Search results

[Bug] : MiMoV2 (FP8 over MXFP4) fails to load: KeyError: 'model.layers.1.mlp.experts.w2_weight_scale' on v0.5.13-rocm720-mi30x

Status: Open.

#28673 In sgl-project/sglang;

· ppatel-eng opened on Jun 18, 2026

[RFC] Disaggregated request preprocessing

Status: Open.

#28669 In sgl-project/sglang;

· sagearc opened on Jun 18, 2026

Deprecate and delete --moe-runner-backend cutlass, dense FP4 GEMM, FP4 quantization (in sgl-kernel)

RFC

Status: Open.

#28663 In sgl-project/sglang;

· b8zhong opened on Jun 18, 2026

[Bug] GLM-5.2 crash with pipeline-parallel

Status: Open.

#28659 In sgl-project/sglang;

· xhejtman opened on Jun 18, 2026

[Bug] HiCacheFile file backend: single flat directory causes ENOSPC at scale (millions of cached pages)

Status: Open.

#28653 In sgl-project/sglang;

· hpcnt-albert opened on Jun 18, 2026

[Docs] Production metrics page out dated

Status: Open.

#28638 In sgl-project/sglang;

· hzxuzhonghu opened on Jun 18, 2026

[Feature] Cross-backend KV transfer QoS — plumb request priority to P→D transfer and HiCache

Status: Open.

#28631 In sgl-project/sglang;

· catyans opened on Jun 18, 2026

[Bug] main-cann8.5.0-910b image fails to import torch_npu due to torch / torch_npu ABI mismatch

Status: Open.

#28628 In sgl-project/sglang;

· ChefWu551 opened on Jun 18, 2026

[RFC] Add SM89/L20 support for DeepSeek-V4-Flash-FP8

Status: Open.

#28618 In sgl-project/sglang;

· xesdiny opened on Jun 18, 2026

[Bug] DP attention crashes at KV store for qwen2.py models (Qwen2.5): KV head count split by tp_size vs full DP KV pool

Status: Open.

#28615 In sgl-project/sglang;

· whycoming opened on Jun 18, 2026

[QST] About DeepSeek's DeepEP v2

Status: Open.

#28610 In sgl-project/sglang;

· Rachmanino opened on Jun 18, 2026

[Bug] [Disaggregation] ZMQ sockets are not cleaned up after prefill node failure, causing port conflicts with NCCL on prefill restart

Status: Open.

#28596 In sgl-project/sglang;

· Xavier1994 opened on Jun 18, 2026