All pull requests

Pull requests

• Fix unaligned Vec4 half access through local temporary arrays

  cla signed
  #6653·cyyever opened Oct 29, 2026Contributor·

  ·

• Per-table HBM drain interval

  cla signedmeta-exported
  #6652·FriedCosey opened Oct 28, 2026·

  ·
  1

• Fix TBE cache signed comparisons (#6650)

  cla signedmeta-exported
  #6650·q10 opened Oct 28, 2026Contributor·

  ·
  1

• [ROCm][CI] Move ROCm CI window to 10.0 and 10.1

  ciflow/rocmcla signedmodule: rocm
  #6649·akashveramd opened Oct 28, 2026Collaborator·

  ·
  1

  

• Fix host operator signed comparisons

  cla signedmeta-exported
  #6647·q10 opened Oct 28, 2026Contributor·

  ·
  1

• Replace the embedding kernel signature wrapper with an alias (#6646)

  cla signedmeta-exported
  #6646·q10 opened Oct 28, 2026Contributor·

  ·
  1

• Consolidate CPU reference conversions with deduced input types (#6635)

  cla signedmeta-exported
  #6645·q10 opened Oct 28, 2026Contributor·

  ·
  1

• Replace sparse bucketize launch macros with template lambdas (#6639)

  cla signedmeta-exported
  #6643·q10 opened Oct 28, 2026Contributor·

  ·
  1

• Consolidate CPU reference conversions with deduced input types

  cla signed
  #6635·cyyever opened Oct 28, 2026Contributor·

  ·
  1

• Keep allgather worker startup outside generated case deadlines

  cla signedmeta-exported
  #6634·q10 opened Oct 28, 2026Contributor·

  ·
  1

• Accumulate optimizer-none GPU test reference in FP32

  cla signedmeta-exported
  #6631·q10 opened Oct 27, 2026Contributor·

  ·
  1

• Use 256-thread blocks for the jagged dense-output kernel on CUDA (#6629)

  cla signedmeta-exported
  #6629·pdesupinski opened Oct 27, 2026·

  ·
  1

• Restore two existing nbit cache FakeTensor checks

  cla signedmeta-exported
  #6628·q10 opened Oct 27, 2026Contributor·

  ·
  1

• Enable portable Triton ROSS request-only LCE

  cla signedmeta-exported
  #6626·Flippedyue opened Oct 27, 2026·

  ·

• Enable request-only TLX LCE

  cla signedmeta-exported
  #6625·Flippedyue opened Oct 27, 2026·

  ·

• Restore whole-row checkpoints by embedded IDs (#6624)

  cla signedmeta-exported
  #6624·q10 opened Oct 27, 2026Contributor·

  ·
  1

• [ROCm][CI] Prevent ROCm CI on main from canceling prior merges

  ciflow/rocmcla signedmodule: rocm
  #6619·akashveramd opened Oct 27, 2026Collaborator·

  ·
  2

  

• KVEmbeddingInference: serve group-wise INT4 (e4m3 scales) rows (#6618)

  cla signedmeta-exported
  #6618·wilsonyqm-debug opened Oct 27, 2026·

  ·
  1

• serving_qdq_group_size: store enrichment cache rows as their group-wise INT4 serving values (#6617)

  cla signedmeta-exported
  #6617·wilsonyqm-debug opened Oct 27, 2026·

  ·
  1

• Group-wise INT4 with float8_e4m3fn block scales: quantize/dequantize ops (#6616)

  cla signedmeta-exported
  #6616·wilsonyqm-debug opened Oct 27, 2026·

  ·
  1

• Store enrichment cache rows as their INT4 serving values (serving_qdq_bit_rate) (#6615)

  cla signedmeta-exported
  #6615·wilsonyqm-debug opened Oct 27, 2026·

  ·
  1

• Keep NaN shared-memory initialization out of FakeTensor kernels

  cla signedmeta-exported
  #6611·q10 opened Oct 26, 2026Contributor·

  ·
  1

• Register fake implementations for nbit cache lookups

  cla signedmeta-exported
  #6610·q10 opened Oct 26, 2026Contributor·

  ·
  1

• Restore nbit CPU/GPU parity operator checks

  cla signedmeta-exported
  #6609·q10 opened Oct 26, 2026Contributor·

  ·
  1

• Support the SGEMM example on HIP

  cla signedmeta-exported
  #6608·q10 opened Oct 26, 2026Contributor·

  ·
  1