HuggingFace inference operator: robustness cleanups in provider fallback and media-generation polling
- Dominant language
- Scala
- Stars
- 314
- Forks
- 187
- Avg merge
- 1d 21h
- Merged PRs (30d)
- 214
Description
### Task Summary
Two brittle spots in the operator's runtime provider-fallback / polling path.
### Context / checklist
- Replicate/Wavespeed poll handlers mutate the private `resp._content` and `resp.status_code` to synthesize a 502 - brittle; use a proper response object or a small result type instead.
- 400/422 are treated as "retryable", so a genuinely malformed payload is replayed against every provider before giving up (wasteful, and it obscures the real error).
File: `.../huggingFace/codegen/HuggingFaceCodegenBase.scala` (`_call_provider` poll handlers, `_post_with_fallback` retryable set).
### Required test
Per item, as addressed.
### Related
Part of the operator's fallback loop (see the provider-memoization task, #7202).
### Task Type
- [x] Refactor / Cleanup
- [ ] DevOps / Deployment / CI
- [ ] Testing / QA
- [ ] Documentation
- [ ] Performance
- [ ] Other
Contributor guide
Assessment
This issue has not been assessed yet.