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
- [PD] Fix prefill warmup hang: advance a skip-insert chunked request's prefix
- [Bug] Mooncake DFlash draft KV transfer fails with asymmetric P/D TP on GLM-5.3-Flash
- [Doc] MiMo-V2.6: minimal B300 commands, verify H200, add B200/GB300
- Npu/kv checksum
- [Bug] DeepSeek-V4.1-Flash + DSPARK: unbounded SparsePrefillWorkspace allocation OOM-crashes the whole TP group
- [GLM-5.3] Capture the KDA extend in breakable prefill CUDA graphs
- [Perf] Padded decode CUDA-graph slots cost more with longer contexts on the triton backend (bs 5 on the bs-8 graph: 27.7 -> 35.9 ms/token at 30K)
- [Bug][AMD] aiter unified attention: Gemma 2/3 batched generations run away (Gemma 3 4B GSM8K 0.57 vs 0.83 on triton)
- [Bug] `--json-model-override-args '{"rope_scaling": ...}'` drops `rope_theta` for models without a per-model fallback: Llama-3.2-3B-Instruct GSM8K 161 → 106 (first 200)
- [DO NOT MERGE][CI] Move 1-gpu-large tests that pass on the 5090 to 1-gpu-small
- Docs
- Python not yet supported