Skip to content

[ExecuTorch][WebGPU] SDPA: skip QK contraction for fully-masked causal tiles #20492

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
JulianCloudNTH wants to merge 2 commits into
base: gh/JulianCloudNTH/62/base
Choose a base branch
from
+7 −3
Open

[ExecuTorch][WebGPU] SDPA: skip QK contraction for fully-masked causal tiles #20492

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
40acc76
Update
JulianCloudNTH Jun 24, 2026
173487b
Update
JulianCloudNTH Jun 25, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 3 additions & 1 deletion backends/webgpu/runtime/ops/sdpa/sdpa_compute_attn_weights.wgsl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -85,9 +85,11 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
acc[2] = vec4<f32>(0.0, 0.0, 0.0, 0.0);
acc[3] = vec4<f32>(0.0, 0.0, 0.0, 0.0);

// Skip fully-masked causal tiles; mirrors Vulkan attn_weights_tiled.glsl.
let skip_tile = c0 > s0 + (TM - 1u) + params.input_pos;
var d4: u32 = 0u;
loop {
if (d4 >= params.D) {
if (d4 >= params.D || skip_tile) {
break;
}
var q: array<vec4<f32>, TM>;
Expand Down
6 changes: 4 additions & 2 deletions backends/webgpu/runtime/ops/sdpa/sdpa_compute_attn_weights_wgsl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,7 @@
namespace executorch::backends::webgpu {

// @generated from sdpa_compute_attn_weights.wgsl - DO NOT EDIT.
// wgsl-sha256: e9bdc2272bba2716655e96be0597571701b4ec1496f78a85660e29d108f655f9
// wgsl-sha256: d177264689e6c50e1794a0599808f3cfe6f30ba99c5084d3c8324da4b9f89d10
inline constexpr const char* kSdpaComputeAttnWeightsWGSL = R"(
@group(0) @binding(0) var<storage, read_write> t_attn_weights: array<f32>;
@group(0) @binding(1) var<storage, read> t_q: array<f32>;
Expand Down Expand Up @@ -102,9 +102,11 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
acc[2] = vec4<f32>(0.0, 0.0, 0.0, 0.0);
acc[3] = vec4<f32>(0.0, 0.0, 0.0, 0.0);

// Skip fully-masked causal tiles; mirrors Vulkan attn_weights_tiled.glsl.
let skip_tile = c0 > s0 + (TM - 1u) + params.input_pos;
var d4: u32 = 0u;
loop {
if (d4 >= params.D) {
if (d4 >= params.D || skip_tile) {
break;
}
var q: array<vec4<f32>, TM>;
Expand Down
Loading