Fixes for release

dev/packaging/apex_conda/README.md

@@ -11,6 +11,9 @@ nothing to do here.
 2. Go into the `inside` directory and clone apex with
 `git clone https://github.com/NVIDIA/apex.git`.
 
+Move to the appropriate commit.
+`git checkout 1f2aa9156547377a023932a1512752c392d9bbdf`.
+
 3. You may want to `docker pull pytorch/conda-cuda:latest`.
 
 4. Run `bash go.sh` in this directory. This takes ages

dev/packaging/apex_conda/inside/build.py

@@ -24,6 +24,9 @@
     "1.7.0": ["cu101", "cu102", "cu110"],
     "1.7.1": ["cu101", "cu102", "cu110"],
     "1.8.0": ["cu101", "cu102", "cu111"],
+    "1.8.1": ["cu101", "cu102", "cu111"],
+    "1.9.0": ["cu102", "cu111"],
+    "1.9.1": ["cu102", "cu111"],
 }
 
 CUDA_HOMES = {
@@ -72,6 +75,15 @@ def pytorch_versions_for_python(python_version):
 
             print()
             print("python", python_version, "pytorch", ptv, "cuda", cuv, flush=True)
+
+            apex_file_name = (
+                f"apex-{VERSION}-py{python_version_nodot}_{cuv}_pyt{ptv_nodot}.tar.bz2"
+            )
+
+            if os.path.exists(f"./inside/packaging/{apex_file_name}"):
+                print(f"Package: {apex_file_name} already found")
+                continue
+
             args = [
                 "conda",
                 "build",
@@ -87,8 +99,10 @@ def pytorch_versions_for_python(python_version):
             if python_version == "3.9" or cuv == "cu111":
                 args.insert(4, "conda-forge")
                 args.insert(4, "-c")
+
             subprocess.check_call(args)
-            file = f"/opt/conda/conda-bld/linux-64/apex-{VERSION}-py{python_version_nodot}_{cuv}_pyt{ptv_nodot}.tar.bz2"
+
+            file = f"/opt/conda/conda-bld/linux-64/{apex_file_name}"
             shutil.copy(file, "inside/packaging")
 
 print("DONE")

dev/packaging/apex_pip/README.md

@@ -11,6 +11,9 @@ nothing to do here.
 2. Go into the `inside` directory and clone apex with
 `git clone https://github.com/NVIDIA/apex.git`.
 
+Move to the appropriate commit.
+`git checkout 1f2aa9156547377a023932a1512752c392d9bbdf`.
+
 3. You may want to `docker pull pytorch/conda-cuda:latest`.
 
 4. Run `bash go.sh` in this directory. This takes ages
@@ -19,7 +22,6 @@ and writes packages to `inside/output`.
 5. You can upload the packages to s3, along with basic html files
 which enable them to be used, with `bash after.sh`.
 
-
 In particular, if you are in a jupyter/colab notebook you can
 then install using these wheels with the following series of
 commands.

dev/packaging/apex_pip/inside/a.sh

@@ -34,6 +34,9 @@ declare -A CONDA_CUDA_VERSIONS=(
     ["1.7.0"]="cu101 cu102 cu110"
     ["1.7.1"]="cu101 cu102 cu110"
     ["1.8.0"]="cu101 cu102 cu111"
+    ["1.8.1"]="cu101 cu102 cu111"
+    ["1.9.0"]="cu102 cu111"
+    ["1.9.1"]="cu102 cu111"
 )
 
