Skip to content

docs: diagnose actual AutoModel GPU placement - #3740

Merged
LauraGPT merged 1 commit into
mainfrom
codex/gpu-device-troubleshooting-20260930
Sep 30, 2026
Merged

LauraGPT merged 1 commit into
mainfrom
codex/gpu-device-troubleshooting-20260930

Conversation

@LauraGPT

@LauraGPT LauraGPT commented Sep 30, 2026 •

Copy link
Copy Markdown
Collaborator

Summary

Related to #3738. The reporter has since confirmed the environment remedy and closed the support issue; a runtime fallback warning remains a separate follow-up.

  • Replace the troubleshooting FAQ's stale generic PyTorch upgrade / FunASR 1.3.26 commands with links to the maintained installation guides and interpreter-specific checks.
  • Add matching English and Chinese instructions for distinguishing PyTorch CUDA build support, runtime CUDA availability, requested/resolved device and actual loaded tensor placement.
  • Inspect ASR, VAD, punctuation and speaker modules separately, including buffers, disabled modules, empty modules and non-PyTorch objects.
  • Explain why whole-card utilization/memory, an installed CUDA toolkit or a working llama.cpp backend cannot alone prove that this Python process uses CUDA.

Only docs/troubleshooting.md and docs/troubleshooting_zh.md change. No runtime behavior, dependencies, model defaults, deployment settings or generated site files change.

Type of change

  • Bug fix
  • Documentation
  • Example or demo
  • Runtime or deployment
  • Benchmark or evaluation
  • Model/training change

Validation

Exact signed head: 03e18dccb1c9228d49e2c59e6922546d34b22d8a, based on 12e417f4fa4490296ea2d81c5352ea2de8e4cac2.

  • Executed both published Markdown diagnostic snippets on tiny CPU PyTorch objects with CUDA hidden and model/hub access disabled. Covered parameters, buffer-only modules, disabled modules, non-PyTorch objects, empty modules and absent package metadata; a substituted version lookup also exercised the metadata-present branch. Both locale snippets are identical.
  • Native product-site suite: 519 passed, no skips.
  • The workflow's model-free documentation-contract selection: 315 passed, 3 deselected, no skips. The initial run had one Node-dependent skip; rerunning with the already-installed Node 22.23.2 on PATH exercised it successfully. The three exclusions match the workflow selection; this is not a full FunASR test-suite claim.
  • Two independent native site builds each generated 117 product pages and validated 206 total pages. Recursive comparison found identical output.
  • git diff --check: passed. A separate Codex reviewer checked the two exact file hashes, source fallback semantics and translation parity; no actionable P1/P2 findings. This is not human maintainer approval.

The remote observation wrapper timed out after receiving some test output. Terminal receipts were independently re-read for every native test/build/validation/comparison process; no completed work was restarted merely because observation expired.

No ASR model was downloaded or loaded, no real transcription or GPU inference was run, and no service was deployed. The issue's Windows/RTX 3080 behavior is not reproduced or claimed fixed. The diagnostic helper is documentation, not a new runtime API or a newly committed unit-test module. This contribution was generated and reviewed with Codex.

  • python -m compileall funasr examples tests (not run; documentation-only change)
  • Docs or links checked
  • Runtime/deployment command tested (no ASR inference or deployment)

User impact

New Python SDK users can inspect effective placement before reinstalling packages or treating a low utilization graph as a GPU failure. Both language versions direct installation questions to the maintained, platform-aware guide.

Notes for reviewers

Review the device-versus-utilization distinction and the documented CPU fallback against AutoModel.build_model. The original user issue still needs reporter output. The documentation catalogue already renders these two source files into the bilingual product site; no manual deployment or generated-file update is part of this PR.

Merged Outcome

Merged as e9d002b54d0ba0592797d13d6ffd742c7d41ca75. Exact-head Product site run 36682086015 completed successfully, including build-and-validate, legacy-docs-links and browser. The merge tree and both document bytes match the reviewed/tested head; the GitHub merge signature is verified. This used the existing administrator exemption through the normal SHA-guarded merge API, without changing protection rules or creating an approval.

The reporter of #3738 independently reported that selecting a CUDA-enabled PyTorch build resolved their usage problem and closed the issue. This is reporter confirmation, not our hardware reproduction. Their suggestion for a visible fallback warning remains a separate runtime follow-up. No production website deployment is claimed.

Post-merge checks on e9d002b54d0ba0592797d13d6ffd742c7d41ca75: Product site 36684747811 succeeded with all three jobs, including browser; Update API Documentation 36684747740 also succeeded. These are exact-merge CI results, not a claim that the production website was deployed.

Signed-off-by: LauraGPT <18321252+LauraGPT@users.noreply.github.com>
@LauraGPT
LauraGPT merged commit e9d002b into main Sep 30, 2026
3 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