You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Fixed TOCTOU race condition in parallel_request_limiter.py. TPM was only updated after request completion, so concurrent requests all read current_tpm=0 and bypassed the limit. Fix: reserve estimated tokens (max_tokens) at pre-call time, reconcile with actual usage on completion, release on failure.
Files changed
litellm/proxy/hooks/parallel_request_limiter.py, 1 new test file (4 tests)
Status
⏳ OPEN — awaiting maintainer review
Next steps
Maintainers may request changes. Watch for CI results and review comments.
Proposed Jain's fairness index as the metric for their dynamic fair-share feature. Linked to standalone implementation as reference. Offered to write a follow-up PR if maintainers are interested.
Status
⏳ POSTED — awaiting maintainer response
Next steps
If they express interest, write a PR adding fairness index to their Prometheus callback system.