ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-25 07:34:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

iq2_kl: CUDA dequantize

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-07-10 14:22:37 +03:00
parent dfae22680b
commit d0f85cb9c2
9 changed files with 165 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
src/llama.cpp
View File

@@ -4435,6 +4435,7 @@ struct llama_model_loader {
case GGML_TYPE_IQ2_K: ftype = LLAMA_FTYPE_MOSTLY_IQ2_K; break;
case GGML_TYPE_IQ2_K_R4:ftype = LLAMA_FTYPE_MOSTLY_IQ2_K_R4;break;
case GGML_TYPE_IQ3_KS: ftype = LLAMA_FTYPE_MOSTLY_IQ3_KS; break;
case GGML_TYPE_IQ2_KL: ftype = LLAMA_FTYPE_MOSTLY_IQ2_KL; break;
case GGML_TYPE_IQ3_K: ftype = LLAMA_FTYPE_MOSTLY_IQ3_K; break;
case GGML_TYPE_IQ3_K_R4:ftype = LLAMA_FTYPE_MOSTLY_IQ3_K_R4;break;
case GGML_TYPE_IQ4_K: ftype = LLAMA_FTYPE_MOSTLY_IQ4_K; break;
@@ -5177,6 +5178,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_IQ2_K: return "IQ2_K - 2.375 bpw";
case LLAMA_FTYPE_MOSTLY_IQ2_K_R4: return "IQ2_K_R4 - 2.375 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_KS: return "IQ3_KS - 3.1875 bpw";
case LLAMA_FTYPE_MOSTLY_IQ2_KL: return "IQ2_KL - 2.6875 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_K: return "IQ3_K - 3.4325 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_K_R4: return "IQ3_K_R4 - 3.4325 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_KL: return "IQ3_KL - 4 bpw";
@@ -19129,7 +19131,7 @@ static ggml_type change_type_if_necessary(ggml_type new_type, int nx, int ny) {
new_type == GGML_TYPE_IQ3_XXS_R4 || new_type == GGML_TYPE_IQ2_XXS_R4 || new_type == GGML_TYPE_IQ2_XS_R4 ||
new_type == GGML_TYPE_IQ2_S_R4|| new_type == GGML_TYPE_IQ3_S_R4|| new_type == GGML_TYPE_IQ3_KS ||
new_type == GGML_TYPE_IQ2_KT || new_type == GGML_TYPE_IQ3_KT || new_type == GGML_TYPE_IQ4_KT ||
new_type == GGML_TYPE_IQ5_KS || new_type == GGML_TYPE_IQ5_KS_R4) {
new_type == GGML_TYPE_IQ5_KS || new_type == GGML_TYPE_IQ5_KS_R4|| new_type == GGML_TYPE_IQ2_KL) {
if (nx % QK_K != 0) {
LLAMA_LOG_WARN("\n\n%s : tensor cols %d x %d are not divisible by %d, required for %s", __func__, nx, ny, QK_K, ggml_type_name(new_type));
convert_incompatible_tensor = true;
@@ -19161,6 +19163,7 @@ static ggml_type change_type_if_necessary(ggml_type new_type, int nx, int ny) {
case GGML_TYPE_Q3_K_R4:
case GGML_TYPE_IQ2_K:
case GGML_TYPE_IQ2_K_R4:
case GGML_TYPE_IQ2_KL:
case GGML_TYPE_IQ3_KS:
case GGML_TYPE_IQ3_K:
case GGML_TYPE_IQ3_K_R4:
@@ -19297,7 +19300,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M ||
ftype == LLAMA_FTYPE_MOSTLY_IQ1_M || ftype == LLAMA_FTYPE_MOSTLY_IQ2_K || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K ||
ftype == LLAMA_FTYPE_MOSTLY_IQ2_KS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_KS ||
ftype == LLAMA_FTYPE_MOSTLY_IQ2_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4 ||
ftype == LLAMA_FTYPE_MOSTLY_IQ2_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ2_KL ||
ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M_R4 ||
ftype == LLAMA_FTYPE_MOSTLY_IQ1_S_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M_R4 ||
ftype == LLAMA_FTYPE_MOSTLY_IQ2_KT || ftype == LLAMA_FTYPE_MOSTLY_IQ3_KT) {
@@ -19504,6 +19507,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_KS && qs.model.hparams.n_gqa() >= 2) {
new_type = GGML_TYPE_IQ4_KS;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_KL && qs.model.hparams.n_gqa() >= 2) {
new_type = GGML_TYPE_IQ4_KS;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_K_R4 && qs.model.hparams.n_gqa() >= 2) {
new_type = GGML_TYPE_IQ4_K_R4;
}
@@ -19549,6 +19555,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_IQ3_S) new_type = GGML_TYPE_Q4_K;
else if (new_type == GGML_TYPE_IQ3_K) new_type = GGML_TYPE_IQ4_K;
else if (new_type == GGML_TYPE_IQ3_KS) new_type = GGML_TYPE_IQ4_KS;
else if (new_type == GGML_TYPE_IQ2_KL) new_type = GGML_TYPE_IQ4_KS;
else if (new_type == GGML_TYPE_IQ3_S_R4) new_type = GGML_TYPE_Q4_K_R4;
else if (new_type == GGML_TYPE_Q3_K_R4) new_type = GGML_TYPE_Q4_K_R4;
else if (new_type == GGML_TYPE_Q4_K || new_type == GGML_TYPE_IQ4_XS) new_type = GGML_TYPE_Q5_K;
@@ -19681,6 +19688,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
new_type = GGML_TYPE_Q5_K; // should the IQ_K quants be applied here as the new type for the IQ_K ftypes ?
// also, this condition could be reproduced on attn_q, eventually with Q4_K instead of Q5_K.
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_KL) {
new_type = GGML_TYPE_IQ4_KS;
}
} else {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K ) new_type = GGML_TYPE_Q3_K; // This list could be generalized and streamlined
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) new_type = GGML_TYPE_IQ3_S;
@@ -19918,6 +19928,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
case LLAMA_FTYPE_MOSTLY_IQ2_K: default_type = GGML_TYPE_IQ2_K; break;
case LLAMA_FTYPE_MOSTLY_IQ2_K_R4:default_type = GGML_TYPE_IQ2_K_R4;break;
case LLAMA_FTYPE_MOSTLY_IQ3_KS: default_type = GGML_TYPE_IQ3_KS; break;
case LLAMA_FTYPE_MOSTLY_IQ2_KL: default_type = GGML_TYPE_IQ2_KL; break;
case LLAMA_FTYPE_MOSTLY_IQ3_K: default_type = GGML_TYPE_IQ3_K; break;
case LLAMA_FTYPE_MOSTLY_IQ3_K_R4:default_type = GGML_TYPE_IQ3_K_R4;break;
case LLAMA_FTYPE_MOSTLY_IQ3_KL: default_type = GGML_TYPE_IQ3_K; break;