Driss Guessous

Not sure if my issues is the same or how to repro, but on large projects (I work on pytorch) ~33,000 code files the @-reference does not work, is this...

I can repro the work arounds: 1. flex_attention = torch.compile(flex_attention, dynamic=True, mode='max-autotune') compile w/ max-autotune. Compile will take longer but you will get better performance (and we will pick a...

The problem is that the amount of shmem used is dependent on the specific score mod and masked mod used. And the available shared memory is dependent on what GPU...