ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-11 22:40:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

Adding IQ3_KS quants (#566)

Browse Source 
* iq3_ks: basics

* iq3_ks: CUDA dequantize

* iq3_ks: CUDA mmvq

* iq3_ks: mmq

* iq3_ks: faster mmq

* iq3_ks: Zen4

* iq3_ks: AVX2 convert to q8_k_r8

This gives usPP-512 = 360 t/s.

* iq3_ks: AVX2 GEMM/GEMV

* iq3_ks: NEON GEMM/GEMV

* iq3_ks: NEON convert to q8_k_r8

This gives us PP-512 = 164 t/s.

* iq3_ks: Metal dequantize

* iq3_ks: Metal gemv - pathetic performance

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2025-07-02 09:27:47 +02:00
committed by GitHub
parent 46f2e5d249
commit 3248a35992
22 changed files with 1040 additions and 65 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
src/llama.cpp
View File

@@ -4374,6 +4374,7 @@ struct llama_model_loader {
case GGML_TYPE_IQ5_KS: ftype = LLAMA_FTYPE_MOSTLY_IQ5_KS; break;
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_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;
@@ -5115,6 +5116,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_IQ5_KS: return "IQ5_KS - 5.25 bpw";
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_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";
@@ -18642,7 +18644,7 @@ static ggml_type change_type_if_necessary(ggml_type new_type, int nx, int ny) {
new_type == GGML_TYPE_IQ4_K_R4|| new_type == GGML_TYPE_Q8_K_R8 || new_type == GGML_TYPE_IQ3_K_R4||
new_type == GGML_TYPE_IQ2_K_R4|| new_type == GGML_TYPE_IQ5_K_R4|| new_type == GGML_TYPE_IQ4_KS_R4 ||
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_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) {
if (nx % QK_K != 0) {
@@ -18676,6 +18678,7 @@ 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_IQ3_KS:
case GGML_TYPE_IQ3_K:
case GGML_TYPE_IQ3_K_R4:
case GGML_TYPE_IQ4_KSS:
@@ -18810,7 +18813,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
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_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_IQ3_XXS_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M_R4 ||
ftype == LLAMA_FTYPE_MOSTLY_IQ1_S_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M_R4 ||
@@ -19003,6 +19006,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
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_KS && 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;
}
@@ -19059,6 +19065,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (new_type == GGML_TYPE_Q2_K_R4 || new_type == GGML_TYPE_IQ3_XXS_R4) new_type = GGML_TYPE_IQ3_K_R4;
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_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;
@@ -19185,7 +19192,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
ftype == LLAMA_FTYPE_MOSTLY_IQ5_KS || ftype == LLAMA_FTYPE_MOSTLY_IQ5_KS_R4 ||
ftype == LLAMA_FTYPE_MOSTLY_IQ2_K || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K || ftype == LLAMA_FTYPE_MOSTLY_IQ3_KL ||
ftype == LLAMA_FTYPE_MOSTLY_Q4_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS_R8 ||
ftype == LLAMA_FTYPE_MOSTLY_Q3_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_KT ||
ftype == LLAMA_FTYPE_MOSTLY_Q3_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_KT || ftype == LLAMA_FTYPE_MOSTLY_IQ3_KS ||
ftype == LLAMA_FTYPE_MOSTLY_Q2_K_R4|| ftype == LLAMA_FTYPE_MOSTLY_IQ4_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K_R4 ||
ftype == LLAMA_FTYPE_MOSTLY_IQ2_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S_R4) {
new_type = GGML_TYPE_Q5_K; // should the IQ_K quants be applied here as the new type for the IQ_K ftypes ?
@@ -19427,6 +19434,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
case LLAMA_FTYPE_MOSTLY_IQ5_KS: default_type = GGML_TYPE_IQ5_KS; break;
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_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;