All pull requests

Pull requests

• [Perf] Use finer gfx1250 index-score chunks for small decode grids

  #2873·xytpai opened Oct 31, 2026Contributor·

  ·
  7

• feat(mono): shared DSA indexer mono launch; Kimi-K3 mono to 64 rows, plain decode and ReplaySSM

  ci:atom
  #2871·valarLip opened Oct 31, 2026Collaborator·

  ·
  7

• [MHA] Unified-attention prefill: run the kernel over the scheduled rows only

  #2870·vgokhale opened Oct 31, 2026·

  ·

• [Qwen3.5] c2048 DEP4: stop prefix-resumed requests decoding garbage (GDN prefill eager in DP prefill graphs, drain before state copies), unified-attention pad-row OOM

  #2868·vgokhale opened Oct 30, 2026·

  ·

• feat(api): constrain JSON output and expose deployment capabilities

  #2862·ZhangLirong-amd opened Oct 30, 2026Collaborator·

  ·
  2

• feat(v41): replay the decoder graph on over-band prefills (opt-in)

  #2860·carlushuang opened Oct 30, 2026Collaborator·4 tasks done·

  ·
  1

• [DSV4.1] Opt-in head-blocked query RoPE in the fused rope/window seam

  #2859·yhl-amd opened Oct 30, 2026Collaborator·

  ·
  5

• Fix standalone Inductor/Triton cache persistence across container restarts

  #2858·xytpai opened Oct 30, 2026Contributor·

  ·
  34

• feat(docker): move the default ROCm line from 7.2.4 to the 7.14.1 pip SDK

  ci:atom
  #2857·junyyang-amd opened Oct 30, 2026Contributor·5 tasks·

  ·
  1

• feat(api): return Chat token logprobs with DSpark support

  #2856·ZhangLirong-amd opened Oct 30, 2026Collaborator·

  ·

• perf(attn): unified attention for prefix-hit FP8 prefill + start-up warm-up

  ci:atom
  #2855·ZhiweiYan-96 opened Oct 30, 2026Contributor·

  ·
  2

• Qwen3.5 MTP: gather the fc output with aiter's custom all-gather instead of RCCL

  ci:atom
  #2854·ZhiweiYan-96 opened Oct 30, 2026Contributor·

  ·
  2

• MTP draft mid-steps in CUDA graphs for MRoPE models (Qwen3.5)

  ci:atom
  #2853·ZhiweiYan-96 opened Oct 30, 2026Contributor·

  ·
  4

• perf(qwen3.5): prefill all-reduce + norm via INT4 quick-reduce for 2k-8k-token steps

  ci:atom
  #2852·ZhiweiYan-96 opened Oct 30, 2026Contributor·

  ·

• perf(qwen3.5): compute the GDN gate inside the ReplaySSM kernel

  ci:atom
  #2851·ZhiweiYan-96 opened Oct 30, 2026Contributor·

  ·
  2

• [MiniMax-M3.1] Keep the DSpark draft KV to its sliding window, and send only that over P/D

  #2850·Jasen2201 opened Oct 30, 2026Contributor·

  ·
  7

• feat(eplb): add DeepInfer hot expert selection

  #2848·JiaoliangYu opened Oct 30, 2026Contributor·

  ·
  12

• perf(qwen4_exp): Qwen3.8-Flash-Next TP prefill follow-ups: drafter FlyDSL QSA, row-sharded indexer selection, sequence-parallel drafter

  #2847·zovonoir opened Oct 30, 2026Contributor·9 of 10 tasks·

  ·
  93

• [PD] Unify KV staging/landing and add MiniMax PP-to-TP support

  #2846·MengqingCao opened Oct 30, 2026Contributor·

  ·
  4

• perf(flydsl pa_decode): head-256 batch-1/2 plan ceiling and 48-row pipe for MTP-5 verify (opt-in)

  ci:atom
  #2845·ZhiweiYan-96 opened Oct 30, 2026Contributor·

  ·
  1

• fix(api): align reasoning controls and frontend protocol contracts

  ci:atom
  #2844·ZhangLirong-amd opened Oct 30, 2026Collaborator·

  ·
  22

• fix(attention): keep MTP-5 decode on FlyDSL pa_decode past the gluon query-length limit

  ci:atom
  #2843·ZhiweiYan-96 opened Oct 30, 2026Contributor·

  ·
  9

• fix(state checkpoints): keep block 0 a zero null page and zero released PAGE units

  ci:atom
  #2842·ZhiweiYan-96 opened Oct 30, 2026Contributor·

  ·
  29

• feat(dsv41): add gfx1250 mono backend and separate architecture implementations

  #2840·yifehuan opened Oct 30, 2026Contributor·

  ·
  32

• [MiniMax-M3.1] Allocate index cache only on top-k computing layers

  #2839·MengqingCao opened Oct 30, 2026Contributor·

  ·