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
- [Roadmap] vLLM Roadmap Q4 2026
- [Bug]: CPU MXFP4 MoE (gpt-oss) broken on AMD EPYC after AMX gate
- [Performance]: Triton decode with 64-token KV tiles beats FLASH_ATTN at kernel level on H100, but only +2.6% end to end
- [Performance]: LoRA shrink launches empty split-K programs with the default config (split_k=64 > ceil(K/block_k))
- [Bug]: CPU KV offload crashes with Xid 31 on NVIDIA vGPU — cuMemcpyBatchAsync faults even with a single valid descriptor
- [Bug]: TP ranks can route the same token to different MoE experts because Inductor autotunes the fused add+RMSNorm reduction per rank
- [Bug][ROCm]: `ENV HSA_ENABLE_IPC_MODE_LEGACY=1` in the ROCm Dockerfiles makes `hipIpcGetMemHandle` fail on in-tree amdgpu kernels
- [Bug]: Rust frontend can't start with the official Qwen3-ASR checkpoints (no tokenizer.json, nested vocab_size)
- [Spec Decode][DFlash2] Every public 27B draft pairing we could construct is a net regression vs full-vocab MTP — pairing validation + last-stage memory asks
- [Spec Decode] RecoverSSM: commit all GDN KV cache groups with one set of launches
- Docs
- Python not yet supported