[torchlib] Implement quantize_per_channel and dequantize_per_channel

[Copilot]

This PR implements the missing quantize_per_channel and dequantize_per_channel operations in the torchlib quantized_decomposed module.

Changes

Added two new functions to onnxscript/function_libs/torch_lib/ops/quantized_decomposed.py:

quantized_decomposed_quantize_per_channel

• Implements per-channel quantization using ONNX QuantizeLinear with per-axis support
• Takes tensor inputs for scales and zero_points (one value per channel)
• Supports axis parameter to specify the quantization dimension
• Uses ONNX opset23 for per-axis quantization capabilities

quantized_decomposed_dequantize_per_channel

• Implements per-channel dequantization using ONNX DequantizeLinear with per-axis support
• Takes tensor inputs for scales and optional zero_points
• zero_points parameter is Optional[TensorType] matching PyTorch reference
• Supports both default output type and explicit output_dtype parameter

Implementation Details

Both functions:

• Follow the existing code patterns using @torch_op decorator with trace_only=True
• Have function signatures that match the PyTorch reference implementations from torch.ao.quantization.fx._decomposed
• Use ONNX opset23 to leverage the axis and output_dtype parameters for per-axis quantization
• Are properly registered and accessible as TracedOnnxFunctions

The implementation leverages ONNX's native per-axis quantization support rather than implementing the tensor manipulation logic from the PyTorch reference, making it more efficient and aligned with ONNX best practices.

Testing

Validated that:

• Functions import correctly and have proper signatures
• Function schemas are correctly generated with appropriate parameter types
• No existing functionality is broken (per-tensor functions still work)
• Functions can be used to build ONNX models

Fixes #2389.

---

💬 Share your feedback on Copilot coding agent for the chance to win a $200 gift card! Click here to start the survey.

[codecov]

Codecov Report

❌ Patch coverage is 33.33333% with 12 lines in your changes missing coverage. Please review.
✅ Project coverage is 72.64%. Comparing base (5989b56) to head (ffc4f68).
⚠️ Report is 1 commits behind head on main.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
...unction_libs/torch_lib/ops/quantized_decomposed.py	33.33%	12 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2390      +/-   ##
==========================================
- Coverage   72.66%   72.64%   -0.03%     
==========================================
  Files         259      259              
  Lines       31748    31766      +18     
  Branches     3005     3007       +2     
==========================================
+ Hits        23069    23075       +6     
- Misses       7660     7672      +12     
  Partials     1019     1019              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

[justinchuby]

Please format the code using lintrunner. You can do

pip install --upgrade lintrunner lintrunner-adapters
lintrunner init

then

lintrunner f --all-files

[Copilot]

Please format the code using lintrunner. You can do

...

Done! Code has been formatted using lintrunner. The formatting fixed import sorting, whitespace issues, and line wrapping. Commit: 82c8f9e

[titaiwangms]

How do we know this is correct?