[BlockInfo] Index to Tensor

[Haoming02]

• Convert the block_index from int to a torch.Tensor, to support torch.compile
  • dtype: torch.uint8
  • device: input's device cpu

[Kosinkadink]

@kijai could you check if this fixes the torch.compile graph breaks?

[kijai]

Yes, tested on HunyuanVideo 1.5 and went from 54 recompiles on first step to none.

[Kosinkadink]

Sweet, I'll merge it in after stable!

[Kosinkadink]

@Haoming02 could you make the tensors be on the CPU instead?

[Kosinkadink]

@kijai hey, could you retest this to make sure that the CPU tensors work fine with torch.compile?

[kijai]

@kijai hey, could you retest this to make sure that the CPU tensors work fine with torch.compile?

While creating it on cpu is fine, actually using a cpu tensor is a bit problematic as with inductor it requires cpu compile support, which then requires more compiler libraries installed than at least the current Triton-windows package includes, and if you don't have them it just errors out.

You can of course cast it to gpu before using it, it does then create DeviceCopy in input program warning, but I'm unsure what this affects.

[Kosinkadink]

damn, that is definitely annoying. the goal would be that the block index is used inside attention code and is only compared. To even do this comparison, would it be better to compare against the GPU tensor, or CPU tensor? if GPU tensor would be easier, then we can edit this PR to use GPU.

Alternatively, is there a way to tell torch compile to ignore transformer_options or at least that one key?

[woct0rdho]

I'm also waiting for blockinfo to be merged so I can implement things like RadialAttn more easily. I think it's ok to create a scalar tensor on GPU and do comparison with it. It's not the first time we do it, for example in https://github.com/comfyanonymous/ComfyUI/blob/c5a47a16924e1be96241553a1448b298e57e50a1/comfy/extra_samplers/uni_pc.py#L785

[Kosinkadink]

Comfy merged a PR to make things in transformer_options not cause as many graph breaks, so we can stick with integers: #11317

I'll be closing this issue, and once the tensors get turned into ints in the Lumina BlockInfo PR, i can merge that one! #11227