[ExecuTorch][WebGPU] Enable FlashDecoding by default for decode SDPA (runtime shape gate)#20586
Merged
Merged
Conversation
…(runtime shape gate) Pull Request resolved: #20544 **Makes split-KV FlashDecoding the default decode-attention path** (it was shipped dormant behind a default-OFF compile flag). FD is the fastest WebGPU SDPA decode arm (**+178% vs naive**, M4 Pro, isolated op); this turns it on for production and selects it at runtime by a shape-capability predicate. {F1991715077} **Problem:** the FD kernel is correct and measured (+178%) but compile-gated OFF, so no production build used it. A device-limit gate (web-llm-style `maxStorageBufferBindingSize`) was considered but is dead code here: FD's resource needs (workgroup size 64, 512 B shared memory, 5 storage bindings) are all below WebGPU's baseline minimum limits, and FD binds the same K/V caches as the materialized fallback — so no spec-compliant device can run materialized decode but fail FD. The only selection criterion with real effect is shape. **Solution:** enable FD by default and select it at runtime on shape, not device. - **Before:** `EXECUTORCH_BUILD_WEBGPU_SDPA_FD` default OFF; FD code unlinked; every decode used the materialized QK/softmax/AV path. - **After:** flag default ON (kept as a build-time kill-switch); decode (`S == 1`, static input_pos) with head dim `<= kSdpaFdMaxHeadDim` uses FD; other shapes (including head dim > 128) fall through to the materialized path. **Implementation:** - `Sdpa.cpp`: extend the FD selection predicate with `D <= kSdpaFdMaxHeadDim` so unsupported head dims fall through instead of throwing. - `SdpaFdDecode.h`: expose `kSdpaFdMaxHeadDim` (FD's lane-owns-D reach) as the single source of truth; `SdpaFdDecode.cpp` ties it to `WG_SIZE * MAX_D_PER_LANE` with a `static_assert`. - `CMakeLists.txt` (fbcode + xplat): flip the option default to ON; OFF remains a kill-switch that drops all FlashDecoding code. - `test_webgpu_native_ci.sh`: drop the now-redundant explicit `=ON` flag so CI builds and tests the default. - Mirrors Vulkan `backends/vulkan/runtime/graph/ops/impl/SDPA.cpp` shape-based kernel selection (`is_single_token`); no device-adaptive gate, matching the Vulkan delegate. **Constraints:** decode-only (`S == 1`), static input_pos (dynamic-pos decode still uses the materialized path); fp32, buffer-only; the FD kernels are unchanged by this diff. Co-authored with Claude Code. ghstack-source-id: 397864676 @exported-using-ghexport @diff-train-skip-merge Differential Revision: [D109520722](https://our.internmc.facebook.com/intern/diff/D109520722/)
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20586
Note: Links to docs will display an error until the docs builds have been completed. ⏳ No Failures, 87 PendingAs of commit 597d658 with merge base 55a71e6 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
This PR needs a
|
JulianCloudNTH
approved these changes
Jun 28, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This PR was created by the merge bot to help merge the original PR into the main branch.
ghstack PR number: #20544 by @JulianCloudNTH
^ Please use this as the source of truth for the PR details, comments, and reviews
ghstack PR base: https://github.com/pytorch/executorch/tree/gh/JulianCloudNTH/64/base
ghstack PR head: https://github.com/pytorch/executorch/tree/gh/JulianCloudNTH/64/head
Merge bot PR base: https://github.com/pytorch/executorch/tree/main
Merge bot PR head: https://github.com/pytorch/executorch/tree/gh/JulianCloudNTH/64/orig
@diff-train-skip-merge