[voxtral_tts] enable MLX backend#19177
Open
seyeong-han wants to merge 8 commits intopytorch:mainfrom
Open
Conversation
seyeong-han
requested review from
kirklandsign,
larryliu0820 and
lucylq
as code owners
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/19177
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: ❌ 1 New Failure, 2 Unrelated FailuresAs of commit ba5b038 with merge base f3e49ff ( NEW FAILURE - The following job has failed:
BROKEN TRUNK - The following jobs failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
This PR needs a
|
seyeong-han
changed the title
seyeong-han
changed the title
Adds Apple Silicon MLX export and runner wiring for Voxtral TTS while keeping codec lowering portable for waveform correctness. Made-with: Cursor
Clarify that the runner exposes streaming flags for MLX builds, while this branch only reports offline MLX performance because codec decoding still falls back to portable CPU. Made-with: Cursor
Made-with: Cursor
Keep the advanced-indexing fix reviewable by isolating gather permutation logic, tightening MLX test availability checks, and updating the Qwen MoE CI cast budget to match the corrected graph. Made-with: Cursor
seyeong-han
force-pushed
the
voxtral-tts-mlx
branch
from
3f98344 to
fca9ecf
Compare
Excluded per PR feedback; parity testing will be handled separately.
The original intro said weights are loaded directly from safetensors, which was ambiguous — at inference the C++ runner loads .pte files, not safetensors. Localize the claim to export.
metascroy
reviewed
May 1, 2026
metascroy
reviewed
May 1, 2026
Move MLX index regression coverage into the consolidated op tests and document why F.unfold needs PyTorch-compatible gather ordering. Co-authored-by: Cursor <cursoragent@cursor.com>
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
Enable Voxtral TTS on the ExecuTorch MLX backend for Apple Silicon.
This PR adds MLX export support for the LM/flow methods and the codec decoder,
make voxtral_tts-mlxand CMake preset wiring, README instructions, a one-shot MLX E2E script, and MLX parity/regression tests. The native MLX codec fix is in the backend advanced-indexing lowering used by the codec'sF.unfold/im2colconv rewrite.The shared runner exposes
--streamingand--speakerfor MLX builds. Offline MLX synthesis is validated here with the native MLX codec artifact; streaming uses the same codec artifact and avoids the old portable CPU codec fallback.--backend mlx --qlinear-codecis still rejected because MLX codec quantization is not yet validated.Benchmark
Apple Silicon MLX benchmark using bf16 + 4w linear + 8w embedding export, with native MLX LM/flow and native MLX codec:
Average generation RTF:
0.811337(0.761774warm-run average). Average process wall:3.82s(3.14swarm-run average). WAV quality check: peak0.42575, clipped samples0. Apple Speech transcribed the benchmark WAV asHello how are you today.Test plan
conda run -n et-mlx python -m pytest -q examples/models/voxtral_tts/test_mlx_parity.py8 passed in 395.04sPYTHONPATH=/Users/younghan conda run -n et-mlx python -m pytest -q backends/mlx/test/test_runtime_ops.py2 passed in 4.82s--backend mlx --dtype bf16 --export-target codec/tmp/voxtral_tts_mlx_native_codec_256/codec_decoder.pte(289.2 MB)/tmp/voxtral_tts_mlx_final/model.pteand/tmp/voxtral_tts_mlx_native_codec_256/codec_decoder.pte0.811337, average process wall3.82sFINAL Hello how are you today