Feature: update to torch 2.5.0

.circleci/docker/Dockerfile

@@ -1,6 +1,6 @@
-ARG PYTORCH="1.8.1"
-ARG CUDA="10.2"
-ARG CUDNN="7"
+ARG PYTORCH="2.5.0"
+ARG CUDA="12.4"
+ARG CUDNN="9"
 
 FROM pytorch/pytorch:${PYTORCH}-cuda${CUDA}-cudnn${CUDNN}-devel
 

.circleci/test.yml

@@ -108,7 +108,7 @@ jobs:
             pip install wheel
       - run:
           name: Install PyTorch
-          command: pip install torch==<< parameters.torch >>+cpu torchvision==<< parameters.torchvision >>+cpu -f https://download.pytorch.org/whl/torch_stable.html
+          command: pip install torch==<< parameters.torch >>+cpu torchvision==<< parameters.torchvision >>+cpu -f https://download.pytorch.org/whl/torch -f https://download.pytorch.org/whl/torchvision
       - run:
           name: Build MMEngine from source
           command: pip install -e . -v
@@ -137,7 +137,7 @@ jobs:
         type: string
       cuda:
         type: enum
-        enum: ["10.1", "10.2", "11.1", "11.7", "11.8"]
+        enum: ["10.1", "10.2", "11.1", "11.7", "11.8", "12.1", "12.4"]
       cudnn:
         type: integer
         default: 7
@@ -267,9 +267,9 @@ workflows:
             - lint
       - build_cpu:
           name: maximum_version_cpu
-          torch: 2.1.0
-          torchvision: 0.16.0
-          python: 3.9.0
+          torch: 2.5.0
+          torchvision: 0.20.0
+          python: 3.11.9
           requires:
             - minimum_version_cpu
       - hold_integration_test:
@@ -296,9 +296,9 @@ workflows:
             - hold
       - build_cuda:
           name: maximum_version_gpu
-          torch: 2.1.0
-          cuda: "11.8"
-          cudnn: 8
+          torch: 2.5.0
+          cuda: "12.4"
+          cudnn: 9
           requires:
             - hold
   merge_stage_test:

.gitignore

@@ -109,6 +109,9 @@ venv.bak/
 .vscode
 .idea
 .DS_Store
+.aim
+.dvc
+.dvcignore
 
 # custom
 *.pkl

docker/dev/Dockerfile

@@ -1,6 +1,6 @@
-ARG PYTORCH="1.8.1"
-ARG CUDA="10.2"
-ARG CUDNN="7"
+ARG PYTORCH="2.5.0"
+ARG CUDA="12.4"
+ARG CUDNN="9"
 
 FROM pytorch/pytorch:${PYTORCH}-cuda${CUDA}-cudnn${CUDNN}-devel
 

docker/release/Dockerfile

@@ -1,6 +1,6 @@
-ARG PYTORCH="1.8.1"
-ARG CUDA="10.2"
-ARG CUDNN="7"
+ARG PYTORCH="2.5.0"
+ARG CUDA="12.4"
+ARG CUDNN="9"
 
 FROM pytorch/pytorch:${PYTORCH}-cuda${CUDA}-cudnn${CUDNN}-devel
 

mmengine/optim/optimizer/builder.py

@@ -160,15 +160,9 @@ def register_bitsandbytes_optimizers() -> List[str]:
 
 
 def register_transformers_optimizers():
-    transformer_optimizers = []
-    try:
-        from transformers import Adafactor
-    except ImportError:
-        pass
-    else:
-        OPTIMIZERS.register_module(name='Adafactor', module=Adafactor)
-        transformer_optimizers.append('Adafactor')
-    return transformer_optimizers
+    return [
+        'Adafactor',
+    ]
 
 
 TRANSFORMERS_OPTIMIZERS = register_transformers_optimizers()

mmengine/testing/_internal/distributed.py

@@ -94,8 +94,15 @@ def wrapper(self):
     # or run the underlying test function.
     def __init__(self, method_name: str = 'runTest') -> None:
         super().__init__(method_name)
-        fn = getattr(self, method_name)
-        setattr(self, method_name, self.join_or_run(fn))
+        try:
+            fn = getattr(self, method_name)
+            setattr(self, method_name, self.join_or_run(fn))
+        except AttributeError as e:
+            if method_name != 'runTest':
+                # we allow instantiation with no explicit method name
+                # but not an *incorrect* or missing method name
+                raise ValueError(f'no such test method in {self.__class__}:'
+                                 f' {method_name}') from e
 
     def setUp(self) -> None:
         super().setUp()
@@ -345,12 +352,13 @@ def _check_return_codes(self, elapsed_time) -> None:
             if first_process.exitcode == skip.exit_code:
                 raise unittest.SkipTest(skip.message)
 
-        # Skip the unittest since the raised error maybe not caused by
-        # the tested function. For example, in CI environment, the tested
-        # method could be terminated by system signal for the limited
-        # resources.
-        self.skipTest(f'Skip test {self._testMethodName} due to '
-                      'the program abort')
+        if first_process.exitcode != 0:
+            # Skip the unittest since the raised error maybe not caused by
+            # the tested function. For example, in CI environment, the tested
+            # method could be terminated by system signal for the limited
+            # resources.
+            self.skipTest(f'Skip test {self._testMethodName} due to '
+                          'the program abort')
 
     @property
     def is_master(self) -> bool:

requirements/tests.txt

@@ -1,4 +1,4 @@
-aim<=3.17.5;sys_platform!='win32'
+aim;sys_platform!='win32'
 bitsandbytes
 clearml
 coverage