sglang
https://github.com/sgl-project/sglang
Python
SGLang is a high-performance serving framework for large language models and multimodal models.
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 supported1 Subscribers
Add a CodeTriage badge to sglang
Help out
- Issues
- [Perf] [NVIDIA] deepseek_v4: Token-shard SM90 sparse-MLA prefill under attention TP
- [Perf] [XPU] minimax-h3: Use bit-exact fused XPU kernels in the DiT
- [Bug] More model loaders drop checkpoint FP8 KV-cache scales (Qwen1.5/2-MoE, Qwen2.5-VL, Qwen3-Next, AFMoE, Gemma 3n, Gemma 4 Unified, Bailing-V2.5, DBRX, HRM-Text, Nanbeige, ...)
- [Fix] models: map checkpoint KV-cache scale names in more model loaders
- [Bug] Gemma 4 and Gemma 3n KV-shared layers read the FP8 KV cache with scale 1.0 instead of the writing layer's scales
- [Fix] gemma: KV-shared layers use the KV-cache scales of the layer whose cache they read
- [Fix] spec: load FP8 KV-cache scales in the DFlash and LLaMA EAGLE-3 draft loaders
- [Bug] TorchNativeLlamaForCausalLM / TorchNativePhi3ForCausalLM load no checkpoint tensors and no KV-cache scales
- [Fix] torch_native_llama: load checkpoint tensors and KV-cache scales
- [Bug] NVFP4 W4A4 linear keeps the largest weight global scale of fused q/k/v and gate/up layers without a check
- Docs
- Python not yet supported