All pull requests

Pull requests

• Let the tests configure alongside Highway's

  #1038·yupengtang opened Sep 21, 2026·

  ·
  3

• Move MergeOnlineSoftmax into a new per-target header, so it can be reused across

  #1034·copybara-service[bot] opened Sep 16, 2026·

  ·

• Internal change.

  #1033·copybara-service[bot] opened Sep 14, 2026·

  ·

• Reduce local KV memory in BF16 Flash attention

  #1020·Mikyx-1 opened Sep 6, 2026·

  ·
  6

• Fix tiled KV cache allocation and support prefix cloning.

  #1015·copybara-service[bot] opened Sep 4, 2026·

  ·
  1

• Add Highway SIMD logit masking kernel for constrained decoding.

  #1014·copybara-service[bot] opened Sep 4, 2026·

  ·
  1

• Add W8A8 int8 MatMul with optional rotations

  #1003·Mikyx-1 opened Aug 31, 2026·

  ·
  11

• Fix crash/abort in matmul when partition size exceeds kMaxNC.

  #958·copybara-service[bot] opened Jul 19, 2026·

  ·
  1

• Add direct HuggingFace safetensors loader for Gemma 4 (E2B/E4B)

  #919·ssfdre38 opened May 21, 2026·

  ·
  5

• Internal change

  #913·copybara-service[bot] opened May 11, 2026·

  ·

• internal changes

  #899·copybara-service[bot] opened Apr 21, 2026·

  ·

• Internal Change

  #868·copybara-service[bot] opened Mar 11, 2026·

  ·
  1

• Change to use faster exponent function

  #847·copybara-service[bot] opened Feb 26, 2026·

  ·

• Fixes to activations and tensor params

  #747·copybara-service[bot] opened Oct 28, 2025·

  ·

• [Feature] ISS-60: Implement Self Extend

  #431·jonpsy opened Oct 18, 2024·

  ·
  31

• Add configurables for norm/rope/activation/scale/residual connection.

  #287·copybara-service[bot] opened Jul 3, 2024·

  ·