• Thomas O'Malley's avatar
    Reduce Layer.__call__ overhead by ~20% · a33f9b44
    Thomas O'Malley authored
    This is achieved by improving the way masks are handled for inputs and outputs.
    For the common case where masks are not input and are not output, minimal work
    is done now.
    For the masking case, the work done is about the same.
    
    PiperOrigin-RevId: 312871996
    Change-Id: I2e122551bec27d075193e1881bf236d570d25ce4
    a33f9b44
sequential.py 22.4 KB