All pull requests

Pull requests

• [b12x] Keep a relayed RoCEnante stripe's flag behind its own payload

  #107·FujitsuPolycom opened Nov 13, 2026·5 of 12 tasks·

  ·
  1

• [b12x] Broadcast all-gather chunks once per direction on RoCE open chains (proxy ABI 15)

  #106·FujitsuPolycom opened Nov 13, 2026·6 of 12 tasks·

  ·
  1

• b12x RoCE proxy: wait for the doorbell in WFET on aarch64 cores with FEAT_WFxT

  #105·jetnet opened Nov 12, 2026·

  ·
  1

• Add dynamic KDA verification and DCP8 collective support

  #104·voipmonitor opened Nov 12, 2026·7 of 12 tasks·

  ·
  11

• comm/pcie: TP8 two-island all-reduce kernel and runtime for prefill reductions

  #103·tunamitom opened Nov 11, 2026·

  ·
  1

• Accept 9 to 20 rows in the standalone Kimi top-16

  #102·tunamitom opened Nov 11, 2026·

  ·
  1

• Accept 9 to 16 rows in the standalone Kimi top-16

  #101·orangezed opened Nov 11, 2026·

  ·
  1

• Exclude exhausted candidates in the fused paired-gather Kimi top-16

  #100·orangezed opened Nov 11, 2026·

  ·
  1

• Decode CSF expert scales inline in the A4 prefill GEMM

  #96·yatesdr opened Nov 10, 2026·3 of 7 tasks·

  ·
  1

• Add opt-in prefill paths: direct Trellis table for 48-row blocks and expanded lowering for K-padded MXFP8 A16

  #95·orangezed opened Nov 10, 2026·

  ·
  2

• Use measured Trellis fused-MoE launch shapes for prefill token counts

  #94·orangezed opened Nov 10, 2026·

  ·
  2

• Add opt-in two-island transports to the b12x PCIe BF16 two-shot all-reduce

  #93·orangezed opened Nov 10, 2026·

  ·
  3

• Add opt-in small-M decode paths to the b12x Trellis fused W4A16 MoE

  #92·orangezed opened Nov 10, 2026·

  ·
  3

• Add opt-in push, acquire-barrier and two-island transports to b12x PCIe DCP gathers

  #91·orangezed opened Nov 10, 2026·

  ·
  2

• Add a one-launch KDA speculative verify to b12x gdn_decode

  #90·orangezed opened Nov 10, 2026·

  ·
  3

• Select PCIe DMA ring pieces per reduction shape at world 8

  #89·orangezed opened Nov 10, 2026·

  ·
  3

• Tune MXFP8 A16 small-M decode defaults and add opt-in decode search/PDL switches

  #88·orangezed opened Nov 10, 2026·

  ·
  2

• Add an opt-in push transport for the b12x TP8 DCP LSE reduce-scatter

  #87·orangezed opened Nov 10, 2026·

  ·
  2

• Bound windowed PV operands, fix native MXFP8 QK addressing, and guard per-token MSA page reads

  #79·brandonmmusic-max opened Nov 6, 2026·7 of 11 tasks·

  ·
  1

• b12x fused MoE: TrellisMX (procedural NPK) source format with an in-tree P8 runtime

  #69·brandonmmusic-max opened Nov 1, 2026·7 of 11 tasks·

  ·
  1

• b12x mHC: keep FP32 mixing weights in the TF32 projection

  #68·brandonmmusic-max opened Nov 1, 2026·7 of 11 tasks·

  ·
  1

• Evict planned Triton kernels from compiler workers for every JIT module

  #67·Raydius opened Oct 29, 2026·2 of 5 tasks·

  ·
  1

• Race route top-k geometry, vector SwiGLU and short-convolution tiles

  #64·original-el8 opened Oct 28, 2026·6 of 11 tasks·

  ·
  1

• Keep route top-k gathers in bounds for rows without a selectable expert

  #63·original-el8 opened Oct 28, 2026·6 of 11 tasks·

  ·
  1

• Add serialized PCIe profiles and interleaved startup sampling

  #62·voipmonitor opened Oct 28, 2026·5 of 11 tasks·

  ·
  4