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
Add a CodeTriage badge to vllm
Help out
- Issues
- [Bug]: dtype="float32" crashes unquantized MoE models at startup on SM89 — Triton fused-MoE picks a 128 KiB shared-memory kernel (Required 131072 > limit 101376)
- [ROCm][Distributed] Enable custom all-reduce on RDNA4 over PCIe peer-to-peer
- [Bug]: Olmo-Hybrid-7B fails to load with default settings — GDN fused conv1d weight loader indexes `dims[loaded_shard_id]` while `loaded_shard_id` defaults to `None`
- [Bug]: dtype="float32" + enable_lora crashes LFM2 startup — lora_expand Triton op asserts `weight.dtype in [torch.float16, torch.bfloat16]`
- [Bug]: quantization="mxfp4" on CUDA reports "No MXFP4 MoE backend supports the deployment configuration" while listing only AMD backends (AITER / amd-quark)
- [ROCm][Perf][GLM-5.3-Flash] Mono decode layer for GLM-5.3-Flash on MI355X
- [Bug]: NixlConnector P/D: first handshake per remote engine stalls all DP decode ranks (THP compaction in `_build_fa_remote`); late reads after lease expiry return HTTP 200
- [DeepEPv2] Add VLLM_DEEPEP_V2_NUM_SMS to set the dispatch/combine SM count
- [Bugfix][MoE] Fix online-FP8 NaN from skipped padding slots in TritonExperts
- [Bug]: `lightning_attention` prefill applies one decay step too few to the stored state (tokens after position 256 wrong, and every token when continuing from a non-zero `kv_history`)
- Docs
- Python not yet supported