All issues

Search results

[Bug] test_expert_pack_mxfp4.py intermittently hangs while loading JIT extension

Status: Open.

#38408 In sgl-project/sglang;

· nvpohanh opened on Sep 8, 2026

[Bug] bench_gdn_replayssm_decode.py never advances write_pos, so it times only the non-flush phase and overstates the speedup by up to 21%

Status: Open.

#38398 In sgl-project/sglang;

· davidli1515 opened on Sep 8, 2026

[Bug] /v1/score returns validation error for single-label cross-encoder rerankers (e.g. bge-reranker-v2-m3)

Status: Open.

#38360 In sgl-project/sglang;

· ChengjieLi28 opened on Sep 7, 2026

[RFC / Feature]: Zero-Copy In-Place KV-Cache Compactor via In-Situ Cycle Permutations (1.65x Speedup, 0 Bytes Aux VRAM)

Status: Open.

#38357 In sgl-project/sglang;

· aemre-cetin opened on Sep 7, 2026

[decode] Spec-decode with large draft-token counts OOMs mid-traffic on the decode instance - no upfront capacity check (verify workspace vs KV pool is zero-sum)

Status: Open.

#38341 In sgl-project/sglang;

· Leslie360 opened on Sep 7, 2026

[RFC] Gluon MegaMoE: SGLang Integration and Multi-Node Support

Status: Open.

#38334 In sgl-project/sglang;

· Sulfur6 opened on Sep 7, 2026

[Bug] Chunked Prefill + Radix Insert race corrupts KV pages (QSA, Qwen3.8-Flash-Next)

Status: Open.

#38319 In sgl-project/sglang;

· andreasknopke opened on Sep 7, 2026

[Playground] Verified cell: h100 / flash-official / fp4 / high-throughput / single

Status: Open.

#38312 In sgl-project/sglang;

· mtc-20 opened on Sep 7, 2026

TP2 hang with HiCache, breakable prefill CUDA graphs, and FlashInfer MNNVL on B300

Status: Open.

#38300 In sgl-project/sglang;

· nvpohanh opened on Sep 7, 2026

[Bug] fp8e4nv not supported on A100 (SM80) when serving Qwen3.8-Flash-Next-FP8

Status: Open.

#38291 In sgl-project/sglang;

· genghisun opened on Sep 7, 2026

[Bug] glm5_next qkvbfg fusion disabled on FP8 GLM-5.3-Flash: gate checks global quant_config, not per-layer

Status: Open.

#38253 In sgl-project/sglang;

· linkedlist771 opened on Sep 7, 2026

[Bug] DeepseekV4ForCausalLM hook rejects --speculative-algorithm NEXTN because the NEXTN->EAGLE alias is resolved after model-specific hooks

Status: Open.

#38236 In sgl-project/sglang;

· Champollion9012 opened on Sep 6, 2026