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 supported51 Subscribers
View all SubscribersAdd a CodeTriage badge to vllm
Help out
- Issues
- [BugFix] Pick the DP world-group port at bind time via the coordination store
- [Bugfix] Forward SwiGLU clamp/alpha/beta in compressed-tensors W4A4 MXFP4 MoE
- [Feature][Model] Support stable-window-aware KV reuse for Qwen3-ASR realtime
- [Model] Kimi-K3: all requests degenerate to a repeated token after long-context prefill (NaN logits; packed KDA prefill suspected)
- DeepSeek-V4-Flash-0731: KV cache holds only 150K tokens in 7.7 GiB (56 bytes/token), max_model_len capped at ~121344 on H20 TP=2
- [PD][NixlPush][Bugfix] Notify D of NIXL writes that failed to post so it stops waiting
- [Bug]: CUDA graph memory profiler charges an mnbt-linear profiling transient as "graph memory" → KV cache pool −18–24% on hybrid GDN models since v0.26.0 (the real FULL graphs cost ~24 MiB each)
- [ROCm][Kimi-K3] Enable opt-in K3 SiTUv2 A8W4 routed MoE
- [Doc] Use wheel index for CUDA variant installation
- [Bug][Spec Decode] DSpark speculative decoding broken on H200 nightly
- Docs
- Python not yet supported