Pull requests

• [Fix] openai: Apply --preferred-sampling-params on /v1/chat/completions

  #45611·saitarrun opened Oct 21, 2026·

  ·
  1

  1

• [CI] [NPU] nightly-test: Switch nightly images to nightly-cann9.2.0

  npu
  #45609·pllimax opened Oct 21, 2026Contributor·1 of 5 tasks·

  ·
  1

• [Feat] memory: Allow prefill CUDA graph bs below chunked_prefill_size under post-capture KV sizing

  memory-pool
  #45608·Eliaaazzz opened Oct 21, 2026·3 of 5 tasks·

  ·

• [Fix] scheduler: Align truncated hybrid Mamba chunks to KV pages

  #45607·jongukc opened Oct 21, 2026·5 tasks done·

  ·
  1

• [Fix] disaggregation: Give parallel disaggregation samples distinct rooms

  #45605·jongukc opened Oct 21, 2026·5 tasks done·

  ·
  1

• [Fix] mamba: Skip zero-length Mamba checkpoint reads

  unified-radix-cache
  #45604·jongukc opened Oct 21, 2026·5 tasks done·

  ·
  2

• [npu]fix memfabric and hccl init conflict

  hicachenpu
  #45603·jamesjxliu opened Oct 21, 2026Contributor·5 tasks·

  ·
  1

• [Perf] [AMD] deterministic: Allow AITER prefill attention on gfx950

  #45601·aarnetalman opened Oct 21, 2026Contributor·4 of 5 tasks·

  ·

• [Feat] linker: Add prefetch mode for the Mooncake external linker

  documentationhicacheunified-radix-cache
  #45600·huangtingwei9988 opened Oct 21, 2026Collaborator·5 tasks·

  ·
  2

• [Perf] [AMD] batch-invariant: Segmented split-K bf16 matmul for gfx950

  deterministic
  #45599·aarnetalman opened Oct 21, 2026Contributor·4 of 5 tasks·

  ·

• [BugFix] validate_buckets_rule: raise ValueError instead of assert False on malformed rules

  #45598·mlleo opened Oct 21, 2026Contributor·2 of 4 tasks·

  ·
  1

• [Doc] [NPU] GLM: Add Ascend cookbook deployment recipes

  documentation
  #45597·stellaxcpeng opened Oct 21, 2026Contributor·

  ·
  2

• [CP] Run FlashAttention zigzag prefill as one combined launch

  #45596·wjinxu opened Oct 21, 2026·4 of 5 tasks·

  ·
  1

• [Fix] docker: Bump sgl-nvidia-mega to 3333e4c so the megakernel runs on main

  #45595·BBuf opened Oct 21, 2026Collaborator·3 of 5 tasks·

  ·
  1

• [Doc] [NPU] cookbook: add A3 deployment scenario to Qwen3.8

  documentation
  #45594·jczhaang opened Oct 21, 2026·4 tasks done·

  ·
  3

• [Fix] disagg: Do not gate PD decode radix prefix match on tree SWA

  #45592·alphabetc1 opened Oct 21, 2026Collaborator·

  ·

• [Perf] moe: Add H200 block-FP8 Triton configs for M1-8192

  #45591·Jzz1943 opened Oct 21, 2026Contributor·3 of 5 tasks·

  ·
  2

• [Fix] [NPU] tests: Prefetch the sgl_eval MMLU archive and pin the Mamba SSM dtype

  hicachenpu
  #45589·pllimax opened Oct 21, 2026Contributor·5 tasks·

  ·
  3

• [Feat] GLM: Match reasoning and tool-call markers by output token id

  run-ci
  #45587·JustinTong0323 opened Oct 21, 2026Collaborator·

  ·
  1

• [Fix] dsv4: Write idle SWA stores to an idle request-window layout

  deepseek
  #45586·hnyls2002 opened Oct 21, 2026Collaborator·

  ·

• [Fix] [SM100] quantization: Honor deep_gemm for 32x32 block-FP8 linears

  #45585·hnyls2002 opened Oct 21, 2026Collaborator·

  ·

• [Fix] server: Serialize multi-tokenizer general warmups

  #45584·Phoenix3334 opened Oct 21, 2026Contributor·3 of 4 tasks·

  ·
  1

• [Perf] kda: Size the speculative ReplaySSM ring by running requests

  documentationjit-kernelmemory-pool
  #45583·vineelbhat opened Oct 21, 2026Contributor·5 tasks done·

  ·
  1

• [Perf] dsv4: Publish prefill read-done under the request window and on V4.1

  deepseekhicacheunified-radix-cache
  #45582·hnyls2002 opened Oct 21, 2026Collaborator·

  ·

• [Perf] dsv4: Let the decoder tail leave the prefill graph under DP attention with MegaMoE

  deepseek
  #45581·hnyls2002 opened Oct 21, 2026Collaborator·

  ·