[Issue]: arch_supports_fp8() for gfx950 and gfx1201?

[kasper93]

Problem Description

We can read here that FP8 is enabled only for gfx942. According to my understanding gfx950 and gfx1201 should also support it on architecture level.

flash-attention/flash_attn/flash_attn_triton_amd/utils.py

Lines 774 to 776 in ea8fe36

	@functools.cache
	def arch_supports_fp8():
	return is_hip() and get_arch() in ('gfx942')

Maybe there are other blockers that has to be resolved before enabling FP8 for those platforms in flash-attention. You can consider this issue as bug report if this is as simple as updating this check or feature request if more changes are needed, but eventually I think FP8 can be enabled for gfx950 and gfx1201.