removed check for ConvTranspose3D on MPS #145366
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/145366
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 3f843fe with merge base 3cbc8c5 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
|
|
|
Any updates on this? |
|
I'm using my branch for some time now and it works like a charm. Any official updates would be appreciated @kulinseth @malfet |
|
Hey @mlaves! How can I test your branch? Should I follow the From Source installation instructions? |
Just clone my branch and do |
Thanks! The following commands eventually worked! git clone --depth 1 --branch convtranspose_mps_remove_check https://github.com/mlaves/pytorch.git
cd pytorch
git submodule update --init --recursive
python3 -m venv venv
source ./venv/bin/activate
pip install -r requirements.txt
python3 setup.py develop |
|
Any updates on this? Trying to run 3D UNET and getting the same error. |
|
Hi! Thanks for your work here! I managed to run the PARIETAL tool by building this branch. I also tried the implementation proposed earlier in #116580 and noticed that the other PR is much faster: 70s versus 20s. (M1 Pro) I tested other models and experienced a similar performance difference. Did you notice the same results in your tests? I understand that PR #116580 wasn't ready to be merged, but this seems like a huge improvement to be considered. Results Comparisonusing #145366 (this PR)python -c "import torch; print(torch.__version__)"
2.7.0a0+git3f843fe
time python parietal.py --input_scan tests/example/T1.nii.gz --output_scan out.nii.gz
python parietal.py --input_scan tests/example/T1.nii.gz --output_scan 4.54s user 3.59s system 11% cpu 1:11.13 total
time python parietal.py --input_scan tests/example/T1.nii.gz --output_scan out.nii.gz
python parietal.py --input_scan tests/example/T1.nii.gz --output_scan 4.70s user 4.06s system 12% cpu 1:11.89 totalusing #116580I used https://github.com/LalithShiyam/pytorch-mps which is #116580 with some deps fixes python -c "import torch; print(torch.__version__)"
2.3.0a0+gitffda73c #
time python parietal.py --input_scan tests/example/T1.nii.gz --output_scan out.nii.gz
python parietal.py --input_scan tests/example/T1.nii.gz --output_scan 4.15s user 4.94s system 46% cpu 19.735 total
time python parietal.py --input_scan tests/example/T1.nii.gz --output_scan out.nii.gz
python parietal.py --input_scan tests/example/T1.nii.gz --output_scan 4.22s user 3.23s system 35% cpu 20.800 total
time python parietal.py --input_scan tests/example/T1.nii.gz --output_scan out.nii.gz
python parietal.py --input_scan tests/example/T1.nii.gz --output_scan 4.17s user 3.23s system 30% cpu 24.133 total
time python parietal.py --input_scan tests/example/T1.nii.gz --output_scan out.nii.gz
python parietal.py --input_scan tests/example/T1.nii.gz --output_scan 4.10s user 3.49s system 34% cpu 21.816 total |
|
@rogerbramon Thank you for your testing and benchmarking! I'd like to clarify that my approach and @mattiaspaul's are complementary rather than conflicting. PyTorch currently utilizes MPSGraph to implement GPU-accelerated In contrast, PR #116580 takes a different approach by cleverly reimplementing |
|
Thanks @mlaves for the info. So what would be the next steps to move it forward? |
Well, we try to get some attention from the maintainers to review this PR. |
|
@mlaves I think next step is for you to sign the CLA, as pull requests can not be accepted otherwise |
|
/easycla |
I signed the CLA, but it's still shown as missing here. |
Hi @mlaves thanks for unblocking transpose Conv3d hopefully this gets integrated soon - I guess you're using this for 3D UNets right? @rogerbramon if you have a good use case that benefits from my speeded-up version I'm happy to see whether we can get that running / integrated as well. The big question is whether one wants to add an option to run Conv3D-as-Conv2D when speed is an important aspect and fall back to native Conv3D elsewhere. This would add some complexity but could be inline with CuDNN/CUDA fast/slow implementations in PyTorch. |
Fixes #130256
I removed
TORCH_CHECK(input_t.dim() < 5, "ConvTranspose 3D is not supported on MPS");as it is actually supported.