WARNING 12-19 11:47:57 [_attention_dispatch.py:454] Re-registered SAGE attention backend to enable context parallelism with FP8 Attention in cache-dit. You can disable this behavior by: export CACHE_DIT_ENABLE_CUSTOM_ATTN_DISPATCH=0.
INFO 12-19 11:47:57 [_attention_dispatch.py:612] Flash Attention 3 not available, skipping _FLASH_3 backend registration.
[rank0]: Traceback (most recent call last):
[rank0]: File "/workspace/dev/vipshop/cache-dit/src/cache_dit/parallelism/transformers/native_diffusers/context_parallelism/__init__.py", line 22, in <module>
[rank0]: _maybe_register_custom_attn_backends()
[rank0]: File "/workspace/dev/vipshop/cache-dit/src/cache_dit/parallelism/transformers/native_diffusers/context_parallelism/attention/__init__.py", line 14, in _maybe_register_custom_attn_backends
[rank0]: from ._attention_dispatch import (
[rank0]: ImportError: cannot import name '_flash_attention_3' from 'cache_dit.parallelism.transformers.native_diffusers.context_parallelism.attention._attention_dispatch' (/workspace/dev/vipshop/cache-dit/src/cache_dit/parallelism/transformers/native_diffusers/context_parallelism/attention/_attention_dispatch.py). Did you mean: '_sage_attention'?
[rank0]: During handling of the above exception, another exception occurred:
