Issue evidence · AI API · BerriAI/litellm

litellm_settings.request_timeout never fires when upstream is silent from the first byte

GitHub issue: https://github.com/BerriAI/litellm/issues/38358

Reproduced

With request_timeout set, a request to an upstream that accepts the connection but never writes a byte hangs instead of failing at the configured budget.

What Badgr ran

Ran LiteLLM 1.104.2 as a proxy with litellm_settings.request_timeout: 8 and an Anthropic-style deployment pointed at a local socket that accepts TCP and never replies; sent POST /v1/messages with stream true and with stream false, client limit 45 s.

What came back

  • Streaming request: no response at all within 45 s (client timeout).
  • Non-streaming request: no response within 45 s either.

The issue reports 1.97.0 and 1.98.0; this is a later release (1.104.2) and still hangs. Whether it eventually returns after more than 45 s was not tested.

Checked 2026-10-15 in Badgr’s local development environment.

← All issue evidence