ci: update runner for IPU (#17183)

.azure/ipu-tests.yml

@@ -43,23 +43,31 @@ variables:
 jobs:
   - job: testing
     # how long to run the job before automatically cancelling
-    timeoutInMinutes: "15"
+    timeoutInMinutes: "20"
     pool: graphcore-ipus
     workspace:
       clean: all
 
     steps:
-    - script: |
-        ls -la /mnt/public/packages
-        ls -la /opt/poplar
-        tar -xvzf /opt/poplar/${{ variables.poplar_sdk }}.tar.gz
-      displayName: "Extract Poplar SDK"
 
     - script: |
         set -eux
+        # ls -la /mnt/public/packages
+        ls -la /opt/poplar
+        tar -xzf /opt/poplar/${{ variables.poplar_sdk }}.tar.gz
         pip debug --verbose
         pip install ${{ variables.poplar_sdk }}/poptorch-*ubuntu*.whl
-      displayName: "Install poptorch"
+      displayName: "Poplar SDK: Extract & Install"
+
+    - bash: |
+        # enable scripts basically just set/modify some environment variables
+        source ${{ variables.poplar_sdk }}/poplar-ubuntu*/enable.sh
+        gc-info --list-devices
+        printenv
+        python --version
+        pip --version
+        pip list
+      displayName: 'Image info & GraphCore'
 
     - script: |
         set -eux
@@ -78,17 +86,18 @@ jobs:
       displayName: "Reset IPU devices"
 
     - bash: |
-        for fpath in `ls requirements/**/*.txt`; do \
+        for fpath in `ls requirements/pytorch/*.txt`; do \
           python ./requirements/pytorch/adjust-versions.py $fpath; \
         done
-        pip install -e .[extra,examples,test]
 
+        pip install .[test] -f /mnt/public/packages
+        pip uninstall -y neptune-client  # it is not clear what version is compatible
         pip list
       env:
         PACKAGE_NAME: "pytorch"
         FREEZE_REQUIREMENTS: "1"
         GIT_TERMINAL_PROMPT: "1"
-      displayName: 'Install dependencies'
+      displayName: 'Install package & dependencies'
 
     - bash: |
         python requirements/collect_env_details.py

.github/checkgroup.yml

@@ -173,7 +173,7 @@ subprojects:
       - "build-cuda (3.9, 1.12, 11.6.1)"
       - "build-cuda (3.9, 1.13, 11.7.1)"
       - "build-cuda (3.10, 2.0, 11.7.1)"
-      - "build-ipu (3.9, 1.13)"
+      - "build-ipu (3.8, 1.13)"
       - "build-NGC"
       - "build-pl (3.9, 1.11, 11.3.1)"
       - "build-pl (3.9, 1.12, 11.6.1)"

.github/workflows/ci-dockers-pytorch.yml

@@ -145,7 +145,7 @@ jobs:
       matrix:
         include:
           # the config used in 'dockers/ci-runner-ipu/Dockerfile'
-          - {python_version: "3.9", pytorch_version: "1.13"}
+          - {python_version: "3.8", pytorch_version: "1.13"}
     steps:
       - uses: actions/checkout@v3
       - uses: docker/setup-buildx-action@v2

.github/workflows/ci-tests-pytorch.yml

@@ -116,7 +116,9 @@ jobs:
     - name: Install package & dependencies
       run: |
         pip install -q pip -U
-        pip install .[extra,test] -U "pytest-timeout" -f ${TORCH_URL} ${TORCH_PREINSTALL} -f ${PYPI_CACHE} --prefer-binary
+        pip install .[extra,test] -U \
+          "pytest-timeout" -r requirements/_integrations/accelerators.txt \
+          -f ${TORCH_URL} ${TORCH_PREINSTALL} -f ${PYPI_CACHE} --prefer-binary
         pip list
     - name: Dump handy wheels
       if: github.event_name == 'push' && github.ref == 'refs/heads/master'

requirements/_integrations/accelerators.txt

@@ -0,0 +1,2 @@
+# validation HPU connectors
+lightning-habana>=0.1.0rc0

requirements/pytorch/adjust-versions.py

@@ -20,7 +20,7 @@ def find_latest(ver: str) -> Dict[str, str]:
     ver = re.search(r"([\.\d]+)", ver).groups()[0]
     # in case there remaining dot at the end - e.g "1.9.0.dev20210504"
     ver = ver[:-1] if ver[-1] == "." else ver
-    print(f"finding ecosystem versions for: {ver}")
+    print(f"\n\n\nfinding ecosystem versions for: {ver}")
 
     # find first match
     for option in VERSIONS:

requirements/pytorch/test.txt

@@ -18,6 +18,3 @@ uvicorn<0.19.1  # for `ServableModuleValidator`
 
 tensorboard>=2.9.1, <2.12.0  # for `TensorBoardLogger`
 protobuf<=3.20.1 # strict  # an extra is updating protobuf, this pin prevents TensorBoard failure
-
-# validation HPU connectors
-lightning-habana>=0.1.0rc0

src/lightning/pytorch/loggers/neptune.py

@@ -35,8 +35,9 @@ from lightning.pytorch.utilities.model_summary import ModelSummary
 from lightning.pytorch.utilities.rank_zero import rank_zero_only
 
 # neptune is available with two names on PyPI : `neptune` and `neptune-client`
-_NEPTUNE_AVAILABLE = RequirementCache("neptune")
+_NEPTUNE_AVAILABLE = RequirementCache("neptune>=1.0")
 _NEPTUNE_CLIENT_AVAILABLE = RequirementCache("neptune-client")
+
 if _NEPTUNE_AVAILABLE:
     # >1.0 package structure
     import neptune