Issues

• [Bug] DSpark SPS profiler mis-reads the attention-DP rank count from /server_info

  sgl-project/sglang#42601 · Filter by author apex-mochen opened on Oct 5, 2026

• [Bug] chain_speculative_sampling_triton rejects every draft whose q rounds just above 1.0 (since #37134)

  sgl-project/sglang#42568 · Filter by author dg1kjd opened on Oct 5, 2026
  1

• [Bug][ROCm] FLUX.2-dev fails on gfx1151: NVIDIA PTX inline assembly in residual_gate_add kernel

  sgl-project/sglang#42564 · Filter by author vvvdwbvvv opened on Oct 5, 2026

• [Bug] Deprecated --dp-size N --enable-dp-attention resolves to dp_size=1 (attn_dp_size=N), so embedding integrations (NVIDIA Dynamo) register only DP rank 0; still present after #42348

  sgl-project/sglang#42560 · Filter by author janbernloehr opened on Oct 5, 2026

• [Feature] Support Whisper speech translation (task="translate" / /v1/audio/translations)

  sgl-project/sglang#42559 · Filter by author haseebkhawaja opened on Oct 5, 2026

• [Bug] DeepSeek-V4 HiCache: SWA host pool is sized from the small SWA device pool, so idle prefixes miss the host tier long before the KV host pool fills

  sgl-project/sglang#42548 · Filter by author divyvasal opened on Oct 5, 2026

• [Bug] /update_weights_from_disk: is_async / keep_pause / token_step have no effect, num_paused_requests is always 0

  sgl-project/sglang#42544 · Filter by author apex-mochen opened on Oct 5, 2026

• [Bug] big prefill blocks other requests on 2xRTX3090 for qwen3.8 27b

  sgl-project/sglang#42530 · Filter by author vargrant opened on Oct 4, 2026

• [Bug] tree_speculative_sampling_target_only rejects the only token with target mass when the coin equals the largest float32 below 1 and emits token id vocab_size - 1

  sgl-project/sglang#42528 · Filter by author clayboby opened on Oct 4, 2026

• [Bug] EAGLE/MTP same-checkpoint draft keeps redundant embed_tokens/lm_head copies resident during KV pool sizing → under-sized pool, startup OOM

  sgl-project/sglang#42510 · Filter by author paulalesius opened on Oct 4, 2026

• [Bug] Scheduler aborts with double free or corruption inside the idle-loop invariant check (session_held_tokens walk); server hangs permanently afterwards

  sgl-project/sglang#42508 · Filter by author LUOSENGWA opened on Oct 4, 2026
  2

• Development Roadmap (2026 Q4)

  sgl-project/sglang#42473 · Filter by author merrymercy opened on Oct 4, 2026

• [Bug] DeepSeek-V4 + HiCache write_through: TP ranks deadlock under concurrent long prefills (scheduler and detokenizer go silent, /health 503)

  sgl-project/sglang#42465 · Filter by author divyvasal opened on Oct 4, 2026
  2

• [Feature] Upgrade Nsight Compute bundled by the CUDA devel base image

  sgl-project/sglang#42459 · Filter by author YouNeedCryDear opened on Oct 4, 2026

• fix(mlx): repeated chat returns unrelated text after native generation

  sgl-project/sglang#42415 · Filter by author coygeek opened on Oct 3, 2026
  1

• [Feature] Cloudflare Clef support

  sgl-project/sglang#42409 · Filter by author RBozydar opened on Oct 3, 2026

• [Bug] --speculative-token-map crashes at TP>1 (device-side assert in init_lm_head): global hot-token ids index a vocab-sharded lm_head

  sgl-project/sglang#42397 · Filter by author xin-osaka-u opened on Oct 3, 2026

• [Bug] DeepSeekV31Detector streaming drops the arguments when a complete tool call arrives in one delta

  sgl-project/sglang#42373 · Filter by author he-yufeng opened on Oct 3, 2026

• [Bug][KV Cache] nvfp4 KV silently reuses the checkpoint's fp8-calibrated k/v_scale as NVFP4 global scale → deterministic long-context corruption (single-variable proof, sm_120)

  sgl-project/sglang#42369 · Filter by author chengye75-star opened on Oct 3, 2026
  3

• [Bug] SM120: DeepSeek-V4.1-Flash decode CUDA graph capture fails with FlashInfer 0.7.0 because _DECODE_DSV4_DISPATCH is no longer iterable

  sgl-project/sglang#42367 · Filter by author juliankang4 opened on Oct 3, 2026

• [Bug] DeepSeek-V4.1 fails at attention backend init on SM12x (GB10): the candidate indexer requires DeepGEMM paged sparse MQA logits, which are SM100-only

  sgl-project/sglang#42364 · Filter by author juliankang4 opened on Oct 3, 2026

• [Bug] DeepSeek V4 Pro loading on Lustre takes 95 minutes despite prefetch; limiting tensor-copy workers reduces it to 3.3 minutes

  sgl-project/sglang#42361 · Filter by author janbernloehr opened on Oct 3, 2026
  1

• [Bug] DeepSeek V3 tool-call parser streams arguments "{}" when the closing fence and end tokens arrive in one stream event (MTP / --stream-interval > 1)

  sgl-project/sglang#42327 · Filter by author wenqiw777 opened on Oct 3, 2026

• [Bug] Anthropic /v1/messages: inline system messages are still merged into the top system block on system-first templates (Qwen), so the prefix cache misses every turn

  sgl-project/sglang#42272 · Filter by author rgrashel opened on Oct 2, 2026

• response_format + tools on glm47: tool calls silently dropped, model invents a JSON answer

  sgl-project/sglang#42269 · Filter by author justicel opened on Oct 2, 2026