Bug Description

After PR #974 (fd92d96) bumped torchao from >=0.14.1,<0.16.0 to >=0.16.0,<0.17.0, model initialization fails with:

RuntimeError: Failed to import diffusers.models.autoencoders.autoencoder_oobleck because of the following error:
name 'logger' is not defined

Root Cause

1. torchao 0.16.0 removed the torchao.dtypes.uintx.uint4_layout module (which contained UInt4Tensor)
2. diffusers 0.36.0 (the currently resolved version, since diffusers is unpinned in pyproject.toml) imports this removed module at module level in diffusers/quantizers/torchao/torchao_quantizer.py:89
3. The ImportError handler at line 93 calls logger.warning(), but logger is not defined until line 126 — a bug in diffusers 0.36.0
4. This causes a NameError that propagates up and crashes VAE model loading

diffusers 0.37.0+ fixed this by:

• Moving logger definition before the function call
• Adding a version gate: if is_torchao_version(">", "0.15.0"): pass to skip the UInt4Tensor import entirely for torchao 0.16.0+

Reference: huggingface/diffusers#11018

Impact

• Gradio server starts but model initialization fails — users see the UI but cannot generate music
• The LLM handler initializes fine, but the DiT handler (which needs AutoencoderOobleck from diffusers) crashes

Proposed Fix

Pin diffusers>=0.37.0 in pyproject.toml to ensure compatibility with torchao>=0.16.0.

diffusers was previously unpinned ("diffusers" with no version constraint), which allowed uv to resolve to 0.36.0 — a version incompatible with torchao 0.16.0.

Reproduction

# With torchao>=0.16.0 and diffusers<0.37.0 installed:
python -c "from diffusers.models import AutoencoderOobleck"
# → NameError: name 'logger' is not defined

Environment

• torch: 2.10.0+cu128
• torchao: 0.16.0
• diffusers: 0.36.0
• GPU: NVIDIA H200