ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-22 23:49:23 +00:00
Code Issues Packages Projects Releases Wiki Activity

WIP + adding iq3_kl quantization mix

Browse Source
This commit is contained in:
Iwan Kawrakow
2024-10-08 13:56:29 +03:00
parent 4c76471979
commit 1dd6c40c15
11 changed files with 110 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
src/llama.cpp
View File

@@ -4498,6 +4498,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_IQ4_XXS: return "IQ4_XXS - 4.06 bpw";
case LLAMA_FTYPE_MOSTLY_IQ2_K: return "IQ2_K - 2.375 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_K: return "IQ3_K - 3.4325 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_KL: return "IQ3_KL - 4 bpw";
case LLAMA_FTYPE_MOSTLY_IQ4_K: return "IQ4_K - 4.5 bpw";
case LLAMA_FTYPE_MOSTLY_IQ5_K: return "IQ5_K - 5.5 bpw";
case LLAMA_FTYPE_MOSTLY_IQ6_K: return "IQ6_K - 6.6 bpw";
@@ -15699,12 +15700,15 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_Q4_K : qs.model.hparams.n_gqa() >= 2 ? GGML_TYPE_IQ3_K
: !qs.has_imatrix ? GGML_TYPE_IQ3_S : GGML_TYPE_IQ3_XXS;
}
else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K) && qs.model.hparams.n_gqa() >= 2) {
else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S) && qs.model.hparams.n_gqa() >= 2) {
new_type = GGML_TYPE_IQ4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_K && qs.model.hparams.n_gqa() >= 2) {
new_type = GGML_TYPE_IQ4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_KL) {
new_type = qs.model.hparams.n_gqa() >= 2 ? GGML_TYPE_IQ5_K : GGML_TYPE_IQ4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) {
new_type = qs.model.hparams.n_gqa() >= 2 ? GGML_TYPE_IQ5_K : GGML_TYPE_IQ4_K;
}
@@ -15781,6 +15785,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) {
new_type = arch == LLM_ARCH_FALCON ? GGML_TYPE_Q4_K : GGML_TYPE_Q5_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_KL) {
new_type = use_more_bits(i_layer, n_layer) ? GGML_TYPE_IQ4_K : GGML_TYPE_IQ3_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) {
if (arch == LLM_ARCH_FALCON) {
new_type = i_layer < n_layer/16 ? GGML_TYPE_Q6_K :
@@ -15821,6 +15828,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L ) new_type = GGML_TYPE_Q5_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M ) new_type = GGML_TYPE_IQ4_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_K ) new_type = GGML_TYPE_IQ3_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_KL ) new_type = GGML_TYPE_IQ4_K;
}
} else {
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q4_K;
@@ -15840,6 +15848,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS && (i_layer >= n_layer/8 && i_layer < 7*n_layer/8)) {
new_type = GGML_TYPE_IQ3_XXS;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_KL && use_more_bits(i_layer, n_layer)) {
new_type = GGML_TYPE_IQ4_XS;
}
++qs.i_ffn_gate;
}
else if (name.find("ffn_up") != std::string::npos) {
@@ -15848,6 +15859,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS && (i_layer >= n_layer/8 && i_layer < 7*n_layer/8)) {
new_type = GGML_TYPE_IQ3_XXS;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_KL && use_more_bits(i_layer, n_layer)) {
new_type = GGML_TYPE_IQ4_K;
}
++qs.i_ffn_up;
}
@@ -16014,6 +16028,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
case LLAMA_FTYPE_MOSTLY_IQ4_XXS: default_type = GGML_TYPE_IQ4_XXS; break;
case LLAMA_FTYPE_MOSTLY_IQ2_K: default_type = GGML_TYPE_IQ2_K; break;
case LLAMA_FTYPE_MOSTLY_IQ3_K: default_type = GGML_TYPE_IQ3_K; break;
case LLAMA_FTYPE_MOSTLY_IQ3_KL: default_type = GGML_TYPE_IQ3_K; break;
case LLAMA_FTYPE_MOSTLY_IQ4_K: default_type = GGML_TYPE_IQ4_K; break;
case LLAMA_FTYPE_MOSTLY_IQ5_K: default_type = GGML_TYPE_IQ5_K; break;
case LLAMA_FTYPE_MOSTLY_IQ6_K: default_type = GGML_TYPE_IQ6_K; break;