[TorchDISC] setup torch-disc building environment and CI job #158
Conversation
Yancey1989
changed the title
Yancey1989
force-pushed
the
torch_disc_build_env
branch
from
7774287
to
aac1e66
Compare
Yancey1989
force-pushed
the
torch_disc_build_env
branch
from
75052d2
to
f827338
Compare
| # CPU-TORCH181: | ||
| # uses: ./.github/workflows/cpu_reusable.yml | ||
| # with: | ||
| # extra_envs: -e TORCH_BLADE_BUILD_WITH_CUDA_SUPPORT=OFF |
There was a problem hiding this comment.
Nit: Delete these codes if we don't need them
There was a problem hiding this comment.
We should remove the comment symbol when merging to main in the future, so I keep these jobs in the current feature branch.
| apt-get install -y clang-8 clang++-8 python3.8 python3.8-dev && \ | ||
| rm /usr/bin/python && \ | ||
| ln -s /usr/bin/python3.8 /usr/bin/python && \ | ||
| python -m pip install --upgrade pip && \ |
There was a problem hiding this comment.
Use multi-stage command would improve cache hit rate, such as:
RUN apt-get install lib1 lib2
RUN python -m pip install wheel1 whee2
|
|
||
|
|
||
| setup( | ||
| name='torch_disc', |
There was a problem hiding this comment.
I prefer to eventually build only one wheel between torch_blade and torch_disc since they share a lot of things. And users of PyTorch may be confused why we should have two package to support inference and training.
But I think it's fine currently in our POC version.
There was a problem hiding this comment.
Agree with that, but a latency problem is that torch-disc depends on lazy_tensor_staging feature branch of PyTorch which is different from the current version TorchBlade.
This PR builds the TorchDISC pybind library `_torch_disc.so` with Torch LTC, Torch Blade, and official PyTorch source code. And also add a CI job to test it with a rough way: `import _torch_disc`. The CI job includes two parts to reduce the CI time: 1. building devel docker: install PyTorch with `lazy_tensor_staging` branch from source code. 2. build and test torch-disc with the devel image from step1. Develops can build from zero with the guide https://github.com/alibaba/BladeDISC/blob/de3a386114dfcd6038c16c9fbe0737c0a6c5f6ef/torch_disc/README.md
…#158) This PR builds the TorchDISC pybind library `_torch_disc.so` with Torch LTC, Torch Blade, and official PyTorch source code. And also add a CI job to test it with a rough way: `import _torch_disc`. The CI job includes two parts to reduce the CI time: 1. building devel docker: install PyTorch with `lazy_tensor_staging` branch from source code. 2. build and test torch-disc with the devel image from step1. Develops can build from zero with the guide https://github.com/alibaba/BladeDISC/blob/de3a386114dfcd6038c16c9fbe0737c0a6c5f6ef/torch_disc/README.md
This PR builds the TorchDISC pybind library
_torch_disc.sowith Torch LTC, Torch Blade, and official PyTorch source code.And also add a CI job to test it with a rough way:
import _torch_disc.The CI job includes two parts to reduce the CI time:
lazy_tensor_stagingbranch from source code.Develops can build from zero with the guide https://github.com/alibaba/BladeDISC/blob/de3a386114dfcd6038c16c9fbe0737c0a6c5f6ef/torch_disc/README.md