flush() awaited join() on whatever queue the worker held, even one bound to
an event loop that has since closed. Its unfinished counter is never
decremented on the new loop, so the first flush() after a loop change hung
until pytest-timeout killed it and every later one raised "is bound to a
different event loop" from the queue's Event. The CircleCI unit job has
been red on every branch since the first tests that flush without
enqueueing landed, and an SDK script that flushes from a second
asyncio.run() hangs the same way.
flush() now goes through start() first, which carries the tasks stranded
on the previous loop onto the current one and guarantees a worker there to
drain them, the same loop-change handling every other entry point already
had.
Register s3_access_key_id, s3_secret_access_key and s3_encryption_key_id as
LiteLLM-owned batch params so they are no longer forwarded to Bedrock as
additionalModelRequestFields (which 400s ordinary chat on a batch-configured
deployment), keep them on CredentialLiteLLMParams so the batch/file paths
still receive them, and redact the S3 credential key names in debug logs.
Resolves LIT-8290
Co-authored-by: yucheng <yucheng@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
The router coverage gate wants every router.py function reached by name from
a router test. The two per-endpoint attempt functions were only reached through
their callers, so each now has a direct test proving the per-request controls
carrier never reaches the provider call and every hop's stream comes back wrapped.
Combines the Redis cluster bridge changes with the Azure Blob variant and
keeps account URL query parameters when building the container URL
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
The /v1/responses and /v1/messages streaming wrappers only ever wrapped the
primary's stream, so a hop reached through the regular fallback chain had no
mid-stream handler: its failure re-raised, or the outer wrapper retried the
same entry with a fresh attempted set and never reached the rest of the list.
Every attempt of the chain now runs through a per-endpoint attempt function
that wraps its own stream, mirroring chat completions, and the per-request
fallback and retry overrides ride a frozen carrier so each hop's re-entry
still sees them after the retry layer pops them.
The stream chunk builder started its per-chunk accumulators at 0 and adopted only nonzero counts, then fell back to litellm's tokenizer whenever the accumulated value was falsy, so a provider that reported an explicit 0 for prompt or completion tokens was billed the estimate instead. The accumulators now start at None, a usage chunk that reports a count marks it reported (a later chunk's 0 never replaces a reported nonzero), and the estimate only runs when no chunk reported the count. The Anthropic message_start cursor reset now yields None so the estimate still covers a cancelled stream, and Ollama chat streaming only attaches usage on the done chunk when both counts are present instead of inventing 0/0 on every chunk
