jhcross

test scriptify MultiheadAttention

3

Summary: We introduce a new test suite where we will iteratively ensure that various model components are TorchScript compliant (JIT-able). Differential Revision: D18799715

more general state init/passing

1

Summary: A more general mechanism for passing states during incremental decoding, which in particular makes requirements for ONNX export more explicit. Differential Revision: D9599067

hybrid: enable no bottleneck

1

Differential Revision: D15507269

cmd-line support for loading mmap datasets

2

Summary: Allows loading data from the Fairseq .idx/.bin format (including most current "mmap" implementation) by specifying the `--fairseq-binary-data-format` flag. (Note that D16867809 added iniernal support for loading legacy .idx /...

return empty tensor instead of None

1

Summary: To allow efficient use of fork/join annotation, we return an empty tensor instead of `None` for `encoder_padding_mask` from transformer encoder in the unmasked/inference case. Note that this slight hack...