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 supported29 Subscribers
View all SubscribersAdd a CodeTriage badge to vllm
Help out
- Issues
- [CPU][Zen] Route Int8 MoE inference through zentorch on AMD
- [Kernel][Perf] Add moe_sum_kernel specializations for topk=5-8
- [Performance]: After enabling MTP on the Qwen3.5-27B model, the number of hit blocks for the prefix cache is one less compared to the scenario with MTP disabled. This is the current implementation. Can we optimize this behavior?
- [Bug]: EngineDeadError: RPC call to execute model timed out on CPU when running google/gemma-4-26B-A4B-it with large concurrent decode batch`
- [KV Offload] Reshape the transfer data model: per group specs and offloaded side alignment offset
- [Feature]: [CPU Backend] Support macOS x86 for CPU backend
- [Aiter][ROCm] QKV-split + QK-RMSNorm + RoPE + KV-cache-write fusion
- [Bug]: vllm serve hangs permanently with data_parallel_size > 1 on multi-node Ray cluster (MoE model, --enable-expert-parallel)
- [Bug]: compressed-tensors kv_cache_scheme overrides kv_cache_dtype to fp8 during MTP speculative decoding, causing CUDA crash on SM89 (L4/RTX4090)
- [Performance]: Cold start takes ~5 min for nvidia/Qwen3.6-35B-A3B-NVFP4 on RTX 5090 (sm_120) with vLLM
- Docs
- Python not yet supported