Matrix-Game-3: --fa_version 0 (SDPA) is not implemented; attention() asserts a flash-attn install instead of falling back #86

Matrix-Game-3/wan/modules/attention.py no longer has the SDPA fallback that the original Wan attention module ships. The non-FA3 branch of flash_attention() is:

else:
    assert FLASH_ATTN_2_AVAILABLE or FLASH_ATTN_3_AVAILABLE

Two consequences:

• Without a flash-attn install, every attention call dies on that assertion, so the model cannot run at all on a machine where flash-attn does not build (the CLI's --fa_version 0 suggests SDPA is meant to be selectable).
• With flash-attn 2 installed, --fa_version 0 still runs the flash-attn 2 kernel: version=0 is neither 3 nor handled anywhere, so it falls through to the same branch. --fa_version 3 with only FA2 installed does degrade with a warning, which is the configuration that worked for us on an H100.

Suggested fix: restore the torch.nn.functional.scaled_dot_product_attention path for version == 0 (and as the fallback when neither flash-attn build is importable), with the warning the upstream Wan module prints. Happy to open a PR if that is the intended behaviour.

Related, still open as of today: the multi-GPU path passes fa_version= to attention() whose parameter is named version (#80, fix in #61), so --ulysses_size > 1 raises TypeError on a clean checkout.