Move more utils to TorchAOBaseTensor#784
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/784
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 8586f6a with merge base 6080986 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
jerryzh168
force-pushed
the
base-class
branch
from
4c8eaac to
f70e229
Compare
Summary: This moves over _implements, _dispatch__torch_dispatch__, _dispatch__torch_function__, _register_layout_cls and _get_layout_tensor_constructor to `TorchAOBaseTensor` so when people inherit from this, they can get these utils directly Test Plan: python test/quantization/test_quant_api.py python test/integration/test_integration.py rely on CI for other tests Reviewers: Subscribers: Tasks: Tags:
jerryzh168
force-pushed
the
base-class
branch
from
f70e229 to
60444e7
Compare
|
can this handle the other ones from #710? feels like there should be a way to set it up so that things like detach are handlable with a combination of tensor_flatten and the constructor? |
yeah will do in a future PR |
|
|
||
| class TestOptim(TestCase): | ||
| @pytest.mark.skipif(not TORCH_VERSION_AT_LEAST_2_3, reason="requires PyTorch >= 2.3") | ||
| @pytest.mark.skipif(not TORCH_VERSION_AT_LEAST_2_4, reason="requires PyTorch >= 2.3") |
There was a problem hiding this comment.
cc @gau-nernst multiple levels of inheritance breaks pytorch 2.3 version (https://github.com/pytorch/ao/actions/runs/10689912664/job/29633120953?pr=784), so I had to use 2.4+ for this
There was a problem hiding this comment.
Personally I would prefer to keep supporting PyTorch 2.3 for low-bit optimizers. @msaroufim What do you think?
From the CI link you show, it seems like only 4-bit optim breaks, while 8-bit and Fp8 optim still work? Anyway, if we keep this change, also need to update the docs.
There was a problem hiding this comment.
After some investigation, the errors only happen to 4-bit optim, and because 4-bit optim uses dynamic-shape compile. Some strange interactions between inheritance and dynamic-shape 🤔. Might have some fix soon..
* Move more utils to TorchAOBaseTensor Summary: This moves over _implements, _dispatch__torch_dispatch__, _dispatch__torch_function__, _register_layout_cls and _get_layout_tensor_constructor to `TorchAOBaseTensor` so when people inherit from this, they can get these utils directly Test Plan: python test/quantization/test_quant_api.py python test/integration/test_integration.py rely on CI for other tests Reviewers: Subscribers: Tasks: Tags:
4 participants
Summary:
This moves over implements, dispatch__torch_dispatch, dispatch__torch_function_, _register_layout_cls and _get_layout_tensor_constructor to
TorchAOBaseTensorso when people inherit from this, they can get these utils directlyTest Plan:
python test/quantization/test_quant_api.py
python test/integration/test_integration.py
rely on CI for other tests
Reviewers:
Subscribers:
Tasks:
Tags: