Transformers should be pinned to 4.46.0? ImportError: cannot import name 'Gemma2FlashAttention2' from 'transformers.models.gemma2.modeling_gemma2'

[josephrocca]

I think the transformers version should be pinned like transformers==4.46.0. Currently if I install, and then run:

python3 -m deepcompressor.app.diffusion.ptq configs/model/flux.1-schnell.yaml --output-dirname reference

I get:

Traceback (most recent call last):
  File "<frozen runpy>", line 198, in _run_module_as_main
  File "<frozen runpy>", line 88, in _run_code
  File "/opt/conda/lib/python3.11/site-packages/deepcompressor/app/diffusion/ptq.py", line 10, in <module>
    from deepcompressor.app.llm.nn.patch import patch_attention, patch_gemma_rms_norm
  File "/opt/conda/lib/python3.11/site-packages/deepcompressor/app/llm/nn/__init__.py", line 3, in <module>
    from .struct import LlmModelStruct, LlmTransformerBlockStruct, LlmTransformerStruct
  File "/opt/conda/lib/python3.11/site-packages/deepcompressor/app/llm/nn/struct.py", line 10, in <module>
    from transformers.models.gemma2.modeling_gemma2 import (
ImportError: cannot import name 'Gemma2FlashAttention2' from 'transformers.models.gemma2.modeling_gemma2' (/opt/conda/lib/python3.11/site-packages/transformers/models/gemma2/modeling_gemma2.py)

But if I run:

pip3 install transformers==4.46.0

Then that fixes it. I've tested this on a Runpod H100, and my local RTX 4090, and it's the same.

I'm following this exactly, but the way: https://github.com/mit-han-lab/deepcompressor/tree/main/examples/diffusion Trying reproduce Schnell quantization.

However, after fixing the above issue, I then run into this issue:

• Getting Error when converting flux.dev with the latest version of deepcompressor #50 (comment)