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
Add a CodeTriage badge to vllm
Help out
- Issues
- First boot of a new `(max-num-batched-tokens, max-num-seqs)` shape is granted a ~5% smaller KV cache than subsequent identical boots
- [Feature] Support downloading models from AtomGit
- [Fix] Raise clear error for unknown source fps in Qwen3-VL video backend
- [Bugfix][Multimodal] Guard GLMGA video sampling against zero source fps
- [Bug][Feature][KV-offloading]: shared-mmap CPU region assumes all ranks on one host — multi-node engines hang at init; node-local disk tier reference implementation
- [Bug][ROCm] GLM-5.3-Flash kpool indexer overwrites its own KV cache: backend advertises a kernel block size the kernels do not use
- [Bug]: Speculative decoding (mtp and dflash) silently disables prefix-cache hits for hybrid GDN models on nightly; worked on v0.24.0
- [Bugfix][GDN] Run Qwen GDN cores as eager breaks in breakable CUDA graph
- Responses API input_item_parsing only backfills id when the key is absent, not when it is explicitly null
- [Bugfix] MoE MBU: estimate distinct activated experts instead of perfect-load-balancing upper bound
- Docs
- Python not yet supported