Search results

[R3] Long collectives with --enable-return-routed-experts on Kimi K2 EP=DP=TP=32

Status: Open.

#24456 In sgl-project/sglang;

· ByronHsu opened on May 5, 2026

[Cuda Graph] Kimi K2 DPEPTP32 cuda graph 128 uses 20GB memory

Status: Open.

#24453 In sgl-project/sglang;

· ByronHsu opened on May 5, 2026

[Bug] [AMD] Accuracy issue with EAGLE speculative decoding and aiter's unified attention for Qwen3.5 FP8 and MXFP4 models

amd

Status: Open.

#24440 In sgl-project/sglang;

· hubertlu-tw opened on May 5, 2026

[Feature] Port to Metal Performance Shaders for MacOS

Status: Open.

#24428 In sgl-project/sglang;

· shyamalschandra opened on May 5, 2026

[Bug] Speculative decoding incorrectly permits Outlines grammar-constrained requests

Status: Open.

#24413 In sgl-project/sglang;

· YzXiao101 opened on May 5, 2026

[Bug] EAGLE3 Tree Decoding CUDA error: an illegal memory access was encountered

Status: Open.

#24402 In sgl-project/sglang;

· Dogacel opened on May 5, 2026

[Bug] [GPU] Qwen3.5 failed on SWE-benchmark with cublas error

Status: Open.

#24365 In sgl-project/sglang;

· LilyPankratova opened on May 4, 2026

[Bug] Qwen3.6 hybrid Mamba: Mamba pool 9-18 GiB allocation OOM on RTX PRO 6000 Blackwell + WSL2

Status: Open.

#24364 In sgl-project/sglang;

· noMugop opened on May 4, 2026

[Bug] CUDA illegal memory access in flashinfer prefill replay during NGRAM speculative decoding on Blackwell sm_120 (RTX PRO 6000)

Status: Open.

#24361 In sgl-project/sglang;

· mittsh opened on May 4, 2026

Triton RMSNorm kernel underutilizes T4/A10G/A100 SMs for large contexts (NCU data)

Status: Open.

#24353 In sgl-project/sglang;

· Umang-projects opened on May 4, 2026

[Bug] DSV4 MTP metadata code

Status: Open.

#24352 In sgl-project/sglang;

· nvjullin opened on May 4, 2026

[Feature] Update PyTorch to 2.12.0, torchvision to 0.27.0, triton to 3.7.0

Status: Open.

#24351 In sgl-project/sglang;

· johnnynunez opened on May 4, 2026