All pull requests

Pull requests

• Require strongly typed inputs in the public CPU FP8 embedding API

  cla signed
  #6512·cyyever opened Oct 18, 2026Contributor·

  ·

• Remove unused FP8 rowwise scratch allocation

  cla signed
  #6511·cyyever opened Oct 18, 2026Contributor·

  ·

• Remove unused FP8 rowwise host template parameter

  cla signed
  #6510·cyyever opened Oct 18, 2026Contributor·

  ·

• Specialize GPU N-bit quantization kernels for common bit widths

  cla signed
  #6509·cyyever opened Oct 18, 2026Contributor·

  ·

• Vectorize N-bit dequantization with front scale and bias on AVX2

  cla signed
  #6507·cyyever opened Oct 18, 2026Contributor·

  ·

• Enable constexpr convolution and scalar quantization helpers (#6485)

  cla signedmeta-exported
  #6504·q10 opened Oct 17, 2026Contributor·

  ·
  1

• Add laser_key_suffix to EnrichmentPolicy for versioned Laser tiers

  cla signedmeta-exported
  #6499·wilsonyqm-debug opened Oct 17, 2026·

  ·

• Add NEON kernels for fp32 <-> fp16 conversion

  cla signed
  #6498·sremes opened Oct 17, 2026·

  ·
  1

• Validate CK FP8 inputs using the native GPU dtype

  cla signedmeta-exported
  #6497·q10 opened Oct 17, 2026Contributor·

  ·
  1

• Add MI350/MI450 TBE and GIS benchmark harness

  cla signedmeta-exported
  #6496·q10 opened Oct 17, 2026Contributor·

  ·
  1

• Add owner-local direct enrichment cache path (#6495)

  cla signedmeta-exported
  #6495·EddyLXJ opened Oct 16, 2026Contributor·

  ·
  1

• Optimize DRAM enrichment reads and direct cache kernels (#6494)

  cla signedmeta-exported
  #6494·EddyLXJ opened Oct 16, 2026Contributor·

  ·
  1

• Frequency-adaptive learning rate for rowwise_rmsprop_ar: TBE frontend

  cla signedmeta-exported
  #6492·hdmeta opened Oct 16, 2026Contributor·

  ·
  1

• Remove half-type token-pasting macros from FP8 instantiation

  cla signed
  #6488·cyyever opened Oct 16, 2026Contributor·

  ·
  1

• Grid-stride FP8 fixed-NK group setup (#6481)

  cla signedmeta-exported
  #6481·q10 opened Oct 16, 2026Contributor·

  ·
  1

• Replace legacy null CUDA stream constants

  cla signedmeta-exported
  #6479·q10 opened Oct 16, 2026Contributor·

  ·
  1

• Move the ROCm CI to ROCm 7.14 and 10.0 (#6474)

  ciflow/rocmcla signedmeta-exportedmodule: rocm
  #6474·q10 opened Oct 15, 2026Contributor·

  ·
  2

• Avoid eager empty tensors in INT-N-bit inference dispatch (#6473)

  cla signedmeta-exported
  #6473·mrmiywj opened Oct 15, 2026Contributor·

  ·
  1

• Re-enable TBE backward none and SGD writeback tests on ROCm (#6472)

  ciflow/rocmcla signedmeta-exportedmodule: rocm
  #6472·q10 opened Oct 15, 2026Contributor·

  ·
  2

• Suppress type errors for Pyre upgrade (landing group 36)

  cla signedmeta-exported
  #6471·maggiemoss opened Oct 15, 2026Contributor·

  ·
  1

• Migrate codegen inference mode to constraints

  cla signedmeta-exported
  #6469·scottliu77 opened Oct 14, 2026·

  ·

• Detect gfx1250 and select TheRock CI mode

  cla signedmeta-exported
  #6463·q10 opened Oct 14, 2026Contributor·

  ·
  1

• Remove shadowing from standalone GPU operators

  cla signedmeta-exported
  #6457·q10 opened Oct 13, 2026Contributor·

  ·
  1

• Fix TBE forward using stale uvm_cache_stats

  cla signed
  #6454·avbokovoy opened Oct 13, 2026Contributor·

  ·
  1

• Re-enable TBE backward none and SGD writeback tests on ROCm

  ciflow/rocmcla signedmodule: rocm
  #6451·avbokovoy opened Oct 13, 2026Contributor·

  ·
  2