Repository navigation
Phonon: run the Mimi transposed convolutions as convolutions on Vulkan, MLX and CUDA (#25020) - #25020
Merged
Merged
Conversation
psiddh
requested review from
SS-JIA,
digantdesai and
larryliu0820
as code owners
November 13, 2026 05:56
pytorch-bot
Bot
added
the
module: vulkan
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/25020
Note: Links to docs will display an error until the docs builds have been completed. ⏳ No Failures, 2 PendingAs of commit 3bf88e8 with merge base 029e351 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Contributor
|
@psiddh has exported this pull request. If you are a Meta employee, you can view the originating Diff in D131642733. |
This PR needs a
|
meta-codesync
Bot
changed the title
psiddh
force-pushed
the
export-D131642733
branch
from
November 13, 2026 08:50
ca106de to
75fbfe9
Compare
psiddh
added a commit
to psiddh/executorch
that referenced
this pull request
Nov 13, 2026
…n, MLX and CUDA (pytorch#25020) Summary: Pull Request resolved: pytorch#25020 Applies the Mimi rewrite from the previous diff (conv_transpose1d_as_conv1d plus depthwise_conv_transpose_as_mul) to Vulkan, MLX and CUDA as well. Vulkan previously ran the decoder's transposed convolutions through conv_transpose1d_as_2d, which is now used only by --webgpu-compat-ops. The CUDA backend has no Triton kernel for transposed convolution and runs each one as a convolution over a zero-stuffed input, which does stride times the work; the rewrite avoids that. DepthwiseConvTranspose1dAsMul now creates its zeros with x.new_zeros, so they follow the input to the GPU (torch.zeros put them on the CPU and broke the CUDA export), and the two swapped modules are in WEIGHT_MODULES, so --dtype bf16 still computes them in bf16. This diff was written with the help of Claude Code (Claude Opus 5.5). Differential Revision: D131642733
psiddh
force-pushed
the
export-D131642733
branch
from
November 13, 2026 17:28
75fbfe9 to
141726b
Compare
psiddh
added a commit
to psiddh/executorch
that referenced
this pull request
Nov 13, 2026
…n, MLX and CUDA (pytorch#25020) Summary: Pull Request resolved: pytorch#25020 Applies the Mimi rewrite from the previous diff (conv_transpose1d_as_conv1d plus depthwise_conv_transpose_as_mul) to Vulkan, MLX and CUDA as well. Vulkan previously ran the decoder's transposed convolutions through conv_transpose1d_as_2d, which is now used only by --webgpu-compat-ops. The CUDA backend has no Triton kernel for transposed convolution and runs each one as a convolution over a zero-stuffed input, which does stride times the work; the rewrite avoids that. DepthwiseConvTranspose1dAsMul now creates its zeros with x.new_zeros, so they follow the input to the GPU (torch.zeros put them on the CPU and broke the CUDA export), and the two swapped modules are in WEIGHT_MODULES, so --dtype bf16 still computes them in bf16. This diff was written with the help of Claude Code (Claude Opus 5.5). Differential Revision: D131642733
psiddh
added a commit
to psiddh/executorch
that referenced
this pull request
Nov 13, 2026
…n, MLX and CUDA (pytorch#25020) Summary: Pull Request resolved: pytorch#25020 Applies the Mimi rewrite from the previous diff (conv_transpose1d_as_conv1d plus depthwise_conv_transpose_as_mul) to Vulkan, MLX and CUDA as well. Vulkan previously ran the decoder's transposed convolutions through conv_transpose1d_as_2d, which is now used only by --webgpu-compat-ops. The CUDA backend has no Triton kernel for transposed convolution and runs each one as a convolution over a zero-stuffed input, which does stride times the work; the rewrite avoids that. DepthwiseConvTranspose1dAsMul now creates its zeros with x.new_zeros, so they follow the input to the GPU (torch.zeros put them on the CPU and broke the CUDA export), and the two swapped modules are in WEIGHT_MODULES, so --dtype bf16 still computes them in bf16. This diff was written with the help of Claude Code (Claude Opus 5.5). Differential Revision: D131642733
Gasoonjia
approved these changes
Nov 13, 2026
Gasoonjia
left a comment
Gasoonjia
left a comment
Contributor
There was a problem hiding this comment.
