Skip to content

feat: align Gradio and MOSS serving documentation - #3669

Merged
LauraGPT merged 2 commits into
mainfrom
codex/gradio-moss-serving-20260908
Sep 7, 2026
Merged

LauraGPT merged 2 commits into
mainfrom
codex/gradio-moss-serving-20260908

Conversation

@LauraGPT

@LauraGPT LauraGPT commented Sep 7, 2026

Copy link
Copy Markdown
Collaborator

Summary

  • Add explicit FunASR, native vLLM and SGLang Omni protocol profiles to the existing private Gradio HTTP client, preserving the FunASR/SenseVoice/verbose_json default.
  • Add MOSS to FunASR choices, use diarized_json for the pinned native vLLM MOSS path, and preserve the full SGLang served-model ID behind its compact display label. An optional --model sends the operator-configured served ID exactly.
  • Rebuild English/Chinese Gradio guides and consolidate the four-language HTTP README entry points around an isolated Python 3.12 / Gradio 6.26.0 client environment. All six prepared-source recipes pin the tested runtime commit d91d961.
  • Distinguish backend preparation, metadata checks, actual smoke inference, response formats/duration and recording-local speaker labels. Document editable server-side targets, missing authentication, buffered uploads, temporary storage and raw diagnostic risks.
  • Preserve existing headings and links. No new hosted Gradio endpoint, backend inference changes, external VAD, speaker-identity claims or PyPI release.

Verification

  • Runtime RED reproduced 7 failures / 11 passes. Final 21 tests exercise real Gradio components, three live Gradio upload/event paths, actual multipart HTTP requests, exact model/format values and unchanged speaker/segment responses. Missing metadata does not gate transcription.
  • Real CLI browser checks cover FunASR MOSS selection, vLLM, SGLang Omni and a custom vLLM model override at 320/390/1440. All 12 final journeys and screenshots inspected; compact display labels and responsive control widths address two screenshot-discovered layout defects.
  • Documentation RED reproduced 6 failures / 39 passes; final 45 focused contracts pass. Full product/docs regression, generated-site validation/export, bilingual HTTP guide navigation and every rendered code-block clipboard value checked. Exact signed-head runtime/docs tests and byte-identical site rebuild are repeated before push.
  • Both commits have verified SSH signatures and DCO. Candidate source, git bundles and test/publication evidence are backed up.

These are controlled HTTP/UI fixtures, not MOSS acoustic inference or native GPU-server certification. The client remains a trusted private operator demo with no backend Authorization or target allowlist. Browser upload success does not prove microphone permissions on arbitrary origins, model accuracy, identity recognition, production capacity or public-network isolation. Gradio/pandas dependency deprecation warnings are retained in the test record. No unresolved issue is closed and no branch-protection or approval state is altered.

Signed-off-by: zhifu gao <zhifu.gzf@alibaba-inc.com>
Signed-off-by: zhifu gao <zhifu.gzf@alibaba-inc.com>
@LauraGPT
LauraGPT merged commit ef11a37 into main Sep 7, 2026
5 checks passed
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