ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-19 04:40:09 +00:00
Code Issues Packages Projects Releases Wiki Activity

Adding IQ5_KS - 5.25 bpw quants (#422)

Browse Source 
* iq5_ks: basics

* iq5_ks: quantize

* iq5_ks: CUDA dequantize works

* iq5_ks: dot product works on CUDA

* iq5_ks: MMQ works

* iq5_ks: Zen4

* iq5_ks: AVX2

But is is not quite right, just like iq4_k, iq5_k, iq6_k, iq4_ks.
All these need fixing on AVX2.

* iq5_ks: NEON

* iq5_ks: Metal dequantize

* iq5_ks: Metal dot product

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2025-05-15 16:02:39 +03:00
committed by GitHub
parent 3f8c865b92
commit 3d92d7f802
20 changed files with 848 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
src/llama.cpp
View File

@@ -4373,6 +4373,7 @@ struct llama_model_loader {
case GGML_TYPE_IQ4_KS: ftype = LLAMA_FTYPE_MOSTLY_IQ4_KS; break;
case GGML_TYPE_IQ4_KS_R4:ftype = LLAMA_FTYPE_MOSTLY_IQ4_KS_R4; break;
case GGML_TYPE_IQ4_KSS: ftype = LLAMA_FTYPE_MOSTLY_IQ4_KSS; break;
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_K: ftype = LLAMA_FTYPE_MOSTLY_IQ3_K; break;
@@ -5109,6 +5110,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_IQ4_KS: return "IQ4_KS - 4.25 bpw";
case LLAMA_FTYPE_MOSTLY_IQ4_KS_R4:return "IQ4_KS_R4 - 4.25 bpw";
case LLAMA_FTYPE_MOSTLY_IQ4_KSS: return "IQ4_KSS - 4.0 bpw";
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_K: return "IQ3_K - 3.4325 bpw";
@@ -18619,7 +18621,7 @@ 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_IQ5_KS) {
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;
@@ -18661,6 +18663,7 @@ static ggml_type change_type_if_necessary(ggml_type new_type, int nx, int ny) {
case GGML_TYPE_IQ4_K:
case GGML_TYPE_IQ4_K_R4:
case GGML_TYPE_Q4_K_R4:
case GGML_TYPE_IQ5_KS:
case GGML_TYPE_Q4_K: new_type = GGML_TYPE_Q5_0; break;
case GGML_TYPE_IQ5_K:
case GGML_TYPE_IQ5_K_R4:
@@ -19321,6 +19324,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
case LLAMA_FTYPE_MOSTLY_IQ4_KS: default_type = GGML_TYPE_IQ4_KS; break;
case LLAMA_FTYPE_MOSTLY_IQ4_KS_R4:default_type = GGML_TYPE_IQ4_KS_R4;break;
case LLAMA_FTYPE_MOSTLY_IQ4_KSS: default_type = GGML_TYPE_IQ4_KSS; break;
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_K: default_type = GGML_TYPE_IQ3_K; break;