Repository navigation
Conversation
Integrate #3768 with the preserved compatibility correction, object validation and durable CI. Co-authored-by: Li Junhui <lijunhui.me@foxmail.com> Signed-off-by: LauraGPT <zhifu.gzf@alibaba-inc.com>
Integrate gotorion's #3768 with compatible optional access, JSON object validation and durable CI. 185 focused CPU tests and native synthetic FP32/BF16 vLLM CUDA execution passed; not pretrained ASR quality or a full-suite claim. Signed-off-by: LauraGPT <zhifu.gzf@alibaba-inc.com>
|
Thanks @gotorion. Your realtime The integration preserves callers without the optional Namespace attribute, rejects non-object JSON before loading models, and keeps the tests in both NumPy CI matrices. Local verification: 185 focused CPU cases, zero skips; actual L20Z/native vLLM execution in FP32 and BF16 using saved, hashed random weights. The initial undersized GPU fixture failure is retained and corrected, not hidden. This validates CLI/loader boundaries and native config/CUDA generation, not pretrained audio accuracy, every custom architecture or the full repository suite. New-head and final-main CI succeeded. Actual downloaded evidence for each of the four NumPy matrix runs contains 358 passing cases, including all 90 realtime service and 13 JSON validation cases, and the exact source-commit marker. No release, PyPI upload, protection change or administrator bypass is included. |
Summary
Type of change
Validation
python -m compileall funasr examples testsUser impact
People who has their own vLLM backend may use different model implementation.
Notes for reviewers