You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Pretrained PyTorch models are available from multiple sources. Most bottom out at nn.Module, but there doesn't appear to be a convenient way to reflect on an arbitrary nn.Module to get a canonical input signature (like with ONNX).
PyTorch models can be very flexible. It may be worth testing mutliple versions of each model: different data types, different implementations (e.g. attention op variants, different dynamic shape sizes, etc.)
iree-turbine may have some issues on Windows:
tests/torchvision/resnet_test.py::test_resnet50 Windows fatal exception: access violation
Current thread 0x00003e6c (most recent call first):
File "D:\dev\projects\iree-test-suites\pytorch_models\.venv\Lib\site-packages\iree\turbine\aot\support\ir_utils.py", line 332 in create_tensor_global
File "D:\dev\projects\iree-test-suites\pytorch_models\.venv\Lib\site-packages\iree\turbine\aot\support\procedural\globals.py", line 126 in track
File "D:\dev\projects\iree-test-suites\pytorch_models\.venv\Lib\site-packages\iree\turbine\aot\support\procedural\exported_program.py", line 323 in _lift_tensor_to_global
...
Might first focus on export -> compile coverage, then later add runtime execution and comparison with native torch on CPU and other backends
The text was updated successfully, but these errors were encountered:
tests/torchvision/resnet_test.py::test_resnet50 Windows fatal exception: access violation
Current thread 0x00003e6c (most recent call first):
File "D:\dev\projects\iree-test-suites\pytorch_models\.venv\Lib\site-packages\iree\turbine\aot\support\ir_utils.py", line 332 in create_tensor_global
File "D:\dev\projects\iree-test-suites\pytorch_models\.venv\Lib\site-packages\iree\turbine\aot\support\procedural\globals.py", line 126 in track
File "D:\dev\projects\iree-test-suites\pytorch_models\.venv\Lib\site-packages\iree\turbine\aot\support\procedural\exported_program.py", line 323 in _lift_tensor_to_global
...
FWIW, that test segfaults on Linux too, so some issues with my test setup.
Related issues:
I want to delete these old tests: https://github.com/nod-ai/SHARK-TestSuite/tree/main/iree_tests/pytorch/models, but want a replacement first. There are currently just two tests there, opt-125M and resnet50, each imported from "e2eshark" using this script. The e2eshark code exports using iree-turbine with the code in
tools/stubs/turbinemodel.py
and these test cases:resnet50/model.py
,opt-125M/model.py
.I've started work on this on a branch: https://github.com/ScottTodd/iree-test-suites/tree/pytorch-models-setup. Some observations:
Pretrained PyTorch models are available from multiple sources. Most bottom out at
nn.Module
, but there doesn't appear to be a convenient way to reflect on an arbitrarynn.Module
to get a canonical input signature (like with ONNX).PyTorch models can be very flexible. It may be worth testing mutliple versions of each model: different data types, different implementations (e.g. attention op variants, different dynamic shape sizes, etc.)
iree-turbine may have some issues on Windows:
Might first focus on export -> compile coverage, then later add runtime execution and comparison with native torch on CPU and other backends
The text was updated successfully, but these errors were encountered: