Ning

Summary: Repeat title. The long term solution could be to add an option to distinguish intermediate eval. But confirmed with Juan setting it to False shouldn't affect existing stuff. Differential...

Flow side support for separate decoders / Fix max_iter at inference

1

Summary: 1. Add flow side support for separate decoder configs 2. Fix #iterations at inference. Will add adaptive decoding support in a separate diff. Reviewed By: kahne Differential Revision: D18165098

Add validation loss to metric reportor and Remove intermediate eval

1

Differential Revision: D19679308

Summary: Now --target-text-file is mandatory when --source-text-file is specified. However for getting translation without evaluation it's not necessary. Remove the validation. LMK if there are potential concerns. Differential Revision: D15540664

NAT bug fix

1

Summary: Remove a duplicate operation. Line 1838 is doing the same thing. Earlier max_iter > 1 inference is broken internally. This fixes it, though we don't observe much performance gain...

[WIP, don't merge] unity.cpp -> ggml master

2

This is a WIP PR to sync unity.cpp from seamless_communication (https://github.com/facebookresearch/seamless_communication/tree/main/ggml) to examples/ under ggml. Sharing for visibility and looking for early feedback from community & authors. Feel free to...

Asserting over nb[0] as type check causes issue for tensors after permuting

1

Is the intention of asserts like https://github.com/ggerganov/ggml/blob/master/src/ggml.c#L8772 to type check? I have a fp32 tensor as a result of permutation function, so nb[0] is no longer sizeof(float) and triggers the...