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 supported37 Subscribers
View all SubscribersAdd a CodeTriage badge to vllm
Help out
- Issues
- [Kernel][Inkling] Fuse shared-expert partial addition into the Lamport collective
- [Bug]: DiffusionGemma: runtime CUDA OOM under concurrent decode (fp32 canvas×vocab temporaries)
- [Bug]: int8_per_token_head KV + prefix caching corrupts output when the KV pool is pinned at 100% (Gemma-4 hybrid, Triton)
- [Bug] DFlash on SM121 (GB10 / DGX Spark): attention autoselect picks FLASH_ATTN for non-causal draft attention and device-asserts in _vllm_fa2_C.varlen_fwd
- [Bug]: TurboQuant hybrid model crashes at determine_available_memory with 'Unknown cache dtype: auto' on v0.25.0+
- [Bug] DFlashProposer assumes a single KV-cache group; breaks with hybrid sliding+full-attention drafters
- [Bugfix] benchmark_moe: do not abort tuning when a candidate config fails Triton compilation
- [CI] Solidify speculative decoding E2E coverage
- [Bugfix][Spec Decode] Support hybrid-attention DFlash drafters spanning multiple KV cache groups
- [Bug]: With qwen3.5-35b-a3b, the performance is relatively poor both when using dflash and when not using it, but the accepted length of dflash is around 5–6.
- Docs
- Python not yet supported