Add CUDA 11.3 + PyTorch 1.11.0 and 1.12.1#4205
Add CUDA 11.3 + PyTorch 1.11.0 and 1.12.1#4205thiagocrepaldi wants to merge 21 commits intofacebookresearch:mainfrom
Conversation
facebook-github-bot
added
the
CLA Signed
thiagocrepaldi
force-pushed
the
thiagofc/update-build-cuda
branch
29 times, most recently
from
9a76428 to
538226c
Compare
yeah, i saw that and I honestly ran out of tricks to force the right cuda version. It seems that by PEP400 (if I am not mistaken), the "+" is called "local version label" and it is not used for version selection. The pytorch's URL for cu113 also has wheels for a bunch of other cuda versions, and this mismatch is happening |
| - <<: *setupcuda113 | ||
| - <<: *removecuda114 |
There was a problem hiding this comment.
shouldn't it remove 114 first then setup 113?
There was a problem hiding this comment.
Removing 11.4 first and installing 11.3 would result in uninstalling several shared components only to reinstall later, increasing installation time. Keeping the current order does not affect the end result and save some time
|
@wat3rBro everything is green now |
|
@wat3rBro gentle ping |
All non pytorch master pipelines are unchanged 3 new pileines are introduced. * 1) Windows CPU and 2) Linux CPU with pytorch/torchvision nightly builds * 2) Linux with CUDA 11.3, python 3.9 and nightly pytorch/torchvision CUDA 11.3 is needed because PyTorch does not distribute wheels for nightly 11.1 Python 3.9 was needed for the same reason; no torch wheels for 3.6
Signed-off-by: Thiago Crepaldi <[email protected]>
|
Is this still relevant or should we close it? |
5 participants
Currently CI supports only CUDA 11.1, which is sufficient for PyTorch versions up to 1.10.
For PyTorch 1.11+ and nightly builds, the minimum CUDA version is 11.3.
This PR adds 3 new pipelines:
The new pipelines are important to test end-to-end ONNX export tests, in which PyTorch inference results are numerically compared with the ONNX Runtime's . A second benefit is help identifying ONNX export issues earlier in the development cycle.
ps: Nightly builds are only installed in the new pipelines (aka
windows_cpu_build_pytorch_master,linux_cuda113_tests_pytorch_master_python39,linux_cpu_tests_pytorch_master. All existing pipelines are intact.