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

More occurences of n_experts == 8 changed to >= in quant strategies

Browse source
This commit is contained in:
Nexes the Old 2024-08-03 03:04:25 +02:00 • committed by GitHub
parent 7d337d0f89
commit d5779c27ba
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 3 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

6
src/llama.cpp
View file

@ -15362,7 +15362,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
++qs.i_ffn_down;
}
else if (name.find("attn_output.weight") != std::string::npos) {
if (qs.model.hparams.n_expert == 8) {
if (qs.model.hparams.n_expert >= 8) {
new_type = GGML_TYPE_Q5_K;
} else {
if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) new_type = GGML_TYPE_IQ2_XXS;
@ -15443,7 +15443,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
: GGML_TYPE_Q3_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M && (i_layer < n_layer/8 ||
(qs.model.hparams.n_expert == 8 && use_more_bits(i_layer, n_layer)))) {
(qs.model.hparams.n_expert >= 8 && use_more_bits(i_layer, n_layer)))) {
new_type = GGML_TYPE_Q4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) {
@ -15474,7 +15474,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
++qs.i_ffn_down;
} else if (name.find("attn_output.weight") != std::string::npos) {
if (arch != LLM_ARCH_FALCON) {
if (qs.model.hparams.n_expert == 8) {
if (qs.model.hparams.n_expert >= 8) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL ||
ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S ||