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

ggml webgpu: fix workgroup dispatch limit for large batch sizes (#19965)

Browse Source 
* ggml-webgpu: fix workgroup dispatch limit for large batch sizes

WebGPU limits workgroup sizes to 65535 per dimension. Large MUL_MAT
operations with batch sizes exceedeing this limi would fail.

* add compute_2d_workgroups() helper to split total workgroup ID across
X/Y dimensions

* update mul_mat_reg_tile.wgsl to reconstruct linear workgroup ID from 2D
   dispatch

* update mul_mat_subgroup_matrix.wgsl to reconstruct linear workgroup ID
  from 2D dispatch

* update mul_mat.wgsl to compute global index from 2D workgroup
  coordinates

* refactor all three mul_mat dispatch paths to use the shared helper

* ggml-webgpu: add bounds checking for over-dispatched workgroups

2D workgroup dispatch can over-dispatch when total workgroups don't
divide evenly into the 65535 per-dimension limit. Extra workgroups
would compute invalid batch indices, causing memory corruption.

* add batch_idx bound check to mul_mat_reg_tile.wgsl and
mul_mat_subgroup_matrix.wgsl to prevent over-dispatched workgroups
from accessing invalid memory

* fixes test failures with large batch sizes (eg., bs=[128, 1024])

* ggml-webgpu: add back TODO for spliting large sizes into batches

* Optimize 2d workgroup provisioning

* Set some parameters that increase speed

---------

Co-authored-by: Reese Levine <reeselevine1@gmail.com>
This commit is contained in:
Abhijit Ramesh
2026-03-02 19:35:11 -08:00
committed by GitHub
co-authored by Reese Levine
parent 4d828bd1ab
commit 49a7564ac1
4 changed files with 49 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_subgroup_matrix.wgsl
+11 -3
View File
@@ -69,7 +69,8 @@ var<workgroup> shmem: array<f16, SHMEM_SIZE>;
@compute @workgroup_size(TOTAL_WORKGROUP_SIZE)
fn main(@builtin(workgroup_id) wg_id: vec3<u32>,
@builtin(local_invocation_id) local_id: vec3<u32>,
@builtin(subgroup_id) subgroup_id: u32) {
@builtin(subgroup_id) subgroup_id: u32,
@builtin(num_workgroups) num_wg: vec3<u32>) {
let thread_id = local_id.x;
let subgroup_m = subgroup_id % SUBGROUP_M;
@@ -79,9 +80,16 @@ fn main(@builtin(workgroup_id) wg_id: vec3<u32>,
let wg_n_count = (params.n + WG_N_SG_TILE_SIZE - 1) / WG_N_SG_TILE_SIZE;
let wg_per_matrix = wg_m_count * wg_n_count;
let batch_idx = wg_id.x / wg_per_matrix;
let wg_linear = wg_id.y * num_wg.x + wg_id.x;
let wg_in_batch = wg_id.x % wg_per_matrix;
let batch_idx = wg_linear / wg_per_matrix;
let total_batches = params.bs02 * params.broadcast2 * params.bs03 * params.broadcast3;
if (batch_idx >= total_batches) {
return;
}
let wg_in_batch = wg_linear % wg_per_matrix;
let wg_m = wg_in_batch % wg_m_count;
let wg_n = wg_in_batch / wg_m_count;