Fix compatibility with transformers 4.41+ and remove diffusers dependency

[nekoscratch01]

I met the Problem when I tried to integrated this model: Version Conflict Loop

Current code cannot work with modern transformers versions due to a dependency conflict:

1. Code uses old transformers APIs (seen_tokens, get_usable_length) → requires transformers < 4.41
2. diffusers dependency requires transformers >= 4.43
3. Result: No version of transformers satisfies both requirements

Errors encountered:

• With transformers >= 4.41: AttributeError: 'DynamicCache' object has no attribute 'seen_tokens'
• With transformers < 4.41: ImportError: cannot import name 'EncoderDecoderCache' from diffusers

The Solution

1. Update to new transformers API

File: modeling_bailing_moe.py

• Replace Cache.get_usable_length(seq, layer) → Cache.get_seq_length()
• Replace Cache.seen_tokens → cache_length

This makes code compatible with transformers 4.41+

2. Remove diffusers dependency

Files: audio_tokenizer/oobleck_distribution.py (new), audio_tokenizer/modeling_audio_vae.py

• Extract OobleckDiagonalGaussianDistribution class (~50 lines) from diffusers
• Only this one class was needed, but it pulled in entire diffusers library
• Eliminates the version conflict entirely

Result

• Works with transformers 4.55.0
• No dependency conflicts
• Smaller dependency footprint

[yongjie-lv]

Thanks for your interest!
On our end, the environment is transformers==4.52.4 and diffusers==0.33.0, and the entire inference process is working correctly. You could try referring to our requirements.txt file, or using the recently published Docker image and Dockerfile. Please give it another try with one of those.
Thanks again for reaching out. Feel free to ask if you have any more questions!

[nekoscratch01]

Thanks for your response! I’m not sure why it doesn’t work on my end. Currently I am working with a project that tried to integrate Ming into our framework, and I suspected the issue might be related to the latest transformers version.
This PR actually resolves that dependency loop nicely and ensures compatibility with newer transformers releases. Thanks for checking it out!