[ExecuTorch][WebGPU] Add fused SDPA (sdpa_with_kv_cache) with dynamic input_pos#20261
Closed
pytorchbot wants to merge 3 commits into
Closed
[ExecuTorch][WebGPU] Add fused SDPA (sdpa_with_kv_cache) with dynamic input_pos#20261pytorchbot wants to merge 3 commits into
pytorchbot wants to merge 3 commits into
Conversation
…ation) Pull Request resolved: #20201 Backend-agnostic GPU-timestamp infrastructure, split out so the general implementation is foundational (below SDPA) while the SDPA-specific dispatch labeling stays above the SDPA op. Composed of: `WebGPUQueryPool`, a faithful re-port of Vulkan's `vkapi::QueryPool` (`backends/vulkan/runtime/vk_api/QueryPool.{h,cpp}`) — same `ShaderDuration` data model and ticks->ns conversion; three deviations are forced by the WebGPU API (per-dispatch bracketing via a compute-pass `timestampWrites` descriptor since there is no mid-encoder `writeTimestamp`; readback via `resolveQuerySet` + buffer map rather than host-side `vkGetQueryPoolResults`; the `TimestampQuery` capability requested as an explicit device feature, fail-open if the adapter lacks it). `WebGPUDevice` gains timestamp-feature detection, and `WebGPUGraph` gains a per-dispatch `kernel_name` label plus `execute()` bracketing of each compute pass when the pool is active. Opt-in via the `WEBGPU_TIMESTAMP_QUERY` env var; off by default, so the production `execute()` path is byte-identical. The SDPA per-kernel labeling lives in the companion "for SDPA" diff above the SDPA op. Co-authored with Claude. ghstack-source-id: 392975889 @exported-using-ghexport Differential Revision: [D108188287](https://our.internmc.facebook.com/intern/diff/D108188287/)
… input_pos Pull Request resolved: #20086 Adds the fused `sdpa_with_kv_cache` op (QK attention-weights, softmax, attention-output sub-kernels over the KV cache), composing the three enablers below it: the base graph's inter-dispatch buffer passing (scratch buffers + multi-pass execute), the `update_cache` op, and the SymInt live-scalar mechanism. The QK/softmax/AV kernels mirror the Vulkan reference's flat-index/GQA/causal-mask math (NCHW, buffer-only, fp32). `input_pos` is consumed dynamically via the SymInt mechanism: the op reads `symint_buffer()` as a uniform, sizes its scratch + dispatches for the max context length, and registers a resize hook so a single delegate runs an autoregressive decode loop (feed only the new token + advancing `input_pos`) instead of a fixed baked position. Mirrors the Vulkan SymInt = live uniform-buffer design. Tests live in the stacked test-suite diff above (clean op diff here). Authored with assistance from Claude. ghstack-source-id: 392609088 @exported-using-ghexport Differential Revision: [D107595125](https://our.internmc.facebook.com/intern/diff/D107595125/)
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20261
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
JulianCloudNTH
requested review from
JacobSzwejbka,
SS-JIA,
abhinaykukkadapu,
digantdesai,
manuelcandales,
mergennachin,
psiddh,
rascani and
robert-kalmar
as code owners
JulianCloudNTH
approved these changes
Jun 13, 2026
JulianCloudNTH
temporarily deployed
to
upload-benchmark-results
GitHub Actions
Inactive
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: #20086 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/19/base
ghstack PR head: https://github.com/pytorch/executorch/tree/gh/JulianCloudNTH/19/head
Merge bot PR base: https://github.com/pytorch/executorch/tree/gh/JulianCloudNTH/22/orig
Merge bot PR head: https://github.com/pytorch/executorch/tree/gh/JulianCloudNTH/19/orig
@diff-train-skip-merge
cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani