[ExecuTorch][WebGPU] Test byte-packed BOOL tails - #21664
Merged
Merged
Conversation
**Exercise byte-packed BOOL tails through export and native execution** Manifest and export coverage now proves exact `BOOL` storage and numeric `BOOL`-to-fp32 conversion across word-aligned and tail lengths. The native driver preserves `BOOL` metadata and compares canonical raw 0/1 bytes rather than semantic truth values. Key changes: - Length coverage — exercises 1, 4, 5, and 67 elements. - Export assertions — keep the fp32 cast inside the delegate while `INT8` stays outside the `BOOL` route. - Native validation — checks direct inputs, comparison outputs, and raw tail bytes. This is coverage-only; runtime kernels are unchanged. Co-authored-with: Claude Code. Differential Revision: [D114936146](https://our.internmc.facebook.com/intern/diff/D114936146/) ghstack-source-id: 411044629 Pull-Request: #21599
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21664
Note: Links to docs will display an error until the docs builds have been completed. ⏳ No Failures, 16 PendingAs of commit f7045ea with merge base ad3a71f ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
**Add a buffer-only groups=1 Conv1d kernel for Voxtral encoder shapes** The WebGPU backend handled only K=1 pointwise and depthwise Conv1d, leaving non-transposed groups=1 K=3 encoder convolutions unsupported. **Before** - Groups=1 K>1 graphs had no valid WebGPU route. **After** - A direct NCL buffer kernel supports groups=1 convolution with stride, padding, dilation, and optional bias. Key changes: - `add_conv1d_node` — validates fp32 operands, geometry, shapes, and shader index ranges. - `conv1d.wgsl` — mirrors Vulkan output indexing in `backends/vulkan/runtime/graph/ops/glsl/conv1d.glsl:78`, stride/padding at `:86`, and the kernel FMA loop at `:108`. - Dispatch and resize hooks — fold x/y workgroups and recompute live parameters safely. - Generated sources — refresh WGSL headers, registry output, and drift digests. The route remains limited to non-transposed groups=1 fp32 convolution; existing pointwise and depthwise routes are unchanged. Co-authored-with: Claude Code. Differential Revision: [D114936145](https://our.internmc.facebook.com/intern/diff/D114936145/) ghstack-source-id: 411044636 Pull-Request: #21600
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: #21599 by @JCNTH
^ 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/JCNTH/206/base
ghstack PR head: https://github.com/pytorch/executorch/tree/gh/JCNTH/206/head
Merge bot PR base: https://github.com/pytorch/executorch/tree/gh/JCNTH/205/orig
Merge bot PR head: https://github.com/pytorch/executorch/tree/gh/JCNTH/206/orig
Differential Revision: D114936146
@diff-train-skip-merge