Fix target mask at inference time

[lucasnewman]

I'm still working on getting the model to sample correctly, but I noticed a couple of issues while debugging:

1) The passed in conditioning + target duration was ignored, so the model never generated outside of the conditioning sequence without the duration predictor in place. 2) The target mask needs to exclude the conditioning, like what's done at training time with the random span mask.

I also made a fast path when CFG is disabled so we don't compute and throw away the prediction with no conditioning.

[lucidrains]

@lucasnewman thanks for catching the duration bug! :bug:

the CFG change also makes sense, forgot about that :facepalm:

so the cond_mask refers to the prefix conditioning length, while the mask is the self attention mask that should encompass both the condition as well as the target sequence that follows (cond sequence length + target length == duration length)

the way you have it the transformer would not be able to attend to the condition

[lucasnewman]

Ah, ok, my mistake! You can close this if you want!

[lucidrains]

@lucasnewman the other two changes look good! let me merge those in :pray: