All pull requests

Pull requests

• Fix: validate GLM checkpoint storage before staging

  #1752·Inspiron-st opened Nov 9, 2026Collaborator·

  ·
  2

• Perf: simplify GLM Sinkhorn and convolution arithmetic

  #1751·oahuohzz opened Nov 9, 2026Contributor·

  ·
  4

• Refactor: migrate HC Sinkhorn to PyPTO

  #1750·yiyuan-123 opened Nov 9, 2026·

  ·
  1

• Fix: keep CP-CSA test scratch rank-local

  #1749·MaxwellF1 opened Nov 9, 2026Contributor·

  ·
  9

• Add: refine hardware utilization and add FA MFU

  #1748·parsifal-wang-1995 opened Nov 9, 2026Contributor·

  ·
  6

• feat(lowlat): adapt DSpark prefill KV cache

  #1747·zzjc1234 opened Nov 9, 2026Contributor·

  ·
  7

• Add: PyPTO gather_packed alongside the V4.1 extern kernel

  #1746·Little-oil opened Nov 9, 2026Contributor·

  ·
  2

• CI: add complete GLM MoE prefill and decode cases

  #1743·Inspiron-st opened Nov 9, 2026Collaborator·

  ·
  7

• Fix: keep DeepSeek CSA validation scratch rank-local

  #1742·MaxwellF1 opened Nov 9, 2026Contributor·

  ·
  9

• Add: PyPTO register Index-K publication functional version

  #1741·wangqin1723-max opened Nov 9, 2026Collaborator·

  ·
  6

• qwen3_14b: size hard-syncall SPMD launches from device cluster count

  #1739·blacksde opened Nov 9, 2026·

  ·
  8

• CI: make lowlat decode layer failures replayable

  #1738·wangqin1723-max opened Nov 9, 2026Collaborator·

  ·
  2

• Perf: reuse queries and paged KV in the FLASH prefill indexer

  #1737·c247274901 opened Nov 9, 2026Contributor·

  ·
  2

• Perf: Optimize GLM KDA prefill and recurrent decode

  #1733·Inspiron-st opened Nov 7, 2026Collaborator·

  ·
  2

• Perf: accelerate GLM projections and recurrent KDA

  #1730·oahuohzz opened Nov 7, 2026Contributor·

  ·
  10

• Add: Fused C++ MoE return kernel for DSV4.1 decode

  #1726·parsifal-wang-1995 opened Nov 6, 2026Contributor·

  ·
  6

• Refactor: unify DeepSeek V4.1 standalone RMSNorm

  #1695·zhangqi-chen opened Nov 5, 2026Collaborator·

  ·
  2

• Perf: fuse V4.1 packed sparse attention through L1

  #1642·wangqin1723-max opened Nov 2, 2026Collaborator·

  ·
  4

• Fix: recover corrupt cached PyPTO source checkouts

  #1634·chenshengxin2026 opened Oct 31, 2026Contributor·

  ·
  1

• Add: native FP4 scoring for V4.1 decode Indexer

  enhancement
  #1605·lwDavid opened Oct 28, 2026Collaborator·

  ·
  8

  

• Fix: align V4.1 decode HC projection with HF32 rounding

  enhancement
  #1599·lwDavid opened Oct 28, 2026Collaborator·

  ·
  12

  

• Fix: split host and device request mappings

  #1598·high-cloud opened Oct 28, 2026Collaborator·

  ·
  1

  4

• Fix: align DSpark kernel interfaces and reference arithmetic

  #1550·Little-oil opened Oct 22, 2026Contributor·

  ·
  4

• Add: integrate resident Engram lookup into V4.1 forwards

  #1539·high-cloud opened Oct 21, 2026Collaborator·

  ·
  6

• Add Qwen3.8-27B PyPTO text network

  #1503·ndleslx opened Oct 19, 2026Contributor·

  ·
  1