Skip to content

CHANGELOG.md

Latest commit

 

History

History
31 lines (20 loc) · 1.76 KB

CHANGELOG.md

File metadata and controls

31 lines (20 loc) · 1.76 KB

Changelog - sym-bot/diffusers Fork

This changelog documents modifications made to the sym-bot/diffusers fork compared to the upstream huggingface/diffusers.

[Unreleased]

Fixed

  • Graceful fallback for attention backend import failures (src/diffusers/models/attention_dispatch.py)

    Added try-except wrappers around all external attention backend imports to handle ABI mismatches and import failures gracefully. Affected backends:

    • flash_attn (Flash Attention)
    • flash_attn_3 (Flash Attention 3)
    • aiter (AMD Instinct)
    • sageattention (SageAttention)
    • flex_attention (PyTorch Flex Attention)
    • torch_npu (Huawei NPU)
    • torch_xla (TPU/XLA)
    • xformers (Meta xFormers)

    Problem: When an external attention package is installed but compiled against a different PyTorch version, importing it causes an OSError with "undefined symbol" errors. The previous code used importlib.util.find_spec() to check package availability, which only checks if the package exists but not if it can actually be imported.

    Solution: Wrap actual imports in try-except blocks. On failure, log a warning and set _CAN_USE_* flags to False, allowing the code to fall back to native PyTorch SDPA (Scaled Dot-Product Attention) which requires no external dependencies.

    Impact: Diffusers pipelines now work correctly even when external attention packages have ABI mismatches. Users see a warning instead of a crash, and inference continues using PyTorch's built-in attention implementation.

Upstream Sync

This fork is periodically synced with upstream huggingface/diffusers. The modifications above are maintained as patches on top of the upstream codebase.