Fix Dawn nightly: select SDPA attn output by shape, not numel (#20283)#20283
Merged
meta-codesync[bot] merged 1 commit intoJun 15, 2026
Merged
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20283
Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Contributor
|
@JulianCloudNTH has exported this pull request. If you are a Meta employee, you can view the originating Diff in D108625761. |
This PR needs a
|
Gasoonjia
approved these changes
Jun 15, 2026
meta-codesync
Bot
changed the title
JulianCloudNTH
added a commit
to JulianCloudNTH/executorch
that referenced
this pull request
Jun 15, 2026
…h#20283) Summary: The WebGPU Dawn native nightly (`webgpu_native_test`) fails deterministically on the `llama1b_prefill` SDPA config with `FAIL: ambiguous attention output: 3 tensors match numel 262144`, which fails the binary and turns the job red. `sdpa_with_kv_cache` returns three tensors `[k_cache, v_cache, attn_output]`. `test_sdpa_config` identified the attention output purely by element count (`numel == S*Hq*D`). For `llama1b_prefill` (`Hq=32, Hkv=8, D=64, S=128, Cmax=512`) the attention count `S*Hq*D = 128*32*64 = 262144` coincides exactly with each cache count `Cmax*Hkv*D = 512*8*64 = 262144`, so all three outputs match `numel` and the existing ambiguity guard correctly bails before any numeric comparison. The kernel output itself is fine -- the sibling `llama1b_decode` config (same `Hq/Hkv/D`) passes at `~1e-9`; only the test's output-selection heuristic was wrong. The colliding config and the numel selector were introduced together in D107595144. Fix: disambiguate by shape instead of flat count. The attention output is `[1, S, Hq, D]` while each cache is `[1, Cmax, Hkv, D]`; these differ in dims 1-2 even when the flat count collides. Match `dim()==4 && size(1)==S && size(2)==Hq && size(3)==D`, keeping the `attn_matches > 1` ambiguity guard as a backstop. Scope: test-only, one function (`test_sdpa_config`); no kernel, runtime, or export change. Authored with Claude Code. Reviewed By: Gasoonjia Differential Revision: D108625761
JulianCloudNTH
force-pushed
the
export-D108625761
branch
from
3186b53 to
63b3389
Compare
…h#20283) Summary: The WebGPU Dawn native nightly (`webgpu_native_test`) fails deterministically on the `llama1b_prefill` SDPA config with `FAIL: ambiguous attention output: 3 tensors match numel 262144`, which fails the binary and turns the job red. `sdpa_with_kv_cache` returns three tensors `[k_cache, v_cache, attn_output]`. `test_sdpa_config` identified the attention output purely by element count (`numel == S*Hq*D`). For `llama1b_prefill` (`Hq=32, Hkv=8, D=64, S=128, Cmax=512`) the attention count `S*Hq*D = 128*32*64 = 262144` coincides exactly with each cache count `Cmax*Hkv*D = 512*8*64 = 262144`, so all three outputs match `numel` and the existing ambiguity guard correctly bails before any numeric comparison. The kernel output itself is fine -- the sibling `llama1b_decode` config (same `Hq/Hkv/D`) passes at `~1e-9`; only the test's output-selection heuristic was wrong. The colliding config and the numel selector were introduced together in D107595144. Fix: disambiguate by shape instead of flat count. The attention output is `[1, S, Hq, D]` while each cache is `[1, Cmax, Hkv, D]`; these differ in dims 1-2 even when the flat count collides. Match `dim()==4 && size(1)==S && size(2)==Hq && size(3)==D`, keeping the `attn_matches > 1` ambiguity guard as a backstop. Scope: test-only, one function (`test_sdpa_config`); no kernel, runtime, or export change. Authored with Claude Code. Reviewed By: Gasoonjia Differential Revision: D108625761
JulianCloudNTH
force-pushed
the
export-D108625761
branch
from
63b3389 to
717a011
Compare
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.
Summary:
The WebGPU Dawn native nightly (
webgpu_native_test) fails deterministically on thellama1b_prefillSDPA config withFAIL: ambiguous attention output: 3 tensors match numel 262144, which fails the binary and turns the job red.sdpa_with_kv_cachereturns three tensors[k_cache, v_cache, attn_output].test_sdpa_configidentified the attention output purely by element count (numel == S*Hq*D). Forllama1b_prefill(Hq=32, Hkv=8, D=64, S=128, Cmax=512) the attention countS*Hq*D = 128*32*64 = 262144coincides exactly with each cache countCmax*Hkv*D = 512*8*64 = 262144, so all three outputs matchnumeland the existing ambiguity guard correctly bails before any numeric comparison. The kernel output itself is fine -- the siblingllama1b_decodeconfig (sameHq/Hkv/D) passes at~1e-9; only the test's output-selection heuristic was wrong. The colliding config and the numel selector were introduced together in D107595144.Fix: disambiguate by shape instead of flat count. The attention output is
[1, S, Hq, D]while each cache is[1, Cmax, Hkv, D]; these differ in dims 1-2 even when the flat count collides. Matchdim()==4 && size(1)==S && size(2)==Hq && size(3)==D, keeping theattn_matches > 1ambiguity guard as a backstop.Scope: test-only, one function (
test_sdpa_config); no kernel, runtime, or export change.Authored with Claude Code.
Reviewed By: Gasoonjia
Differential Revision: D108625761