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
- [RFC]: Native span pooling for contextual chunk embeddings
- [Qwen4Exp] Gather PLE rows on the host from the checkpoint files (`host_file_gather`)
- [Bugfix] Llama3 JSON tool parser: flush buffered '{...}' as content when it isn't a tool call
- [Performance][Bug]: Tiered Offloading
- vllm serve segfaults deterministically right after weight loading (single L4 GPU, V1 engine) — only avoided by VLLM_TRACE_FUNCTION=1
- [Performance]: MoE decode ~15% slower on SM12x since #56876 (DeepGEMM contiguous-layout alignment 128 instead of 64)
- [Bug][DSA] Complete sparse top-k output after DeepSelect detects a NaN
- [Bug]: GLM-5.x sparse indexer: the replicated key norm is autotuned per rank, so TP ranks can select different candidates and greedy completions change from launch to launch
- [Bug]: MFU/MBU counts speculative-decoding steps as prefill (`is_prefill = num_tokens > 1`)
- [Frontend][Spec Decode] Populate metrics.speculative_decoding on /v1/responses
- Docs
- Python not yet supported