mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 24 Packages Projects Releases Wiki Activity

vulkan: add FA dequant for q4_1, q5_0, q5_1, iq4_nl (#21029)

Browse Source 
Add dequantize4() implementations for Q4_1, Q5_0, Q5_1, and IQ4_NL
in the flash attention base shader. Register them in the shader
generator, pipeline creation, and enable in the scalar/coopmat1 FA
support check.
This commit is contained in:
mkoker
2026-04-07 07:41:29 -04:00
committed by GitHub
parent 482192f12d
commit edd4d9bca5
3 changed files with 119 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/ggml-vulkan.cpp
+16 -8
View File
@@ -3447,11 +3447,19 @@ static void ggml_vk_load_shaders(vk_device& device) {
CREATE_FA(GGML_TYPE_F16, f16, FA_SCALAR, )
CREATE_FA(GGML_TYPE_Q4_0, q4_0, FA_SCALAR, )
CREATE_FA(GGML_TYPE_Q8_0, q8_0, FA_SCALAR, )
CREATE_FA(GGML_TYPE_Q4_1, q4_1, FA_SCALAR, )
CREATE_FA(GGML_TYPE_Q5_0, q5_0, FA_SCALAR, )
CREATE_FA(GGML_TYPE_Q5_1, q5_1, FA_SCALAR, )
CREATE_FA(GGML_TYPE_IQ4_NL, iq4_nl, FA_SCALAR, )
} else {
CREATE_FA(GGML_TYPE_F32, f32, FA_SCALAR, _fp32)
CREATE_FA(GGML_TYPE_F16, f16, FA_SCALAR, _fp32)
CREATE_FA(GGML_TYPE_Q4_0, q4_0, FA_SCALAR, _fp32)
CREATE_FA(GGML_TYPE_Q8_0, q8_0, FA_SCALAR, _fp32)
CREATE_FA(GGML_TYPE_Q4_1, q4_1, FA_SCALAR, _fp32)
CREATE_FA(GGML_TYPE_Q5_0, q5_0, FA_SCALAR, _fp32)
CREATE_FA(GGML_TYPE_Q5_1, q5_1, FA_SCALAR, _fp32)
CREATE_FA(GGML_TYPE_IQ4_NL, iq4_nl, FA_SCALAR, _fp32)
}
#if defined(VK_KHR_cooperative_matrix) && defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
if (device->coopmat1_fa_support) {
@@ -3459,6 +3467,10 @@ static void ggml_vk_load_shaders(vk_device& device) {
CREATE_FA(GGML_TYPE_F16, f16, FA_COOPMAT1, _cm1)
CREATE_FA(GGML_TYPE_Q4_0, q4_0, FA_COOPMAT1, _cm1)
CREATE_FA(GGML_TYPE_Q8_0, q8_0, FA_COOPMAT1, _cm1)
CREATE_FA(GGML_TYPE_Q4_1, q4_1, FA_COOPMAT1, _cm1)
CREATE_FA(GGML_TYPE_Q5_0, q5_0, FA_COOPMAT1, _cm1)
CREATE_FA(GGML_TYPE_Q5_1, q5_1, FA_COOPMAT1, _cm1)
CREATE_FA(GGML_TYPE_IQ4_NL, iq4_nl, FA_COOPMAT1, _cm1)
}
#endif
#if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
@@ -15331,11 +15343,12 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
case GGML_TYPE_F32:
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q8_0:
// supported in scalar and coopmat2 paths
break;
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_IQ4_NL:
// supported in scalar and coopmat2 paths
break;
// K dequants currently disabled because D dimension is rounded up to 256 and runs inefficiently
//case GGML_TYPE_Q2_K:
//case GGML_TYPE_Q3_K:
@@ -15350,12 +15363,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
//case GGML_TYPE_IQ3_XXS:
//case GGML_TYPE_IQ3_S:
//case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ4_NL:
// currently supported only in coopmat2 path
if (!coopmat2) {
return false;
}
break;
default:
return false;
}