To go in more detail
Right now we use PyTorch's default FP32 as default input tensor precision for the initial version of KernelBench.
Imo I think we should support more precision, especially BF16, to enable better TensorCore usage.
We might need to update the associated prompt context and tolerance as well!
To go in more detail
Right now we use PyTorch's default
FP32as default input tensor precision for the initial version of KernelBench.Imo I think we should support more precision, especially
BF16, to enable better TensorCore usage.We might need to update the associated prompt context and tolerance as well!