How does this work with ONNX export and quantization?

[ogencoglu]

Does quantized models here become quantized models in ONNX after conversion? Can you even convert/export them to ONNX? How about other way around? Can you export a sparse model to ONNX and quantize in ONNX afterwards?

[msaroufim]

We haven't really experimented much with ONNX so far. Though we do support export and once you export a model you can use an ONNX backend

1. Step1: Export an AO model https://github.com/pytorch/ao/tree/main/torchao/quantization#workaround-with-unwrap_tensor_subclass-for-export-aoti-and-torchcompile-pytorch-24-and-before-only
2. Step2: Use the ONNX backend https://pytorch.org/tutorials/beginner/onnx/export_simple_model_to_onnx_tutorial.html

If you wanna work through an example and post your progress here, happy to unblock you! We can add some example in the repo

[hahnec]

Using the procedure outlined above, I successfully got a simple quantization model exported. However, when applying the same approach to a custom, nested, large model, torch.export.export() throws the error shown below. To bypass the unwrap_tensor_subclass issue, I upgraded to Torch 2.5.1. Any guidance or suggestions to resolve this would be greatly appreciated.

Traceback (most recent call last): File "<string>", line 1, in <module> File "<torch_path>/torch/export/__init__.py", line 270, in export return _export( ^^^^^^^^ File "<torch_path>/torch/export/_trace.py", line 1017, in wrapper raise e File "<torch_path>/torch/export/_trace.py", line 990, in wrapper ep = fn(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^ File "<torch_path>/torch/export/exported_program.py", line 114, in wrapper return fn(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^ File "<torch_path>/torch/export/_trace.py", line 1880, in _export export_artifact = export_func( # type: ignore[operator] ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "<torch_path>/torch/export/_trace.py", line 1224, in _strict_export return _strict_export_lower_to_aten_ir( ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "<torch_path>/torch/export/_trace.py", line 1333, in _strict_export_lower_to_aten_ir aten_export_artifact = lower_to_aten_callback( ^^^^^^^^^^^^^^^^^^^^^^^ File "<torch_path>/torch/export/_trace.py", line 637, in _export_to_aten_ir gm, graph_signature = transform(aot_export_module)( ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "<torch_path>/torch/_functorch/aot_autograd.py", line 1246, in aot_export_module fx_g, metadata, in_spec, out_spec = _aot_export_function( ^^^^^^^^^^^^^^^^^^^^^ File "<torch_path>/torch/_functorch/aot_autograd.py", line 1480, in _aot_export_function fx_g, meta = create_aot_dispatcher_function( ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "<torch_path>/torch/_functorch/aot_autograd.py", line 522, in create_aot_dispatcher_function return _create_aot_dispatcher_function( ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "<torch_path>/torch/_functorch/aot_autograd.py", line 729, in _create_aot_dispatcher_function raise RuntimeError( RuntimeError: aot_export is not currently supported with traceable tensor subclass. If you need this feature, please comment on <CREATE_ISSUE_LINK>

[jerryzh168]

@hahnec did you use unwrap_tensor_class before calling export?

[hahnec]

@jerryzh168 Yes, I used unwrap_tensor_subclass with version 2.4.x. However, then I upgraded, so the above error is from version 2.5.1 without using unwrap_tensor_subclass.

EDIT I need to mention that my custom model successfully gets converted with torch.onnx.export() without quantization. When using unwrap_tensor_subclass(model) with v2.5.1, I receive:

Exception has occurred: AttributeError 'UnwrapTensorSubclass' object has no attribute 'rebuild_stack' File "<path>/onnx_test2.py", line 28, in <module> model = unwrap_tensor_subclass(model) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ AttributeError: 'UnwrapTensorSubclass' object has no attribute 'rebuild_stack'

[ogencoglu]

@hahnec so what is the difference of quantization before conversion and quantizing the onnx model?

[jerryzh168]

@jerryzh168 Yes, I used unwrap_tensor_subclass with version 2.4.x. However, then I upgraded, so the above error is from version 2.5.1 without using unwrap_tensor_subclass.

EDIT I need to mention that my custom model successfully gets converted with torch.onnx.export() without quantization. When using unwrap_tensor_subclass(model) with v2.5.1, I receive:

Exception has occurred: AttributeError 'UnwrapTensorSubclass' object has no attribute 'rebuild_stack' File "<path>/onnx_test2.py", line 28, in <module> model = unwrap_tensor_subclass(model) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ AttributeError: 'UnwrapTensorSubclass' object has no attribute 'rebuild_stack'

can you do torch.export first before doing onnx export?