[NO CP] [Inductor] Use torch.version.hip to conditionalise out dynamic rblock scaling

ROCm / pytorch

Tensors and Dynamic neural networks in Python with strong GPU acceleration

http://pytorch.org

Other

219 stars 50 forks source link

[NO CP] [Inductor] Use torch.version.hip to conditionalise out dynamic rblock scaling #1448

Closed jataylo closed 1 week ago

jataylo commented 2 weeks ago

The is_hip logic is flakey on 6.2_internal_testing, bringing in this change to avoid unexpected issues due to device_prop.regs_per_multiprocessor not being available on ROCm in this branch. This will be adopted upstream and cherry picked into 6.3 via https://github.com/pytorch/pytorch/pull/129663

This issue was raised in trying to reproduce a defect in PLAT-160450 and will help the investigation into this issue.

jithunnair-amd commented 1 week ago

From offline discussion with @jataylo: Planning to not introduce the device_prop.regs_per_multiprocessor property in rocm6.2_internal_testing so as to not enable a whole new untested piece of logic this late in the ROCm6.2 release cycle. It's safer to just disable the logic properly using torch.version.hip