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
- [Feature] Batch-invariant support for speculative decoding
- [Bug]: Marlin MoE output depends on equivalent within-expert token ordering
- [Metrics] Report shared-prefix tokens lost to a missing sparse-retention checkpoint
- [Bug]: Kimi-K3 CUDA graph capture silently corrupts output at batch=1; three distinct failure modes across cudagraph modes.
- [Bugfix][Kernel] Canonicalize Marlin MoE token order
- [Bug] XPU qnorm/rope kernel: overlapping stores in one program corrupt the KV cache nondeterministically
- [Feature]: Support NVFP4 DeepSeek-V4-Flash-0731 with FP4 KV cache + DSpark speculative decoding on SM121 (DGX Spark)
- [Bug]: 0.27.1 crashes in custom_all_reduce (illegal memory access) during CUDA graph capture on Hopper TP=4; identical config works on 0.26.0
- [Core][DP] Prefix-affinity tie-break in the internal DP load balancer
- Streaming input: per-chunk arrival cost is O(cumulative prompt), making long multimodal sessions quadratic
- Docs
- Python not yet supported