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

ggml-webgpu: Add MMVQ path for Q4/Q8/Q2_K/Q4_K and clean up legacy MUL_MAT pipeline (#23594)

Browse Source 
* ggml-webgpu: Add MMVQ path for Q4/Q8/Q2_K/Q4_K

* Fix to editorconfig checking pass

* Remove mul-mat-legacy pipeline

* Fix to use vendor name as is and add dot_product/vendor to shader_lib_ctx
This commit is contained in:
Masashi Yoshimura
2026-05-25 20:42:49 -07:00
committed by GitHub
parent 54121f7325
commit 1506d39e76
8 changed files with 715 additions and 932 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl
+21 -3
View File
@@ -3,10 +3,18 @@ enable subgroups;
#endif
enable f16;
#ifdef MMVQ
requires packed_4x8_integer_dot_product;
#endif
#define DECLARE_BYTE_LOADERS_SRC0
#include "common_decls.tmpl"
#ifdef MMVQ
#include "mul_mat_vec_q_acc.tmpl"
#else
#include "mul_mat_vec_acc.tmpl"
#endif
struct MulMatParams {
offset_src0: u32,
@@ -28,9 +36,14 @@ struct MulMatParams {
};
@group(0) @binding(0) var<storage, read_write> src0: array<SRC0_TYPE>;
@group(0) @binding(1) var<storage, read_write> src1: array<SRC1_TYPE>;
@group(0) @binding(2) var<storage, read_write> dst: array<f32>;
#ifdef MMVQ
@group(0) @binding(1) var<storage, read_write> src1q: array<q8_1>;
#else
@group(0) @binding(1) var<storage, read_write> src1: array<SRC1_TYPE>;
#endif
@group(0) @binding(2) var<storage, read_write> dst: array<f32>;
// "mul_mat_vec_acc.tmpl" requires params.k, params.m, params.stride_01
@group(0) @binding(3) var<uniform> params: MulMatParams;
@@ -75,10 +88,15 @@ fn main(
let src12_idx = dst2_idx;
let src0_batch_offset = params.offset_src0 + src03_idx * params.stride_03 + src02_idx * params.stride_02;
let src1_idx_base = params.offset_src1 + src13_idx * params.stride_13 + src12_idx * params.stride_12;
let dst_idx_base = params.offset_dst + dst3_idx * dst3_stride + dst2_idx * dst2_stride + row_base;
#ifdef MMVQ
let src1q_idx_base = (src13_idx * params.bs02 * params.broadcast2 + src12_idx) * (params.k / 32u);
let acc = accumulate_vec_q_dot(thread_id, row_base, src0_batch_offset, src1q_idx_base);
#else
let src1_idx_base = params.offset_src1 + src13_idx * params.stride_13 + src12_idx * params.stride_12;
let acc = accumulate_vec_dot(thread_id, row_base, src0_batch_offset, src1_idx_base);
#endif
#ifdef USE_SUBGROUP_REDUCTION
for (var row = 0u; row < OUTPUTS_PER_WG; row++) {