fix(pytorch): Rename layer_scale parameter to avoid quantization error

[ved1beta]

Here's the formatted PR description based on our changes:

Type of Change

• Bug fix
• No API changes (parameter rename only)

Description

This PR fixes an issue where PyTorch models using layer_scale parameters fail during quantization due to tensor conversion errors. The problem occurs when the quantization process attempts to convert tensor-type scale parameters to Python scalars.

Key changes:

1. Renamed problematic layer_scale parameter to layer_gamma to avoid conflicts with scale detection
2. Enhanced _get_module_scale_zeropoint method to better handle scale parameters
3. Added improved logging for scale parameter detection

The fix maintains backward compatibility while resolving the quantization failure.

Expected Behavior & Potential Risk

Expected Behavior:

• Models using the renamed layer_gamma parameter will successfully complete quantization
• Existing scale detection logic works correctly with the new parameter name
• Debug logs provide better visibility into scale parameter detection

Potential Risks:

• Models explicitly looking for layer_scale parameter name might need updates
• Need to communicate the parameter rename in documentation
• May need to provide migration guide for existing models

How has this PR been tested?

Tests have been implemented in test/adaptor/test_pytorch_layer_scale.py with two test cases:

1. test_layer_scale_error: Verifies the original issue by confirming that models with layer_scale parameter fail quantization with the expected error
2. test_layer_gamma_success: Validates that models using the new layer_gamma parameter successfully complete quantization

Test Environment:

• Python 3
• PyTorch latest version
• Neural Compressor latest version
• Test models: ConvEncoder architecture with both original and fixed parameter names
• Input tensor shape: (1, 64, 32, 32)

To reproduce:

cd neural-compressor
python3 test/adaptor/test_pytorch_layer_scale.py

Dependency Change?

No new dependencies were introduced or removed.

• Uses existing PyTorch and Neural Compressor infrastructure
• All changes are internal to the codebase

[Copilot]

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

Comments suppressed due to low confidence (1)

neural_compressor/adaptor/pytorch.py:4175

• The current filter condition excludes any parameter containing 'gamma', which might unintentionally filter out valid parameters. Consider checking explicitly for 'layer_scale' and 'layer_gamma' instead to avoid potential misclassification.

if "scale" in node.target and not any(exclude in node.target for exclude in ["layer_scale", "gamma"]):

[xin3he]

Hi @ved1beta Thanks for your commit.
I notice that the unexpected case is that scale is a tensor list, not one value. If so, we can fix it in a better way.

Suggestion:
float(getattr(model, node.target)) -> getattr(model, node.target).tolist() # it can cover both one value and list value

BTW, currently our CI is blocked by #2171. To resolve the CI issue, please update your branch after #2171 is merged.

[XuehaoSun]

Please fix CI issues and sign DCO
https://dev.azure.com/lpot-inc/neural-compressor/_build/results?buildId=39554&view=logs&j=e5896b99-a49d-517b-218b-3b918f0c116d&t=b18c099a-26ee-5571-9980-67a803d9b7da&l=29800

[ved1beta]

hey i did required changes but had a merge conflict , fixing it

[xin3he]

Thanks, @ved1beta, merged