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

Usage of n_head to discriminate very small models

Browse source 
Of which the size is more sensitive to the non repeating tensors
This commit is contained in:
Nexesenex 2024-08-23 23:27:26 +02:00
parent 16e9c3771a
commit 1bde168c07
1 changed files with 10 additions and 9 deletions
Download patch file Download diff file Expand all files Collapse all files

19
src/llama.cpp
View file

@ -16441,14 +16441,15 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) new_type = GGML_TYPE_Q4_K;
else new_type = GGML_TYPE_IQ4_XS;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_XL || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) {
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_XL || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS ||
ftype == LLAMA_FTYPE_MOSTLY_IQ2_S) {
if (qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_Q6_K;
else if (qs.model.hparams.n_head <= 20) new_type = GGML_TYPE_IQ4_XS;
else if (qs.model.hparams.n_head() <= 20) new_type = GGML_TYPE_IQ4_XS;
else new_type = GGML_TYPE_Q4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XL) {
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_M || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XL) {
if (qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_Q6_K;
else if (qs.model.hparams.n_vocab >= 127999 || qs.model.hparams.n_head <= 20) new_type = GGML_TYPE_Q4_K;
else if (qs.model.hparams.n_vocab >= 127999) new_type = GGML_TYPE_Q4_K;
else new_type = GGML_TYPE_Q5_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) {
@ -16457,7 +16458,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_S || ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) {
if (qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_Q6_K;
else if (qs.model.hparams.n_vocab >= 127999 || qs.model.hparams.n_head <= 20) new_type = GGML_TYPE_Q5_K;
else if (qs.model.hparams.n_vocab >= 127999 || qs.model.hparams.n_head() <= 20) new_type = GGML_TYPE_Q5_K;
else new_type = GGML_TYPE_Q6_K;
}
else if (new_type != GGML_TYPE_Q8_0) {
@ -16488,18 +16489,18 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
new_type = GGML_TYPE_IQ2_S;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) {
if (qs.model.hparams.n_vocab >= 127999 || qs.model.hparams.n_head <= 20) new_type = GGML_TYPE_IQ2_S;
if (qs.model.hparams.n_vocab >= 127999 || qs.model.hparams.n_head() <= 20) new_type = GGML_TYPE_IQ2_S;
else new_type = GGML_TYPE_IQ3_XXS;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XL || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
new_type = GGML_TYPE_IQ3_XXS;
}
else if ( || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) {
if (qs.model.hparams.n_vocab >= 127999 || qs.model.hparams.n_head <= 20) new_type = GGML_TYPE_IQ3_XXS;
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) {
if (qs.model.hparams.n_vocab >= 127999 || qs.model.hparams.n_head() <= 20) new_type = GGML_TYPE_IQ3_XXS;
else new_type = GGML_TYPE_IQ3_S;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XL || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXL) {
if (qs.model.hparams.n_vocab >= 127999 || qs.model.hparams.n_head <= 20) new_type = GGML_TYPE_IQ3_S;
if (qs.model.hparams.n_vocab >= 127999 || qs.model.hparams.n_head() <= 20) new_type = GGML_TYPE_IQ3_S;
new_type = GGML_TYPE_IQ4_XS;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) new_type = GGML_TYPE_IQ3_XXS;