All pull requests

Pull requests

• [ROCm] v0.2.0 feature support on AMD Instinct

  #171·0z5a opened Nov 5, 2026Contributor·

  ·

• Fix missing CUDA coverage note for default profiling activities

  #170·huawanrr opened Nov 4, 2026·

  ·
  1

• [Feature] Add experimental Ascend support for Ouro

  #168·KrystalRay opened Nov 3, 2026·

  ·
  2

• [Kernels] Fuse Ouro RMSNorm and residual with a precision-compatible CUDA path

  #167·bjf-frz opened Nov 3, 2026Collaborator·

  ·

• [KV] Add INT4/NVFP4 reference codecs and cross-loop residual reconstruction

  #166·HowardYangAI opened Nov 3, 2026·

  ·
  6

• [Bugfix]Fix livelock under wcpb scheduling

  #165·LyxWxj opened Nov 2, 2026Contributor·

  ·

• sparse decode attention: CUDA graphs, P/D, asynchronous decode, token pipelines and preemption

  #164·0z5a opened Nov 2, 2026Contributor·

  ·

• feat: FlashInfer FA2 paged decode backend for Ada (SM89)

  #162·sysuTobo opened Nov 2, 2026·4 tasks done·

  ·
  3

• [Docs] Index Ouro and Nanbeige validation evidence

  #161·dorodo0721 opened Nov 1, 2026·

  ·
  1

• Sparse decode attention: per-token reference lifecycle for async self-speculation

  #157·kunkunblueberry opened Oct 27, 2026Contributor·

  ·
  65

• [Docs] Record BeNice! KV sharing and AsymLoop explorations as not qualified

  #155·0z5a opened Oct 27, 2026Contributor·

  ·
  1

• [REFACTOR]M7 3/3: validate sampling and RNG contracts on CUDA

  #153·ycjcx123 opened Oct 26, 2026Contributor·

  ·
  3

• [Engine] Packed prefill resolves automatically, for Huginn too

  #149·0z5a opened Oct 25, 2026Contributor·

  ·

• [Scheduler] Fuse fixed-depth decode loops into one submission, keeping the per-loop prefill share

  #148·0z5a opened Oct 25, 2026Contributor·

  ·
  1

• [KV] Pack recurrent depth planes for partially recurrent models

  #147·0z5a opened Oct 25, 2026Contributor·

  ·

• [Feature] WISE working-set decode for recurrent-depth models

  #143·0z5a opened Oct 23, 2026Contributor·

  ·
  7

• [Doc][M9] PD and KV transfer walkthrough

  #142·YouHengfei opened Oct 23, 2026Contributor·

  ·

• test(nanbeige): qualify recurrent CUDA Graph capture, replay, and fal…

  #139·cs-ai-li opened Oct 22, 2026Contributor·

  ·

• test(nanbeige): qualify chunked prefill, priority preemption, and prefix caching

  #138·cs-ai-li opened Oct 22, 2026Contributor·

  ·
  3

• [Scheduler] Separate request lifecycle from scheduled work

  #137·LyxWxj opened Oct 22, 2026Contributor·

  ·
  7

• Add opt-in loop-level speculative execution

  #133·YuDeng0102 opened Oct 18, 2026·

  ·
  1

• Support Opt-in FP8 KV Cache quantization with FA2-4 backend.

  #129·0z5a opened Oct 15, 2026Contributor·

  ·
  2

• [Feature] Add experimental deferred Ouro deep batching

  #125·Camellia86 opened Oct 14, 2026·

  ·

• docs: record alternate-branch speculation results

  #120·prettygirlisnotme opened Oct 13, 2026·

  ·
  2

• Add a real-checkpoint P->D agreement test for the RFC #111 handoff contract

  #119·jyizheng opened Oct 12, 2026·

  ·