Commit History

2 days ago

  1. make.py: FP4 table now lives in mxfp4

    ArthurZ HF Staff
    Verified signature
    1236014
  2. experts fp8 with 128x128 blocks (grouped kernel), attention 32x32

    ArthurZ HF Staff
    Verified signature
    3a23a8f
  3. every fp8 dim spans at least 2 blocks (grouped kernel)

    ArthurZ HF Staff
    Verified signature
    bc77947
  4. fp8 blocks 32x32 (the triton kernel needs K >= 32)

    ArthurZ HF Staff
    Verified signature
    4ec61d8
  5. tiny Mistral4 with fp8 + nvfp4 layers

    ArthurZ HF Staff
    Verified signature
    ac53c7a