Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

vulkan: Support Q2_0 (#25430)

Browse Source 
* vulkan: Support Q2_0

The backend perf tests for mat-vec-mul weren't very good at first (worse than
q2_k), doubling the rows per workgroup made a big difference.

* reorder

* resolve merge conflict, adjust err threshold for f16->q2_0 set_rows
This commit is contained in:
Jeff Bolz
2026-07-17 08:42:59 +02:00
committed by GitHub
parent 0bd0ec6099
commit 788e07dc91
13 changed files with 266 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl
+23
View File
@@ -46,6 +46,26 @@ f16vec4 dequantFuncQ1_0_v(const in decodeBufQ1_0 bl, const in uint blockCoords[2
(qs_nib & 8u) != 0u ? d : md);
}
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ2_0 {
block_q2_0 block;
};
float16_t dequantFuncQ2_0(const in decodeBufQ2_0 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
const float16_t d = bl.block.d;
const uint idx = coordInBlock[1];
const uint bits = uint(bl.block.qs[idx >> 2]) >> (2u * (idx & 3u));
return (float16_t(bits & 3u) - float16_t(1.0)) * d;
}
f16vec4 dequantFuncQ2_0_v(const in decodeBufQ2_0 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
const float16_t d = bl.block.d;
const uint idx = coordInBlock[1];
const uint bits = uint(bl.block.qs[idx >> 2]);
return f16vec4((vec4(bits & 3u, (bits >> 2u) & 3u, (bits >> 4u) & 3u, bits >> 6u) - 1.0f) * float(d));
}
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ4_0 {
block_q4_0_packed16 block;
};
@@ -1330,6 +1350,9 @@ f16vec4 dequantFuncNVFP4_v(const in decodeBufNVFP4 bl, const in uint blockCoords
#if defined(DATA_A_Q1_0)
#define dequantFuncA dequantFuncQ1_0
#define dequantFuncA_v dequantFuncQ1_0_v
#elif defined(DATA_A_Q2_0)
#define dequantFuncA dequantFuncQ2_0
#define dequantFuncA_v dequantFuncQ2_0_v
#elif defined(DATA_A_Q4_0)
#define dequantFuncA dequantFuncQ4_0
#define dequantFuncA_v dequantFuncQ4_0_v