vllm
https://github.com/vllm-project/vllm
Python
A high-throughput and memory-efficient inference and serving engine for LLMs
Triage Issues!
When you volunteer to triage issues, you'll receive an email each day with a link to an open issue that needs help in this project. You'll also receive instructions on how to triage issues.
Triage Docs!
Receive a documented method or class from your favorite GitHub repos in your inbox every day. If you're really pro, receive undocumented methods or classes and supercharge your commit history.
Python not yet supported62 Subscribers
View all SubscribersAdd a CodeTriage badge to vllm
Help out
- Issues
- [Bugfix][ToolParsers] Close unterminated tool-call arguments on truncated streams
- [KV Offload] Preserve packed layouts for layer-wise KV cache
- [WIP][KV Connector] UMBP layer-wise KV loading in stages
- [Bug]: FlashMLA sparse decode allocates the split-KV workspace again on SM90 (regression of #53413)
- [Bug]: Gemma 4 26B AutoRound MoE scales fail loading with 11-vs-12 shape mismatch
- Rocm/ci amd test areas parity
- [Bug]: Sparse-indexer DCP top-k merge allocates (1 + 2 * dcp) * 8 * T * K bytes per step that the profile run never sees, OOM on the first long prefill
- [Bug]: CUDA graph memory estimate (#53306) reports 2.16 GiB against 1.19 GiB actually captured on GLM-5.3 TP4/DCP4: the profiling pass measures things the real capture does not, and the FULL extrapolation uses the second-largest graph as the per-graph cost
- [Bugfix][KV Connector] Stop P/D transfers short of the last prompt token
- [NIXL] Share in-flight prefix loads across requests
- Docs
- Python not yet supported