SDPA tests: select attn output by shape#20285
Open
JulianCloudNTH wants to merge 1 commit into
Open
Conversation
Summary: Follow-up to D108625761 (pytorch#20283), which fixed `test_sdpa_config`'s ambiguous attention-output selection; this handles the remaining edge cases in the sibling SDPA test functions. `sdpa_with_kv_cache` returns `[k_cache, v_cache, attn_output]`, where the attention output is `[1, S, Hq, D]` and each KV cache is `[1, Cmax, Hkv, D]`. Selecting the attention output by flat element count is ambiguous whenever `S*Hq == Cmax*Hkv` (all three tensors then share a `numel`) — the deterministic `llama1b_prefill` failure D108625761 fixed in `test_sdpa_config`. The three other SDPA test functions — `test_sdpa_replay`, `test_sdpa_dynamic_decode`, and `test_sdpa_incache_decode` — still matched by `numel`; no currently-configured sequence triggers the collision, but they carry the same latent ambiguity. This applies the same shape-based selection: match the attention output as the 4-D tensor `[1, S, Hq, D]` and (for replay and dynamic decode) classify the two caches as `[1, Cmax, Hkv, D]`, keeping the per-step k-vs-v identification by content unchanged. Test-only; no kernel, runtime, or export change. Authored with Claude Code. Differential Revision: D108650388
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20285
Note: Links to docs will display an error until the docs builds have been completed. ❌ 3 New Failures, 4 PendingAs of commit 2762587 with merge base 99ca02f ( NEW FAILURES - The following jobs have failed:
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 D108650388. |
This PR needs a
|
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
1 participant
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:
Follow-up to D108625761 (#20283), which fixed
test_sdpa_config's ambiguous attention-output selection; this handles the remaining edge cases in the sibling SDPA test functions.sdpa_with_kv_cachereturns[k_cache, v_cache, attn_output], where the attention output is[1, S, Hq, D]and each KV cache is[1, Cmax, Hkv, D]. Selecting the attention output by flat element count is ambiguous wheneverS*Hq == Cmax*Hkv(all three tensors then share anumel) — the deterministicllama1b_prefillfailure D108625761 fixed intest_sdpa_config. The three other SDPA test functions —test_sdpa_replay,test_sdpa_dynamic_decode, andtest_sdpa_incache_decode— still matched bynumel; no currently-configured sequence triggers the collision, but they carry the same latent ambiguity. This applies the same shape-based selection: match the attention output as the 4-D tensor[1, S, Hq, D]and (for replay and dynamic decode) classify the two caches as[1, Cmax, Hkv, D], keeping the per-step k-vs-v identification by content unchanged. Test-only; no kernel, runtime, or export change.Authored with Claude Code.
Differential Revision: D108650388