Issues

• [Bug] disaggregation: Qwen EAGLE PD decode can produce corrupted output after retraction

  sgl-project/sglang#48356 · Filter by author zeyuan-lcl opened on Nov 4, 2026
  1

• [Feature]

  sgl-project/sglang#48348 · Filter by author iridiumine opened on Nov 4, 2026

• [Feature] dp-attention: the per-step H2D transfer of global_num_tokens_gpu is overwritten before anything reads it

  sgl-project/sglang#48346 · Filter by author michael7193 opened on Nov 4, 2026

• [RFC] mem_cache: Cross-branch KV page splicing and cursor fork/rewind for hybrid models (Qwen3.5 / GDN)

  sgl-project/sglang#48344 · Filter by author shatianming5 opened on Nov 4, 2026

• [Bug] GLM-5.3-Flash + glm47 tool grammar: enable_thinking=false bans </think>, so generation always runs to max_tokens

  sgl-project/sglang#48336 · Filter by author isaac-tee opened on Nov 4, 2026

• [CI] diffusion: perf checks depend on which runner a job lands on, and in-job retries can't fix a slow host

  sgl-project/sglang#48311 · Filter by author leejnau opened on Nov 4, 2026
  1

• [Feature][RFC] Preserve measured cache-write usage across PD handoff

  sgl-project/sglang#48300 · Filter by author LeslieWylie opened on Nov 4, 2026

• [Bug] DSA models crash on first prefill with --cpu-offload-gb: breakable prefill CUDA graph replays the DSA indexer eager break with CPU-offloaded weights (mat2 is on cpu)

  sgl-project/sglang#48260 · Filter by author janbernloehr opened on Nov 3, 2026

• [RFC] Engine-free render service for sglang-processor over gRPC (Lower/Respond), including Messages and Responses

  sgl-project/sglang#48232 · Filter by author ahg-g opened on Nov 3, 2026

• [Bug] PD + Mooncake: different DSA backends on prefill and decode silently corrupt the decode KV ("Requested address ... not found")

  sgl-project/sglang#48230 · Filter by author mattteochen opened on Nov 3, 2026

• Add pinned model-file checksum verification to SGLang Diffusion loading

  sgl-project/sglang#48225 · Filter by author spencer-ritual opened on Nov 3, 2026
  1

• [Feature] diffusion: TP=2 support for GGUF with dynamic LoRA

  sgl-project/sglang#48215 · Filter by author ruiling-smartbear opened on Nov 3, 2026

• [CI] PD: Mooncake 'Memory region not registered by any active device(s)' fails a whole class on the 8-gpu-h20 runners

  sgl-project/sglang#48206 · Filter by author leejnau opened on Nov 3, 2026

• [CI] diffusion: Wan VAE decode time spikes on the 1-gpu-5090 sm* runners

  sgl-project/sglang#48196 · Filter by author leejnau opened on Nov 3, 2026
  3

• [Bug] --tool-call-parser auto finds no parser on Qwen2.5-0.5B, Qwen3 2507 and Ling-mini-2.0, so tool calls come back as text

  sgl-project/sglang#48172 · Filter by author hivaze opened on Nov 3, 2026
  1

• [Bug] fused_marlin_moe is not run-to-run deterministic on sm_90+ (atomic adds, moe_align_block_size token order); --enable-deterministic-inference does not cover it

  sgl-project/sglang#48150 · Filter by author symbai-lab opened on Nov 3, 2026

• [Bug] Mooncake HiCache backup fails for Qwen3.8-Flash-Next-FP8

  sgl-project/sglang#48145 · Filter by author zeyuan-lcl opened on Nov 3, 2026
  2

• [RFC] KV Cache Optimization for Cosmos3-Nano-Sim-Bimanual

  sgl-project/sglang#48144 · Filter by author xupinjie opened on Nov 3, 2026

• [Bug] RadixArk/Muse-Glimmer-NVFP4 fails at load since #45632 (MXFP8 lm_head), and on Hopper also at DeepGEMM warmup

  sgl-project/sglang#48127 · Filter by author Smallfu666 opened on Nov 3, 2026

• [Bug] ModelOpt NVFP4 MoE crashes at warmup with AttributeError 'g1_scale_c' under the default --moe-runner-backend auto on Blackwell

  sgl-project/sglang#48108 · Filter by author aniskumar-nv opened on Nov 3, 2026
  1

• [Feature] Model registry: eager import of all model modules hides failures (one bad dependency → 0 models registered) and adds ~2 s startup

  sgl-project/sglang#48080 · Filter by author SanjayReddy91 opened on Nov 3, 2026

• [Bug] expert_distribution_recorder records all-zero counts under flashinfer_trtllm (BYPASSED TopK skips the recorder hook)

  sgl-project/sglang#47932 · Filter by author jmeadlock opened on Nov 2, 2026

• [Bug] [mamba/hybrid] MambaPool.copy_from gathers all source temporal states into one temporary (n x state_size) - CUDA OOM under parallel admission bursts

  sgl-project/sglang#47916 · Filter by author jeck00119 opened on Nov 2, 2026
  2

• [Bug] Qwen3.5 GatedDeltaNet: finalize_fused_in_proj builds the fused in_proj weight on CPU under --cpu-offload-gb (mat2 is on cpu)

  sgl-project/sglang#47896 · Filter by author janbernloehr opened on Nov 2, 2026

• [Bug] SMG PowerOfTwoPolicy does not track in-flight requests in regular HTTP routing

  sgl-project/sglang#47868 · Filter by author liruiAXE opened on Nov 2, 2026
  1