Review automatically exported from Phabricator review in Meta.
psiddh
force-pushed
the
export-D131642733
branch
from
November 13, 2026 19:07
141726b to
3602271
Compare
psiddh
added a commit
to psiddh/executorch
that referenced
this pull request
Nov 13, 2026
…n, MLX and CUDA (pytorch#25020) Summary: Pull Request resolved: pytorch#25020 Applies the Mimi rewrite from the previous diff (conv_transpose1d_as_conv1d plus depthwise_conv_transpose_as_mul) to Vulkan, MLX and CUDA as well. Vulkan previously ran the decoder's transposed convolutions through conv_transpose1d_as_2d, which is now used only by --webgpu-compat-ops. The CUDA backend has no Triton kernel for transposed convolution and runs each one as a convolution over a zero-stuffed input, which does stride times the work; the rewrite avoids that. DepthwiseConvTranspose1dAsMul now creates its zeros with x.new_zeros, so they follow the input to the GPU (torch.zeros put them on the CPU and broke the CUDA export), and the two swapped modules are in WEIGHT_MODULES, so --dtype bf16 still computes them in bf16. This diff was written with the help of Claude Code (Claude Opus 5.5). Reviewed By: Gasoonjia Differential Revision: D131642733
psiddh
added a commit
to psiddh/executorch
that referenced
this pull request
Nov 13, 2026
…n, MLX and CUDA (pytorch#25020) Summary: Pull Request resolved: pytorch#25020 Applies the Mimi rewrite from the previous diff (conv_transpose1d_as_conv1d plus depthwise_conv_transpose_as_mul) to Vulkan, MLX and CUDA as well. Vulkan previously ran the decoder's transposed convolutions through conv_transpose1d_as_2d, which is now used only by --webgpu-compat-ops. The CUDA backend has no Triton kernel for transposed convolution and runs each one as a convolution over a zero-stuffed input, which does stride times the work; the rewrite avoids that. DepthwiseConvTranspose1dAsMul now creates its zeros with x.new_zeros, so they follow the input to the GPU (torch.zeros put them on the CPU and broke the CUDA export), and the two swapped modules are in WEIGHT_MODULES, so --dtype bf16 still computes them in bf16. This diff was written with the help of Claude Code (Claude Opus 5.5). Reviewed By: Gasoonjia Differential Revision: D131642733
psiddh
added a commit
to psiddh/executorch
that referenced
this pull request
Nov 14, 2026
…n, MLX and CUDA (pytorch#25020) Summary: Applies the Mimi rewrite from the previous diff (conv_transpose1d_as_conv1d plus depthwise_conv_transpose_as_mul) to Vulkan, MLX and CUDA as well. Vulkan previously ran the decoder's transposed convolutions through conv_transpose1d_as_2d, which is now used only by --webgpu-compat-ops. The CUDA backend has no Triton kernel for transposed convolution and runs each one as a convolution over a zero-stuffed input, which does stride times the work; the rewrite avoids that. DepthwiseConvTranspose1dAsMul now creates its zeros with x.new_zeros, so they follow the input to the GPU (torch.zeros put them on the CPU and broke the CUDA export), and the two swapped modules are in WEIGHT_MODULES, so --dtype bf16 still computes them in bf16. This diff was written with the help of Claude Code (Claude Opus 5.5). Reviewed By: Gasoonjia, shoumikhin Differential Revision: D131642733
psiddh
added a commit
to psiddh/executorch
that referenced
this pull request
Nov 14, 2026
…n, MLX and CUDA (pytorch#25020) Summary: Applies the Mimi rewrite from the previous diff (conv_transpose1d_as_conv1d plus depthwise_conv_transpose_as_mul) to Vulkan, MLX and CUDA as well. Vulkan previously ran the decoder's transposed convolutions through conv_transpose1d_as_2d, which is now used only by --webgpu-compat-ops. The CUDA backend has no Triton kernel for transposed convolution and runs each one as a convolution over a zero-stuffed input, which does stride times the work; the rewrite avoids that. DepthwiseConvTranspose1dAsMul now creates its zeros with x.new_zeros, so they follow the input to the GPU (torch.zeros put them on the CPU and broke the CUDA export), and the two swapped modules are in WEIGHT_MODULES, so --dtype bf16 still computes them in bf16. This diff was written with the help of Claude Code (Claude Opus 5.5). Reviewed By: Gasoonjia, shoumikhin Differential Revision: D131642733
psiddh
force-pushed
the
export-D131642733
branch
from
November 14, 2026 03:19
3602271 to
0876f42
Compare
psiddh
added a commit
to psiddh/executorch
that referenced
this pull request
Nov 14, 2026
