All pull requests

Pull requests

• Normalize backward kernel indices

  cla signedmeta-exported
  #6737·q10 opened Nov 5, 2026Contributor·

  ·
  1

• Cap the float32 jagged add grid and test the actual launch boundary

  cla signedmeta-exported
  #6736·q10 opened Nov 5, 2026Contributor·

  ·
  1

• Cap batched ad-index broadcast grid

  cla signedmeta-exported
  #6735·q10 opened Nov 5, 2026Contributor·

  ·
  1

• Check ID_COUNT per count-bound table (#6734)

  cla signedmeta-exported
  #6734·vinares opened Nov 5, 2026·

  ·
  1

• Add ITERATION and MEM_UTIL thresholds to KVZCHTBEConfig (#6733)

  cla signedmeta-exported
  #6733·vinares opened Nov 5, 2026·

  ·
  1

• Coordinate training eviction across ranks (#6732)

  cla signedmeta-exported
  #6732·vinares opened Nov 5, 2026·

  ·
  1

• Expose the coordinated eviction backend contract (#6731)

  cla signedmeta-exported
  #6731·vinares opened Nov 4, 2026·

  ·
  1

• Centralize eviction trigger-mode decisions (#6730)

  cla signedmeta-exported
  #6730·vinares opened Nov 4, 2026·

  ·
  1

• Make n-bit lane bounds sign-safe

  cla signedmeta-exported
  #6728·q10 opened Nov 4, 2026Contributor·

  ·
  1

• Normalize quantized lookup indices

  cla signedmeta-exported
  #6727·q10 opened Nov 4, 2026Contributor·

  ·
  1

• Avoid saving unused registers in SparseAdagrad JIT kernels

  cla signed
  #6725·cyyever opened Nov 4, 2026Contributor·

  ·
  1

• [ROCm] Source ROCm/Python CI versions from test-infra matrix

  ciflow/binariesciflow/rocmcla signedmodule: rocm
  #6724·akashveramd opened Nov 4, 2026Collaborator·

  ·
  1

• Generate shared FP16 and FP32 GEMM kernels with asmjit

  cla signed
  #6723·cyyever opened Nov 4, 2026Contributor·

  ·
  1

• Cover AVX512-YMM dispatch in FP16 and FP32 GEMM tests

  cla signed
  #6722·cyyever opened Nov 4, 2026Contributor·

  ·
  1

• Fix BF16 rounding and NaN preservation in embedding JIT

  cla signed
  #6720·cyyever opened Nov 4, 2026Contributor·

  ·
  1

• Fix BF16 scalar and x86 conversion rounding and NaNs

  cla signed
  #6719·cyyever opened Nov 4, 2026Contributor·

  ·
  1

• Checkpoint packed_storage tables as raw g8 rows, load across raw formats, stage g8 FP16 publish views (#6712)

  cla signedmeta-exported
  #6712·EddyLXJ opened Nov 3, 2026Contributor·

  ·
  1

• Add default-off owner_cache_exclusive: the owner-local INT8 HBM cache holds the only copy of its rows (#6709)

  cla signedmeta-exported
  #6709·EddyLXJ opened Nov 3, 2026Contributor·

  ·
  1

• Optimize INT4 embedding accumulation and eliminate intermediate buffers

  cla signed
  #6706·gaborkertesz opened Nov 3, 2026Contributor·

  ·
  1

• Turn the HBM lane off on TBEs without an allowlisted DEVICE table

  cla signedmeta-exported
  #6703·FriedCosey opened Nov 3, 2026·

  ·
  1

• Fuse MX4 qcomm codec with neighboring pooled all-to-all ops

  cla signedmeta-exported
  #6701·santoshmo opened Nov 2, 2026·

  ·
  1

• mx4 quantize: cut per-element integer work, bit-exact (#6700)

  cla signedmeta-exported
  #6700·santoshmo opened Nov 2, 2026·

  ·
  1

• Enable C++ Autodeps2 for xplat/deeplearning

  cla signedmeta-exported
  #6699·kkolur76 opened Nov 2, 2026Contributor·

  ·

• Opt out CPU ops on ROCm

  ciflow/rocmcla signedmodule: rocm
  #6698·avbokovoy opened Nov 2, 2026Contributor·

  ·
  1

• Use hardware wave width in ROCm VBE routing tests

  ciflow/rocmcla signedmeta-exportedmodule: rocm
  #6687·q10 opened Nov 1, 2026Contributor·

  ·
  2