All pull requests

Pull requests

• [Serving] Add --max-payload-size flag to raise the HTTP request payload limit

  documentation
  #48195·saitarrun opened Nov 3, 2026·

  ·
  1

  3

• [Fix] [AMD] kimi-k3: Restore the KDA in-proj fusion for quantized che…

  #48194·amd-mvarjoka opened Nov 3, 2026Contributor·1 of 5 tasks·

  ·

• [Fix] deps: add websockets so /v1/realtime works without extras

  dependenciesnpu
  #48193·T0mSIlver opened Nov 3, 2026·1 of 3 tasks·

  ·
  1

• [Chore] mooncake: bump mooncake version to 0.3.14

  mooncake
  #48192·ShangmingCai opened Nov 3, 2026Collaborator·6 tasks done·

  ·
  6

• [Feat] diffusion/realtime: Add a client-side benchmark for realtime video sessions

  diffusion
  #48191·stayinalive181 opened Nov 3, 2026·2 of 3 tasks·

  ·

• [CI] [AMD] rerun-test: Add --amd to rerun tests on their AMD runners

  amddocumentation
  #48190·bingxche opened Nov 3, 2026Collaborator·

  ·
  3

• [Feat] [NVIDIA] deepseek-v4: add opt-in row-wise MoE pipelines

  deepseekdependenciesdocumentationnpu
  #48189·xyDong0223 opened Nov 3, 2026·4 of 6 tasks·

  ·

• [Fix] spec: Keep chain MTP prefill hidden states under DP attention

  #48188·ZhouYuxuanYX opened Nov 3, 2026Contributor·4 of 5 tasks·

  ·

• [Fix] spec: Apply the multi-layer boundary-KV fix to non-hybrid chain MTP drafts

  #48187·ZhouYuxuanYX opened Nov 3, 2026Contributor·3 of 5 tasks·

  ·

• [Fix] rust-server: Parse JSON tool calls for named and required choices

  #48181·bvolpato opened Nov 3, 2026Contributor·3 of 5 tasks·

  ·

• [Fix] openai: Keep one BOS on text-only multimodal prompts

  Multi-modal
  #48180·bvolpato opened Nov 3, 2026Contributor·3 of 5 tasks·

  ·

• [Fix] kimi-k3: Verify DFLASH on cutedsl_mla under DCP

  documentation
  #48178·bvolpato opened Nov 3, 2026Contributor·4 of 5 tasks·

  ·
  1

• [CI] rerun-test: Retry diffusion reruns like the multimodal-gen CI

  #48176·leejnau opened Nov 3, 2026Collaborator·1 task done·

  ·
  1

• [Fix] anthropic: don't reject adaptive thinking when there is no reasoning parser

  #48175·T0mSIlver opened Nov 3, 2026·2 of 3 tasks·

  ·
  1

• [Perf] [NVIDIA] moe: Re-land the JIT moe_finalize_fuse_shared with the shared row read after the PDL wait

  jit-kernelrun-ci
  #48174·DarkSharpness opened Nov 3, 2026Collaborator·

  ·
  6

• [Feat] [AMD] diffusion: Add MXFP6 and A6W4 online GEMM quantization

  diffusionquant
  #48173·avjves opened Nov 3, 2026Contributor·4 of 5 tasks·

  ·

• [Perf] diffusion: Share convrot_int8 input quantization across LTX-2 projections

  diffusionnpu
  #48171·syang-genai opened Nov 3, 2026·4 tasks done·

  ·

• [Test] [NPU] rl: Add pause/resume scheduler idempotency and 1-NPU smoke tests

  npu
  #48169·SeaAsters opened Nov 3, 2026·

  ·
  1

• [Test] [NPU] rl: Add sleep/wake scheduler unit and 1-NPU smoke tests

  npu
  #48168·SeaAsters opened Nov 3, 2026·

  ·
  1

• [Feat] hicache: Register draft Mamba state pool

  hicache
  #48167·sshleifer opened Nov 3, 2026Contributor·

  ·
  1

• [Perf] [NVIDIA] moe: Add L4 Triton fused MoE configs for Qwen3-30B-A3B

  #48166·shijie-lyu opened Nov 3, 2026Contributor·3 of 5 tasks·

  ·

• [Perf] quantization: Speed up Hopper group32 FP8 quantization with FP32 scales

  jit-kernelquant
  #48165·Frank-whw opened Nov 3, 2026·

  ·
  3

• [Fix] kimi-k3: Reuse the attention-residual bank across forward passes

  #48164·amd-mvarjoka opened Nov 3, 2026Contributor·1 of 5 tasks·

  ·

• [Fix] hicache: Register the whole host-pool allocation with Mooncake

  hicachemooncake
  #48163·NUABO opened Nov 3, 2026·1 of 5 tasks·

  ·

• [Feat] [AMD] dsv4: FP4 indexer with natural layout and OPUS MQA logits

  amddeepseekjit-kernelmemory-pool
  #48162·inkcherry opened Nov 3, 2026Contributor·

  ·