Ying Zhang
Ying Zhang
Did you call `target = detect_target()` before initializing `b2b_bmm`? As https://github.com/facebookincubator/AITemplate/issues/511 mentioned, there is a bug that b2b_bmm must be called after Target is defined.
Could you try it using "with" clause? ``` with detect_target() as target: xxxx ``` It's called here: https://github.com/facebookincubator/AITemplate/blob/6bbe03c7128104c8ceb0e88dea09b39429e8cb9e/python/aitemplate/backend/cuda/target_def.py#L224. You could also add some breakpoint() to debug.
Seems that ROCM backend doesn't have dtype support. You could try to modify https://github.com/facebookincubator/AITemplate/blob/6bbe03c7128104c8ceb0e88dea09b39429e8cb9e/python/aitemplate/backend/rocm/conv2d/conv2d.py#L25 to add dtype param, following https://github.com/facebookincubator/AITemplate/blob/6bbe03c7128104c8ceb0e88dea09b39429e8cb9e/python/aitemplate/backend/cuda/conv2d/conv2d.py#L25.
cc AMD folks @fsx950223 @asroy
cc AMD folks @fsx950223 @asroy
cc @thakkarV @hwu36 @ANIKET-SHIVAM
Hi @chsungen , you could set env "LOGLEVEL=DEBUG" when you use AIT to compile. This will generate a lot of debug files for each optimization step. e.g. "memory_planning_pseudo_code.txt" and "memory_planning_graph.txt"...
Thanks @fsx950223 !
cc fx2ait poc @wushirong @frank-wei to take a look.
Maybe worth checking whether the huggingface packages that you use are supported by AIT. AIT SD support hasn't been updated for a while. https://github.com/facebookincubator/AITemplate/tree/main/examples/05_stable_diffusion.