All pull requests

Pull requests

• [Fix] startup: suppress unrelated model import messages

  #46620·merrymercy opened Oct 26, 2026Collaborator·4 of 6 tasks·

  ·

• [Perf] Capture GDN prefill in breakable CUDA graphs

  jit-kernel
  #46619·wirybeaver opened Oct 26, 2026Contributor·

  ·

• [Fix] dspark: Avoid host reads during draft metadata glue capture

  deepseek
  #46618·feng397 opened Oct 26, 2026Contributor·3 of 5 tasks·

  ·

• [Fix] function_call: Preserve valid DeepSeek calls when a sibling is malformed

  deepseek
  #46617·ycheng-allen opened Oct 26, 2026·3 of 5 tasks·

  ·

• docs: fix 'maxium' typo in post-training integration page

  documentation
  #46616·aarushkandukoori opened Oct 26, 2026·

  ·
  2

• [Refactor] pd: Extract NixlTransport behind the common Transport interface

  mooncakenpu
  #46614·jambow0320 opened Oct 26, 2026Contributor·

  ·

• [Perf] diffusion: Cosmos-Dreams rollout-owned prefix K/V buffer

  diffusion
  #46613·OnePunchMonk opened Oct 26, 2026Contributor·

  ·
  1

• [Fix] [AMD] dsv41: Make Engram's input_ids optional for the AMD monokernel

  #46612·AMD-yanfeiwang opened Oct 26, 2026Contributor·3 of 5 tasks·

  ·

• [Fix] forward-hooks: Warn when CUDA graphs skip Python hooks

  #46611·chetaniitbhilai opened Oct 26, 2026·1 of 3 tasks·

  ·
  1

• [Feat] [Intel] mem_cache: Enable the HiCache Mamba host pool on XPU

  hicache
  #46610·kpjeeja opened Oct 26, 2026Contributor·4 tasks done·

  ·
  4

• [Perf] [AMD] moe: Raise the small MoE sort's pair limit from 256 to 1024

  jit-kernel
  #46609·tkarttunen opened Oct 26, 2026·3 of 5 tasks·

  ·

• [Perf] [AMD] moe: Add opt-in AITER MegaMoE-TP for eager TP MoE forwards

  documentation
  #46608·eppaneamd opened Oct 26, 2026·5 tasks done·

  ·
  1

• [Fix] responses: Align K3 input token usage with Chat

  #46607·LeslieWylie opened Oct 26, 2026·

  ·
  1

• [Fix] cache: Roll back host slots when allocation raises

  hicache
  #46606·LeslieWylie opened Oct 26, 2026·

  ·
  1

• [Fix] [XPU] docker: install oneCCL 2022.1.2 graph-mode runtime set

  documentation
  #46605·rahulvijayaraghavan opened Oct 26, 2026Contributor·2 tasks done·

  ·
  1

• [Feat] [NPU] hicache: Support layer-first split host transfers

  hicachememory-pool
  #46604·ZhengXR-L opened Oct 26, 2026·5 tasks·

  ·

• [Perf] diffusion model: Accelerate conditioning hashing with GPU BLAKE3

  diffusionjit-kernel
  #46603·xupinjie opened Oct 26, 2026Contributor·4 of 5 tasks·

  ·
  1

• [Doc] hicache: Fix MLA expansion in HiCache design doc

  documentationhicache
  #46602·bLueriVerLHR opened Oct 26, 2026·3 of 5 tasks·

  ·
  1

• [Fix] diffusion: LingBot QKV sharing, causal RoPE on no-cache path, uneven split guard

  diffusion
  #46601·OnePunchMonk opened Oct 26, 2026Contributor·

  ·

• [Fix] diffusion: HunyuanImage-3 MoE shared MLP and timestep schedule on CUDA

  diffusion
  #46600·OnePunchMonk opened Oct 26, 2026Contributor·2 tasks done·

  ·
  1

• [Feat] disaggregation: Add connex as a KV transfer backend

  documentation
  #46599·yanyinglin opened Oct 26, 2026·4 tasks done·

  ·
  5

• [Perf] [NVIDIA] minimax-sparse: Run the decode indexer CP on Hopper at TP4 and TP8

  #46597·ischencheng opened Oct 26, 2026·4 of 5 tasks·

  ·

• [Feat] metrics: expose scheduler-to-router delivery latency

  documentation
  #46596·yu830 opened Oct 26, 2026·5 tasks done·

  ·
  3

• [Fix] openai: match tokenize reasoning effort to chat

  deepseek
  #46595·YangChen-pro opened Oct 26, 2026Contributor·

  ·
  1

• [Fix] kda: Honor state and replay dtype

  jit-kernelmemory-pool
  #46594·yhyang201 opened Oct 26, 2026Collaborator·

  ·
  4