n_expert >= 8) {
// for the 8-expert model, bumping this to Q8_0 trades just ~128MB
// TODO: explore better strategies
new_type = GGML_TYPE_Q8_0;
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_IQ1_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S ||
ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q6_K;
else new_type = GGML_TYPE_Q8_0;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M && (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)) {
new_type = GGML_TYPE_IQ4_XS;
else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) && (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)) {
new_type = GGML_TYPE_IQ1_M;
}
else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) && (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)) {
new_type = GGML_TYPE_IQ2_XXS;
}
else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS) && (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)) {
new_type = GGML_TYPE_IQ2_XS;
}
else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) && (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)) {
new_type = GGML_TYPE_IQ2_S;
}
else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) &&
(qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)) {
new_type = GGML_TYPE_IQ3_XXS;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K && (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)) {
new_type = GGML_TYPE_Q3_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS && qs.model.hparams.n_gqa() < 2 && qs.model.hparams.n_expert < 2) {
new_type = GGML_TYPE_IQ3_XXS;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS && qs.model.hparams.n_gqa() < 2 && qs.model.hparams.n_expert < 2) {
new_type = GGML_TYPE_IQ2_S;
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M && (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)) {
new_type = GGML_TYPE_IQ4_XS;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L && (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)) {
new_type = GGML_TYPE_Q4_K;
}
else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S ||
ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) && (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)) {
new_type = GGML_TYPE_Q5_K;
}
} else if (name.find("attn_q.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) new_type = GGML_TYPE_IQ3_XXS;