training / experiments /gemma_decoder /debug_backend.py

Commit History

2 days ago

  1. debug_backend: len-mismatch fix (eos slicing)

    agent
    b052374
  2. debug_backend: operator-requested cuDNN SDPA NaN diagnostics (GPU id, seq%128, zero-token repro with cudnn sdp disabled, both-backend grad finiteness)

    agent
    bb51fdf