mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 27 Packages Projects Releases Wiki Activity

Vulkan MMQ Integer Dot Refactor and K-Quant support (#16536)

Browse Source 
* vulkan: add mmq q2_k integer dot support

* Refactor mmq caching

* Reduce mmq register use

* Load 4 quant blocks into shared memory in one step

* Pack q2_k blocks into caches of 32

* Use 32-bit accumulators for integer dot matmul

* Add q4_k mmq

* Add q3_k mmq

* Add q5_k mmq

* Add q6_k mmq

* Add mxfp4 mmq, enable MMQ MUL_MAT_ID

* Fix mmv dm loads
This commit is contained in:
Ruben Ortlam
2025-10-29 14:39:03 +01:00
committed by GitHub
parent 3eb2be1ca5
commit bcf5bda6f5
18 changed files with 941 additions and 418 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl
+3 -3
View File
@@ -120,7 +120,7 @@ layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ2
float16_t dequantFuncQ2_K(const in decodeBufQ2_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
decodeBufQ2_K_packed16 bl16 = decodeBufQ2_K_packed16(bl);
const f16vec2 d = bl.block.d;
const f16vec2 dm = bl.block.dm;
const uint idx = coordInBlock[1];
const uint scalesi = (idx & 0xF0) >> 4; // 0..15
@@ -131,7 +131,7 @@ float16_t dequantFuncQ2_K(const in decodeBufQ2_K bl, const in uint blockCoords[2
qs = unpack8(qs)[idx & 1];
const uint scales = bl.block.scales[scalesi];
float16_t ret = d.x * float16_t(scales & 0xF) * float16_t(qs) - d.y * float16_t(scales >> 4);
float16_t ret = dm.x * float16_t(scales & 0xF) * float16_t(qs) - dm.y * float16_t(scales >> 4);
return ret;
}
@@ -680,7 +680,7 @@ float16_t dequantFuncMXFP4(const in decodeBufMXFP4 bl, const in uint blockCoords
uint32_t qs = bl.block.qs[iqs];
qs >>= shift;
qs &= 0xF;
float16_t ret = float16_t(kvalues_mxfp4[qs] * d);
float16_t ret = float16_t(kvalues_mxfp4[qs] * d * 0.5);
return ret;
}
#endif