All pull requests

Pull requests

• [Bugfix] Scalar GM stores race MTE3 DMA through an incoherent write-back cache (#1304)

  #1834·2Sheldonlaw opened Sep 23, 2026Contributor·

  ·
  1

  7

• [Feat] Migrate add wy_fast_bwd_split operator for Ascend NPU

  #1828·wch0810 opened Sep 22, 2026Contributor·

  ·
  1

• [Feat] Migrate and add new WY Fast operator

  #1827·wch0810 opened Sep 22, 2026Contributor·

  ·
  1

• [Bugfix] Make in-place T.tile.silu safe on AscendC

  #1822·wzzll123 opened Sep 21, 2026Contributor·

  ·
  3

• [Bugfix] Reject non-contiguous flat UB BufferRegions

  #1821·wzzll123 opened Sep 21, 2026Contributor·

  ·
  3

• [Bugfix] Enable automatic C/V handling in Vector operators 🤖

  #1816·platelett opened Sep 20, 2026Contributor·

  ·
  3

• [Fix] PTO: serialize V2C LEFT_RIGHT pipe push when per-AIV rows are not 32B-aligned

  #1815·zwh1025 opened Sep 20, 2026Contributor·

  ·
  1

  5

• ci: add generic multi-device test sharding

  #1812·Fanerly opened Sep 18, 2026Contributor·

  ·
  7

• [Example] Add gqa operator

  #1808·JasonLee-03 opened Sep 16, 2026Contributor·

  ·
  9

• [Fix] Ascend copy: fold leading dims of a >2D GM<->UB region into the DMA row count (#1231)

  #1805·2Sheldonlaw opened Sep 15, 2026Contributor·

  ·
  1

  22

• [Feat] Migrate add Chunk Scaled Dot KKT operator

  #1803·wch0810 opened Sep 15, 2026Contributor·

  ·
  3

• [Feat] Migrate add chunk_o_bwd operator for Ascend NPU

  #1802·wch0810 opened Sep 15, 2026Contributor·

  ·
  1

• [Docs] Add Chinese version of CrossEntropy benchmark summary

  #1801·zwj-sep opened Sep 15, 2026Contributor·

  ·
  1

• [Feat] Migrate add Chunk Delta H operator

  #1799·wch0810 opened Sep 15, 2026Contributor·

  ·
  7

• [Feat] Migrate add Chunk Delta Backward operator

  #1798·wch0810 opened Sep 15, 2026Contributor·

  ·
  3

• 创建 example_roi_align.py

  #1796·jluis666 opened Sep 14, 2026·

  ·
  1

• [Fix] copy_ub_to_gm cannot express compact sub-32B-row sources, corrupting keepdim reduce outputs (#1682)

  #1795·2Sheldonlaw opened Sep 14, 2026Contributor·

  ·
  1

  11

• [example] Add RMSNorm forward improved implementation, achieving a 1.10x speedup over torch_npu

  #1791·VisibleCloudsDeep opened Sep 11, 2026Contributor·

  ·
  2

• [Testing] Align example precision checks with the standard

  #1790·Lt170409 opened Sep 11, 2026·

  ·
  1

• [Example] Add lightning_indexer improved implementation (online TopK, 64.5x)

  #1787·aynewsd opened Sep 10, 2026Contributor·

  ·
  7

• [Fix] Reject non flat-contiguous BufferRegions in T.tile binary ops(#1680)

  #1786·2Sheldonlaw opened Sep 10, 2026Contributor·

  ·
  1

  2

• [Fix] Resolve multi-dim scalar BufferLoad index in T.tile binary ops (#1679)

  #1785·2Sheldonlaw opened Sep 10, 2026Contributor·

  ·
  1

  4

• [test merge]ci optimic 0421 (#902)

  #1784·z00520135 opened Sep 10, 2026Collaborator·

  ·
  1

• Precision and move pr

  #1781·Lt170409 opened Sep 9, 2026·

  ·
  1

• fix(ascend): silent wrong results when T.tile.broadcast src is loaded from GM (M≥2)

  #1777·fredrekelthen opened Sep 8, 2026Contributor·

  ·
  1

  33