 #VERSION=$(python -c "exec(open('${script_dir}/apex/__init__.py').read()); print(__version__)")
@@ -85,6 +88,7 @@ do
             outdir="../output/$tag"
             if [[ -d "$outdir" ]]
             then
+                echo "skipping" "$outdir"
                 continue
             fi
 

dev/packaging/vissl_conda/build_all_conda.sh

@@ -8,8 +8,13 @@ set -ex
 
 rm -rf dev/packaging/vissl_conda/ClassyVision
 git clone https://github.com/facebookresearch/ClassyVision.git dev/packaging/vissl_conda/ClassyVision
+
 rm -rf dev/packaging/vissl_conda/fairscale
 git clone https://github.com/facebookresearch/fairscale.git dev/packaging/vissl_conda/fairscale
+cd dev/packaging/vissl_conda/fairscale
+git reset --hard df7db85cef7f9c30a5b821007754b96eb1f977b6
+cd ../../../../
+
 rm -rf classy_vision
 cp -r dev/packaging/vissl_conda/ClassyVision/classy_vision classy_vision
 rm -rf fairscale

dev/packaging/vissl_conda/vissl/meta.yaml

@@ -38,7 +38,7 @@ test:
   source_files:
     - tests
     - tools
-    - dev/run_quick_tests.sh
+    - dev/run_quick_integration_tests.sh
     - configs
   requires:
     - ca-certificates
@@ -47,7 +47,7 @@ test:
     - tensorboard
   commands:
     - python -m unittest discover -v -s tests
-    - ./dev/run_quick_tests.sh
+    - ./dev/run_quick_integration_tests.sh
 
 build:
   string: py{{py}}

dev/run_quick_tests.sh

@@ -43,36 +43,4 @@ popd
 # - verify that the associated jobs run to the end
 # -----------------------------------------------------------------------------
 
-CFG_LIST=(
-    "test/integration_test/quick_barlow_twins"
-    "test/integration_test/quick_deepcluster_v2"
-    "test/integration_test/quick_pirl"
-    "test/integration_test/quick_simclr"
-    "test/integration_test/quick_simclr_efficientnet"
-    "test/integration_test/quick_simclr_multicrop"
-    "test/integration_test/quick_simclr_regnet"
-    "test/integration_test/quick_swav"
-)
-
-echo "========================================================================"
-echo "Configs to run:"
-echo "${CFG_LIST[@]}"
-echo "========================================================================"
-
-BINARY="python ${SRC_DIR}/tools/run_distributed_engines.py"
-
-for cfg in "${CFG_LIST[@]}"; do
-    echo "========================================================================"
-    echo "Running $cfg ..."
-    echo "========================================================================"
-    CHECKPOINT_DIR=$(mktemp -d)
-    # shellcheck disable=SC2102
-    # shellcheck disable=SC2086
-    CUDA_LAUNCH_BLOCKING=1 $BINARY config=$cfg \
-        config.DATA.TRAIN.DATA_SOURCES=[synthetic] \
-        hydra.verbose=true \
-        config.HOOKS.TENSORBOARD_SETUP.USE_TENSORBOARD=true \
-        config.CHECKPOINT.DIR="$CHECKPOINT_DIR" && echo "TEST OK" || exit
-
-    rm -rf $CHECKPOINT_DIR
-done
+bash "${SRC_DIR}/dev/run_quick_integration_tests.sh"

tests/test_tasks.py

@@ -7,6 +7,7 @@
 import unittest
 
 import pkg_resources
+import torch
 from parameterized import parameterized
 from utils import UNIT_TEST_CONFIGS, SSLHydraConfig
 from vissl.engines.train import train_main
@@ -38,6 +39,10 @@ def test_run(self, config_file_path: str):
             pkg_resources.resource_filename(__name__, "test_data")
         ]
 
+        # Destroy process groups as torch may be initialized with NCCL, which
+        # is incompatible with test_cpu_regnet_moco.yaml
+        torch.distributed.destroy_process_group()
+
         # run training and make sure no exception is raised
         dist_run_id = get_dist_run_id(config, config.DISTRIBUTED.NUM_NODES)
         train_main(

vissl/utils/test_utils.py

@@ -29,8 +29,10 @@ def in_temporary_directory(enabled: bool = True):
         old_cwd = os.getcwd()
         with tempfile.TemporaryDirectory() as temp_dir:
             os.chdir(temp_dir)
-            yield temp_dir
-            os.chdir(old_cwd)
+            try:
+                yield temp_dir
+            finally:
+                os.chdir(old_cwd)
     else:
         yield os.getcwd()