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
- [Model] MuseSpark: Ampere (sm80) + AutoRound/GPTQ INT4 experts
- [Beam] Enable CuTeDSL NVFP4 MegaMoE
- Nemotron-3.5 tool calls lose trailing/edge newlines under the wrapped qwen3_coder format (_trim_wrapping_newlines) — needs a compact converter like #58019
- [Performance]: NIXL KV transfer silently falls back to UCX TCP emulation (~0.56 GB/s) on same-host GPUs without peer-to-peer access and no RDMA NIC
- [Benchmark][Attention] Support int8_per_token_head KV cache in attention_benchmarks
- [Perf][Attention] 16-byte-align per-token-head KV slots so Triton vectorizes loads
- [Bug]: Qwen3.8-Flash-Next (qwen4_exp) deadlocks on RTX 6000D (SM120) — hangs in JIT warmup and at first real request (GDN/hyperconnection kernels); Qwen3.8-27B works fine on the same box
- [Bugfix][Parser] Strip leading whitespace from first content delta af…
- [Spec Decode] Compose grammar-constrained sampling with EAGLE speculative decoding
- [Bug]: moe_backend="cutlass" with online FP8 raises a reason-free "vLLM CUTLASS FP8 MoE backend is disabled for this configuration." — every sibling backend reports the actual rejection reason via _make_log_unsupported
- Docs
- Python not yet supported