Disable docker sandbox (#4649)

* Revert "Disable the Sandbox build in Docker (#4515)"

This reverts commit 661ab6d.

* Remove submodule step from TPU CI

* Update cloudbuild.yaml

* Remove separate TF build step

* Add bazel WORKSPACE and rc

* Update Dockerfile

* throw more compute at the TPU CI build

* more bazel jobs

* Revert to old machine type and default to local builds.

* Revert to default nonsandbox

* Remove jobs limitation

* Try smaller shm

* Fix build of llvm in sandbox

* Mount 7g of ram properly

* TEMP remove pytorch build

* remove install of packages

* confirm shm creation

* Revert machine type

---------

Co-authored-by: Will Cromar <[email protected]>
Co-authored-by: Will Cromar <[email protected]>

.bazelrc

@@ -10,8 +10,6 @@ build --announce_rc
 # TODO(goranpetrovic): figure out visibility of tensorflow libraries.
 build --nocheck_visibility
 
-#build --define open_source_build=true
-
 # We can set this to `standalone` after https://github.com/bazelbuild/bazel/issues/15359 is resolved.
 build --spawn_strategy=sandboxed
 

docker/Dockerfile

@@ -40,7 +40,7 @@ ENV BUNDLE_LIBTPU "${tpuvm}"
 ENV BAZEL_JOBS "${bazel_jobs}"
 
 # This makes the bazel build behave more consistently, but runs slower.
-ENV XLA_SANDBOX_BUILD "0"
+ENV XLA_SANDBOX_BUILD "1"
 ENV XLA_SANDBOX_BASE "/dev/shm"
 
 # To get around issue of Cloud Build with recursive submodule update

docker/experimental/Dockerfile

@@ -70,28 +70,29 @@ WORKDIR /pytorch/xla/
 FROM builder AS artifacts
 
 COPY third_party/ third_party/
-
+COPY tf_patches/ tf_patches/
+COPY .bazelrc .
+COPY .bazelversion .
+COPY WORKSPACE .
 COPY build_torch_xla_libs.sh .
 
 # TODO: Remove this when it's not required anymore
-ENV XLA_SANDBOX_BUILD=0
+ENV XLA_SANDBOX_BUILD=1
 ENV XLA_SANDBOX_BASE "/dev/shm"
 
-ARG tpuvm
-ARG cuda
-ARG tf_cuda_compute_capabilities
-ARG bazel_jobs
-RUN TPUVM_MODE=${tpuvm} XLA_CUDA=${cuda} BAZEL_JOBS=${bazel_jobs} TF_CUDA_COMPUTE_CAPABILITIES=${tf_cuda_compute_capabilities} bash build_torch_xla_libs.sh -O -D_GLIBCXX_USE_CXX11_ABI=1
-
 COPY torch_xla/ torch_xla/
 COPY setup.py .
 COPY xla_native_functions.yaml .
 
 COPY scripts/ scripts/
 
+ARG tpuvm
+ARG cuda
+ARG tf_cuda_compute_capabilities
+ARG bazel_jobs
 ARG build_cpp_tests
 ARG package_version
-RUN TORCH_XLA_VERSION=${package_version} BUILD_CPP_TESTS=${build_cpp_tests} TPUVM_MODE=${tpuvm} BUNDLE_LIBTPU=${tpuvm} XLA_CUDA=${cuda} TF_CUDA_COMPUTE_CAPABILITIES=${tf_cuda_compute_capabilities} python setup.py bdist_wheel
+RUN --mount=type=tmpfs,target=/dev/shm,rw TORCH_XLA_VERSION=${package_version} BUILD_CPP_TESTS=${build_cpp_tests} TPUVM_MODE=${tpuvm} BUNDLE_LIBTPU=${tpuvm} XLA_CUDA=${cuda} TF_CUDA_COMPUTE_CAPABILITIES=${tf_cuda_compute_capabilities} python setup.py bdist_wheel
 
 # Expunge cache to keep image size under control
 RUN bazel clean --expunge

docker/experimental/cloudbuild.yaml

@@ -1,12 +1,4 @@
 steps:
-# We only need to update submodules in triggers. User must update submodule
-# before local runs because .git is not present.
-- name: 'alpine/git'
-  entrypoint: sh
-  args:
-  - -c
-  - |
-    git submodule update --init || echo No git repository found
 - id: 'common-flags'
   name: 'bash'
   args:

test/tpu/cloudbuild.yaml

@@ -6,14 +6,6 @@ substitutions:
   # Location of GKE cluster.
   _CLUSTER_ZONE: 'europe-west4-a'
 steps:
-# We only need to update submodules in triggers. User must update submodule
-# before local runs because .git is not present.
-- name: 'alpine/git'
-  entrypoint: sh
-  args:
-  - -c
-  - |
-    git submodule update --init || echo No git repository found
 - name: 'docker'
   id: build-image
   args: [
@@ -26,8 +18,6 @@ steps:
     '--shm-size=16G',
     '--build-arg',
     'tpuvm=1',
-    '--build-arg',
-    'bazel_jobs=8'
   ]
 - name: 'docker'
   id: push-image