Download Latest Version v0.5.4 -- transport and resilience fixes source code.zip (358.0 kB) Google Add to Preferred Sources
Home / v0.5.4
Name Modified Size InfoDownloads / Week
Parent folder
README.md 2026-10-01 2.2 kB
v0.5.4 -- transport and resilience fixes source code.tar.gz 2026-10-01 329.7 kB
v0.5.4 -- transport and resilience fixes source code.zip 2026-10-01 358.0 kB
Totals: 3 Items   689.9 kB 0

Reliability and transport fixes

  • Use HTTPX2 transports for OpenAI-compatible providers, Azure, and Bedrock. Preserve Bedrock SigV4 signing and refresh credentials per request.
  • Honor num_retries / max_retries across all completion handlers, including streamed request creation. Default retries are zero; num_retries takes precedence.
  • Map ordinary SDK bad-request errors into litelm's exception hierarchy. Recognize actual context overflow rather than any message mentioning tokens or length.
  • Map streaming transport errors without replaying output. Close active streams on exhaustion, failure, cancellation, or explicit close; preserve primary failures if wrapper cleanup also fails.
  • Improve unknown-provider diagnostics, preserve nested OpenRouter model names, and support keyless registered local endpoints.
  • Update locked urllib3 from 2.7.0 to 2.8.0 to address Dependabot alerts [#45]–#47. This refreshes the project lock; it does not add urllib3 as a base dependency.

Upgrade notes

  • Requires openai[httpx2]>=2.54.0,<3 and httpx2>=2.12.0,<3; the Anthropic extra requires anthropic>=1.5.0. HTTPX remains an OpenAI transitive dependency.
  • Non-empty fallbacks= now raises ValueError instead of being silently ignored. Fallback execution is not implemented; empty/None fallbacks remain accepted.
  • Unknown provider prefixes require an explicit endpoint for real calls. Parsing/introspection and mock responses remain permissive.
  • Output-only max_tokens violations now raise BadRequestError, not ContextWindowExceededError.

Verification

  • 580 local tests passed on Python 3.10 and 3.14; 49 upstream public-contract tests passed.
  • 52 selected live provider/DSPy tests passed. Two context-window tests and the retired gpt-3.5-turbo-instruct test were excluded.
  • Bedrock signing/retry coverage uses mocked transports, not live AWS calls.
  • These results do not certify DSPy 3.4 canonical workflows or additional downstream integrations.

Feedback: OpenRouter usability (#19), fallback request (#20), and Hacker News discussion.

Source: README.md, updated 2026-10-01