…n, MLX and CUDA (pytorch#25020) Summary: Applies the Mimi rewrite from the previous diff (conv_transpose1d_as_conv1d plus depthwise_conv_transpose_as_mul) to Vulkan, MLX and CUDA as well. Vulkan previously ran the decoder's transposed convolutions through conv_transpose1d_as_2d, which is now used only by --webgpu-compat-ops. The CUDA backend has no Triton kernel for transposed convolution and runs each one as a convolution over a zero-stuffed input, which does stride times the work; the rewrite avoids that. DepthwiseConvTranspose1dAsMul now creates its zeros with x.new_zeros, so they follow the input to the GPU (torch.zeros put them on the CPU and broke the CUDA export), and the two swapped modules are in WEIGHT_MODULES, so --dtype bf16 still computes them in bf16. This diff was written with the help of Claude Code (Claude Opus 5.5). Reviewed By: Gasoonjia, shoumikhin Differential Revision: D131642733
psiddh
force-pushed
the
export-D131642733
branch
from
November 14, 2026 07:50
0876f42 to
6eff255
Compare
psiddh
added a commit
to psiddh/executorch
that referenced
this pull request
Nov 14, 2026
…n, MLX and CUDA (pytorch#25020) Summary: Applies the Mimi rewrite from the previous diff (conv_transpose1d_as_conv1d plus depthwise_conv_transpose_as_mul) to Vulkan, MLX and CUDA as well. Vulkan previously ran the decoder's transposed convolutions through conv_transpose1d_as_2d, which is now used only by --webgpu-compat-ops. The CUDA backend has no Triton kernel for transposed convolution and runs each one as a convolution over a zero-stuffed input, which does stride times the work; the rewrite avoids that. DepthwiseConvTranspose1dAsMul now creates its zeros with x.new_zeros, so they follow the input to the GPU (torch.zeros put them on the CPU and broke the CUDA export), and the two swapped modules are in WEIGHT_MODULES, so --dtype bf16 still computes them in bf16. This diff was written with the help of Claude Code (Claude Opus 5.5). Reviewed By: Gasoonjia, shoumikhin Differential Revision: D131642733
…n, MLX and CUDA (pytorch#25020) Summary: Applies the Mimi rewrite from the previous diff (conv_transpose1d_as_conv1d plus depthwise_conv_transpose_as_mul) to Vulkan, MLX and CUDA as well. Vulkan previously ran the decoder's transposed convolutions through conv_transpose1d_as_2d, which is now used only by --webgpu-compat-ops. The CUDA backend has no Triton kernel for transposed convolution and runs each one as a convolution over a zero-stuffed input, which does stride times the work; the rewrite avoids that. DepthwiseConvTranspose1dAsMul now creates its zeros with x.new_zeros, so they follow the input to the GPU (torch.zeros put them on the CPU and broke the CUDA export), and the two swapped modules are in WEIGHT_MODULES, so --dtype bf16 still computes them in bf16. This diff was written with the help of Claude Code (Claude Opus 5.5). Reviewed By: Gasoonjia, shoumikhin Differential Revision: D131642733
psiddh
force-pushed
the
export-D131642733
branch
from
November 14, 2026 17:13
6eff255 to
3bf88e8
Compare
psiddh
added a commit
to psiddh/executorch
that referenced
this pull request
Nov 14, 2026
…n, MLX and CUDA (pytorch#25020) Summary: Applies the Mimi rewrite from the previous diff (conv_transpose1d_as_conv1d plus depthwise_conv_transpose_as_mul) to Vulkan, MLX and CUDA as well. Vulkan previously ran the decoder's transposed convolutions through conv_transpose1d_as_2d, which is now used only by --webgpu-compat-ops. The CUDA backend has no Triton kernel for transposed convolution and runs each one as a convolution over a zero-stuffed input, which does stride times the work; the rewrite avoids that. DepthwiseConvTranspose1dAsMul now creates its zeros with x.new_zeros, so they follow the input to the GPU (torch.zeros put them on the CPU and broke the CUDA export), and the two swapped modules are in WEIGHT_MODULES, so --dtype bf16 still computes them in bf16. This diff was written with the help of Claude Code (Claude Opus 5.5). Reviewed By: Gasoonjia, shoumikhin Differential Revision: D131642733
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
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:
Applies the Mimi rewrite from the previous diff (conv_transpose1d_as_conv1d plus depthwise_conv_transpose_as_mul) to Vulkan, MLX and CUDA as well. Vulkan previously ran the decoder's transposed convolutions through conv_transpose1d_as_2d, which is now used only by --webgpu-compat-ops. The CUDA backend has no Triton kernel for transposed convolution and runs each one as a convolution over a zero-stuffed input, which does stride times the work; the rewrite avoids that.
DepthwiseConvTranspose1dAsMul now creates its zeros with x.new_zeros, so they follow the input to the GPU (torch.zeros put them on the CPU and broke the CUDA export), and the two swapped modules are in WEIGHT_MODULES, so --dtype bf16 still computes them in bf16.
This diff was written with the help of Claude Code (Claude Opus 5.5).
Reviewed By: Gasoonjia, shoumikhin
Differential Revision: D131642733