Search results

[Feature] DeepSeek-V4 hybrid KV pool: host/storage integrations assume k/v two-buffer pools (LMCache connector, decode KV offload)

Status: Open.

#31207 In sgl-project/sglang;

· d3nb opened on Jul 14, 2026

[Bug] sgl-router (PD): open circuit breaker still dispatches to decode, producing a permanent fake-dead prefill

Status: Open.

#31206 In sgl-project/sglang;

· d3nb opened on Jul 14, 2026

[Bug] PD prefill silently starves forever when a request's SWA budget exceeds the SWA pool (DeepSeek-V4 hybrid pool)

Status: Open.

#31205 In sgl-project/sglang;

· d3nb opened on Jul 14, 2026

[Bug] Qwen3-VL video frames double-sampled when --mm-process-config sets fps (visual tokens halved since v0.5.11)

Status: Open.

#31200 In sgl-project/sglang;

· SkipperJK opened on Jul 14, 2026

[Bug] Humming quantization probe raises KeyError for configs without quant_method

Status: Open.

#31183 In sgl-project/sglang;

· jeremyzhang866 opened on Jul 14, 2026

[Bug] D node deployed with sglang glm 5.2 pd is throwing an error when using speculative decoding.

Status: Open.

#31178 In sgl-project/sglang;

· yangzhipeng1108 opened on Jul 14, 2026

Add native OLMo3 (Olmo3ForCausalLM) support by reusing the Olmo2 implementation

Status: Open.

#31175 In sgl-project/sglang;

· yihao-liang opened on Jul 14, 2026

[Bug] MiniMax sparse prefill: OOB GPU write when max_seqlen_k < seq_lens.max() (silent topk corruption / Xid 31 / NCCL watchdog kills)

Status: Open.

#31133 In sgl-project/sglang;

· yifjiang opened on Jul 14, 2026

[Feature] any plan to support suffix decoding?

Status: Open.

#31127 In sgl-project/sglang;

· JINO-ROHIT opened on Jul 14, 2026

[Bug] Significant Performance Degradation of Qwen3.5-4B on RTX 5090

Status: Open.

#31120 In sgl-project/sglang;

· SundayVHan opened on Jul 14, 2026

Fix infinite thinking loop caused by cross-chunk tag truncation in Qwen3 streaming

Status: Open.

#31118 In sgl-project/sglang;

· HH1162 opened on Jul 14, 2026

[Bug] custom all-reduce (V2 one-shot push) deadlocks when one communicator is issued concurrently from two CUDA streams

Status: Open.

#31117 In sgl-project/sglang;

· hanyax opened on Jul 14, 2026