Detect Hopper for FlashAttention 3 by compute capability, not GPU name (H200/H800/H20 silently fell back to FA2)

[stayinalive181]

Problem

wan/modules/attention.py enables FlashAttention 3 only when the GPU name contains h100 or hopper:

device_name = torch.cuda.get_device_name(0).lower()
return "h100" in device_name or "hopper" in device_name

torch.cuda.get_device_name() reports product names such as NVIDIA H200, NVIDIA H800 and NVIDIA H20. None of them match, so on those Hopper parts FLASH_ATTN_3_AVAILABLE is False and every attention call silently takes the FlashAttention 2 path, even with flash_attn_interface installed. Nothing is logged, so the slowdown is easy to miss.

Change

Detect Hopper by compute capability, torch.cuda.get_device_capability(0)[0] == 9, which is what the flashattn-hopper build targets (sm_90). H100 behaviour is unchanged; H200/H800/H20 now use FA3 when it is installed; Blackwell (major 10 and 12) still returns False, as the original comment intends. The docstring says why.

Verification

On an H100 80GB SXM the flag evaluates to True before and after the change (same kernels, same output). The device-name strings above are what torch.cuda.get_device_name() returns for those parts, which is what makes the old check fail there. The same name check exists in other Wan-derived repositories; we carry the identical fix in our DreamX-World deployment.

[frankjiang]

@stayinalive181 Thank you for your contribution!