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 supported36 Subscribers
Add a CodeTriage badge to vllm
Help out
- Issues
- [Bug]: --moe-backend flashinfer_b12x roughly doubles peak activation memory vs default, cutting available KV cache ~55%
- [Feature] Add IndexCache support for DeepSeek V4 on Hopper
- [Kernel][MoE] Zero-expert identity: drop redundant zeros_like, use ceil-div grid
- [Bugfix] Guard scaled_fp4_quant against non-contiguous input
- [Bugfix] Gate FlashInferTrtllmNvFp4LinearKernel on sm_10x
- [Bug]: PRIORITY scheduling can silently skip a running request for a full step when the preemption victim was already deferred earlier in the same schedule() call
- [Doc] Document SimpleCPUOffloadConnector in the KV offloading guide
- [Bug]: `all_gather_interleave()` in PaddleOCR-VL/Ernie4.5-VL/GLM-4.1V bypasses vLLM's collective API and crashes with "No backend type associated with device type" on out-of-tree platforms
- [Bug] enable_lora on hybrid-MoE model perturbs outputs of requests WITHOUT lora_request (greedy repetition loop) — fused MoE LoRA implementation
- [Usability] Engine fails to start when available Mamba cache blocks < max_num_seqs (hybrid models + LoRA) — suggest auto-clamping with a warning
- Docs
- Python not yet supported