Skip to content

[Feat][F3] Apply torch.compile to audio encoder of Fun-ASR.#976

Draft
PoTaTo-Mika wants to merge 21 commits into
sgl-project:mainfrom
PoTaTo-Mika:f-pr3
Draft

[Feat][F3] Apply torch.compile to audio encoder of Fun-ASR.#976
PoTaTo-Mika wants to merge 21 commits into
sgl-project:mainfrom
PoTaTo-Mika:f-pr3

Conversation

@PoTaTo-Mika

Copy link
Copy Markdown

Motivation

Speed up Fun-ASR through torch.compile method.

Modifications

stages.py : build a new function that compiles the SANM encoder part.

Related Issues

Addresses the F-PR3 item #924, following the #898

Accuracy Test

To be done.

Benchmark & Profiling

To be done.

Checklist

  • Format your code according with pre-commit.
  • Add unit tests.
  • Update documentation / docstrings / example tutorials as needed.
  • Provide throughput / latency benchmark results and accuracy evaluation results as needed.
  • For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.

CI

CI runs on self-hosted GPU runners and requires a maintainer to add the
run-ci label. Once labeled, every subsequent push re-triggers CI as
long as the label remains. Use /tag-and-rerun-ci higgs or
/tag-and-rerun-ci moss to select a TTS CI model. Draft PRs are skipped even
if labeled.

PoTaTo-Mika and others added 21 commits June 25, 2026 17:42
- Add # Author: attribution block to all Fun-ASR model and test files
  (sglang_omni/models/fun_asr/*, tests/unit_test/fun_asr/*).
- Prefix the Fun-ASR greedy-decoding comment in benchmarks/tasks/tts.py
  with # note (PoTaTo-Mika): as required by the review convention.
- Remove the redundant section-divider comment block above
  AsrBackendConfig in benchmarks/tasks/tts.py.

Co-Authored-By: Claude <noreply@anthropic.com>
Merges 13 upstream commits (c631dc9..966ddcf) covering:
- H20 -> H100 CI migration for Qwen3-Omni/Qwen3-ASR/TTS
- Lazy TensorRef path for large inter-stage tensors
- T1 shared pipeline-state base + TTS migrations
- Moss streaming chunk coalescing + redundant deepstack clone removal
- Higgs RL rollout logprob/action-mask kernels
- Ming-Omni streaming TTS overlap bugfix
- Qwen3-Omni talker_ar default max_running_requests 16->32

Auto-merge resolved tests/README.md and tests/test_model/test_qwen3_asr_ci.py
cleanly: our fun_asr additions and upstream's H100 calibration/fixture
description updates sit in non-overlapping regions.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Bring in PR sgl-project#914 (Support MOSS-Transcribe-Diarize model). Conflicts in
tests/README.md resolved by keeping both the fun_asr/ and
moss_transcribe_diarize/ directory entries.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
@PoTaTo-Mika PoTaTo-Mika requested a review from shuaills as a code owner July 6, 2026 05:31
@PoTaTo-Mika PoTaTo-Mika marked this pull request as draft July 6, 2026 05:31
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant