Add triton dtype for torch.float8_e4m3fnuz to allow explicit casts #8164
Description
Describe the bug
Currently, triton does not have a dtype for torch.float8_e4m3fnuz:
triton/python/triton/language/core.py
Line 380 in aff4b7a
The datatype is supported by the backend and a cast like some_tensor.to(target_pointer.dtype.element_ty) is working.
However, a explicit type cast like some_tensor.to(tl.fp8e4fnuz) is not possible.
This is a problem in situations where the kernel cannot access the correct datatype via a pytorch pointer (one example: vllm-project/vllm#24503 (comment)).
I know, this might be hard to keep it compatible cross-platform, but maybe still worth considering.
Environment details
Triton 3.4.0, MI300