All pull requests

Pull requests

• feat(moe): native MXFP8 W8A8 grouped MoE, including 32-aligned intermediate sizes

  #449·randomvariable opened Sep 30, 2026·

  ·
  18

• Karmic beta to master (Sep 30): #427, #420–#424 (MiMo-V2.6), #440

  #448·voipmonitor opened Sep 30, 2026Contributor·

  ·
  9

• Serve DS4.1 and Kimi-K3 X4T with native W4A16 scale decoding

  #446·voipmonitor opened Sep 29, 2026Contributor·

  ·
  6

• Retain Trellis route-pack programs in prepared MoE plans

  #442·voipmonitor opened Sep 29, 2026Contributor·

  ·
  1

• Allow FP32 split partials in prepared dense MLA

  #441·voipmonitor opened Sep 29, 2026Contributor·

  ·
  1

• attention: V4.1 FP4 KV writer divides like the DeepSeek reference (master counterpart of #435)

  #440·voipmonitor opened Sep 29, 2026Contributor·

  ·
  3

• comm.roce: host-visible completion, a wrap-safe failure flag, and a wedge snapshot

  #438·tobymao opened Sep 27, 2026Contributor·

  ·
  9

• [Research] Isolate optional stable W4A16 route packing

  #437·lukealonso opened Sep 27, 2026Collaborator·

  ·
  1

• perf(moe): full-codeword lookup for prepared K2 Trellis decode

  #436·lukealonso opened Sep 27, 2026Collaborator·

  ·
  2

• feat(qsa): serve sparse attention from an NVFP4 main K/V cache

  #434·yatesdr opened Sep 27, 2026Contributor·

  ·
  18

• Support PCIe one-shot and DMA all-reduce at world size 3; make PDL flags effective

  #427·voipmonitor opened Sep 24, 2026Contributor·

  ·
  4

• perf(moe): 16-byte loads and stores in the W4A16 gated activation

  #424·MadeBy561 opened Sep 23, 2026Contributor·

  ·
  1

• feat(moe): optional FP32 router weights in the W4A16 top-k combine

  #423·MadeBy561 opened Sep 23, 2026Contributor·

  ·
  1

• perf(gemm): small-row GEMV for serialized block-FP8 plans

  #422·MadeBy561 opened Sep 23, 2026Contributor·

  ·
  1

• perf(gemm): tensor-core GEMV backend for wide BF16 decode projections

  #421·MadeBy561 opened Sep 23, 2026Contributor·

  ·
  6

• feat(attention): add paged_decode split-KV decode/verify with KV append

  #420·MadeBy561 opened Sep 23, 2026Contributor·

  ·
  9

• fix(gemm): serve BF16 MLA query rows up to the declared capacity

  #392·tobymao opened Sep 17, 2026Contributor·

  ·
  5

• perf(qsa): reuse representative keys across paired queries

  #387·original-el8 opened Sep 16, 2026Contributor·

  ·
  1

• feat(ple): export prepared internal prefill checkpoints

  #386·original-el8 opened Sep 16, 2026Contributor·

  ·
  4

• fix(dsa_indexer): accept topk 512/1024/2048 in the MXFP4 gates

  #385·fernandaspets opened Sep 16, 2026·

  ·
  41

• fix(preparation): retain prepared launchers and coordinate collectives

  #384·randomvariable opened Sep 16, 2026·

  ·
  23

• Add native SM120 decode-context-parallel primitives for DeepSeek V4.1

  #374·yatesdr opened Sep 15, 2026Contributor·

  ·
  28

• fix(comm): prepare two-shot executables and native row geometry

  #373·voipmonitor opened Sep 14, 2026Contributor·

  ·
  5

• perf(loader): allow filesystem caching for disk tables

  #370·MadeBy561 opened Sep 14, 2026Contributor·

  ·
  3

• test(moe): cover canonical A16 binding on the host

  #369·original-el8 opened Sep 13, 2026Contributor·

  ·
  1