All pull requests

Pull requests

• Restore GPU inplace-update test coverage

  cla signedmeta-exported
  #6785·q10 opened Nov 11, 2026Contributor·

  ·
  1

• Use c10::irange for simple loops in fbgemm_gpu

  cla signedmeta-exported
  #6784·r-barnes opened Nov 10, 2026Contributor·

  ·
  1

• Keep SSD N-bit lookup dummy backing valid and preserve legacy state dictionaries

  cla signedmeta-exported
  #6783·q10 opened Nov 10, 2026Contributor·

  ·
  1

• Persist Adam optimizer states under UVM-compatible checkpoint names

  cla signedmeta-exported
  #6782·Frederick-Zhu opened Nov 10, 2026·

  ·
  1

• Zero-fill cold rows in readonly checkpoint iterators (#6781)

  cla signedmeta-exported
  #6781·Frederick-Zhu opened Nov 10, 2026·

  ·
  1

• Skip experimental CUDA TLX tests appropriately on HIP and CPU

  cla signedmeta-exported
  #6779·q10 opened Nov 10, 2026Contributor·

  ·
  1

• dram_kv_embedding_inplace_update_cpu: take each table's row width from update_offsets

  cla signedmeta-exported
  #6778·FriedCosey opened Nov 10, 2026·

  ·
  1

• Share x86 INT4 autovec dequantization to reduce binary size

  cla signedmeta-exported
  #6774·q10 opened Nov 9, 2026Contributor·

  ·
  1

• Keep backward segment loops at 32 bits

  cla signedmeta-exported
  #6772·q10 opened Nov 9, 2026Contributor·

  ·
  1

• fbgemm: warp-per-sequence block_bucketize kernels for long sequences (#6771)

  cla signedmeta-exported
  #6771·XialuLi opened Nov 9, 2026·

  ·
  1

• Add asynchronous_segmented_complete_cumsum

  cla signedmeta-exported
  #6769·danpog opened Nov 9, 2026·

  ·
  1

• Compile scaled_fp4_quant for SM100 family targets

  cla signed
  #6767·0z5a opened Nov 9, 2026Contributor·

  ·
  1

  1

• Size no-bag overflow regression from the runtime warp

  cla signedmeta-exported
  #6759·q10 opened Nov 8, 2026Contributor·

  ·
  1

• Test requested precision in large-table TBE overflow cases

  cla signedmeta-exported
  #6758·q10 opened Nov 8, 2026Contributor·

  ·
  1

• Use the native dense BMM reference in jagged GPU tests

  cla signedmeta-exported
  #6757·q10 opened Nov 8, 2026Contributor·

  ·
  1

• Apply Blackwell attention architecture skips to the test class

  cla signedmeta-exported
  #6756·q10 opened Nov 8, 2026Contributor·

  ·
  1

• Cap jagged add output grid

  cla signedmeta-exported
  #6755·q10 opened Nov 7, 2026Contributor·

  ·
  1

• Add gated SSD KV lookup optimizations and lookup statistics (#6754)

  cla signedmeta-exported
  #6754·EddyLXJ opened Nov 7, 2026Contributor·

  ·
  1

• Owner-local HBM caches in the DRAM storage format: 16-way set-associative LRU slots and HBM-first trainer row writes (#6746)

  cla signedmeta-exported
  #6746·EddyLXJ opened Nov 6, 2026Contributor·

  ·
  1

• Bind the gfx90a NFP8 device alias to fnuz

  cla signed
  #6745·aryaman-gupta opened Nov 6, 2026Contributor·

  ·
  1

• Share compact SSD storage across pinned snapshots and fence delta versions (#6744)

  cla signedmeta-exported
  #6744·EddyLXJ opened Nov 6, 2026Contributor·

  ·
  1

• Owner-cache hit metrics and KVZCH_ROW_WRITE_DIRECT_HBM: trainer-written rows enter the owner-local HBM cache at write time (#6743)

  cla signedmeta-exported
  #6743·EddyLXJ opened Nov 6, 2026Contributor·

  ·
  1

• VBE metadata: optional row_output_offsets / b_t_map via compute_row_metadata

  cla signedmeta-exported
  #6740·gchalump opened Nov 5, 2026Contributor·

  ·
  1

• Exclusive HBM checkpoint/promote QPS fixes and the mb knobs for INT8/g8 DRAM rows and the exclusive owner cache (#6739)

  cla signedmeta-exported
  #6739·EddyLXJ opened Nov 5, 2026Contributor·

  ·
  1

• INT8 and g8 DRAM rows end to end: int8_storage, raw INT8/g8 checkpoints, raw HBM moves, FP16 publish views, fill-only writes (#6738)

  cla signedmeta-exported
  #6738·EddyLXJ opened Nov 5, 2026Contributor·

  ·