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

ggml-webgpu: improve MTP inference by using mat-vec path for small batches (#24811)

Browse Source 
* ggml-webgpu: improve small batches decoding

* Add barrier to the NUM_COLS loop in mul-mat-vec
This commit is contained in:
Masashi Yoshimura
2026-06-23 17:13:55 +09:00
committed by GitHub
parent 035cd8f9a6
commit 7c908502ea
8 changed files with 680 additions and 589 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_vec.wgsl
+2 -2
View File
@@ -103,7 +103,7 @@ fn main(
#ifdef USE_SUBGROUP_REDUCTION
for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
let subgroup_total = subgroupAdd(acc[row]);
let subgroup_total = subgroupAdd(acc[0][row]);
if (subgroup_invocation_id == 0u) {
partial_sums[partial_index(row, subgroup_id)] = subgroup_total;
}
@@ -126,7 +126,7 @@ fn main(
#ifdef USE_WORKGROUP_REDUCTION
for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
partial_sums[partial_index(row, thread_id)] = acc[row];
partial_sums[partial_index(row, thread_id)] = acc[0][row];
}
workgroupBarrier();