fix

.github/workflows/test_cli_misc.yaml

@@ -58,7 +58,7 @@ jobs:
         run: |
           pip install --upgrade pip
           pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
-          pip install -e .[testing,timm,diffusers,codecarbon]
+          pip install -e .[testing]
 
       - name: Run tests
         run: pytest -s -k "cli and not (cpu or cuda or rocm or mps)"

optimum_benchmark/launchers/base.py

@@ -62,6 +62,12 @@ def device_isolation(self, pid: int, device_ids: Optional[str] = None):
 
     @contextmanager
     def numactl_executable(self):
+        self.logger.info("\t+ Warming up multiprocessing context")
+        dummy_process = Process(target=dummy_target, daemon=False)
+        dummy_process.start()
+        dummy_process.join()
+        dummy_process.close()
+
         self.logger.info("\t+ Creating numactl wrapper executable for multiprocessing")
         python_path = sys.executable
         numactl_path = shutil.which("numactl")
@@ -84,3 +90,7 @@ def numactl_executable(self):
         self.logger.info("\t+ Resetting default multiprocessing executable")
         os.unlink(numa_executable.name)
         set_executable(sys.executable)
+
+
+def dummy_target() -> None:
+    exit(0)

optimum_benchmark/launchers/inline/launcher.py

@@ -12,6 +12,6 @@ def __init__(self, config: InlineConfig):
         super().__init__(config)
 
     def launch(self, worker: Callable[..., BenchmarkReport], worker_args: List[Any]) -> BenchmarkReport:
-        self.logger.warn("The inline launcher is only recommended for debugging purposes and not for benchmarking")
+        self.logger.warning("The inline launcher is only recommended for debugging purposes and not for benchmarking")
         report = worker(*worker_args)
         return report

optimum_benchmark/launchers/process/launcher.py

@@ -23,12 +23,6 @@ def __init__(self, config: ProcessConfig):
         if get_start_method(allow_none=True) != self.config.start_method:
             self.logger.info(f"\t+ Setting multiprocessing start method to {self.config.start_method}")
             set_start_method(self.config.start_method, force=True)
-            # creates the resource tracker with default executable
-            self.logger.info("\t+ Warming up multiprocessing context")
-            dummy_process = Process(target=dummy_target, daemon=False)
-            dummy_process.start()
-            dummy_process.join()
-            dummy_process.close()
 
     def launch(self, worker: Callable[..., BenchmarkReport], worker_args: List[Any]) -> BenchmarkReport:
         child_connection, parent_connection = Pipe()
@@ -48,7 +42,6 @@ def launch(self, worker: Callable[..., BenchmarkReport], worker_args: List[Any])
             else:
                 raise RuntimeError("Could not synchronize with isolated process")
 
-        with ExitStack() as stack:
             if self.config.device_isolation:
                 stack.enter_context(self.device_isolation(isolated_process.pid))
 
@@ -121,21 +114,10 @@ def target(
 
 
 def sync_with_parent(child_connection: Connection) -> None:
-    if child_connection.poll():
-        response = child_connection.recv()
-    else:
-        raise RuntimeError("Received no response from main process")
-
-    if response == "SYNC":
-        return
-    else:
-        raise RuntimeError(f"Received an unexpected response from main process: {response}")
+    child_connection.recv()
+    child_connection.send(0)
 
 
 def sync_with_child(parent_connection: Connection) -> None:
-    parent_connection.send("SYNC")
+    parent_connection.send(0)
     parent_connection.recv()
-
-
-def dummy_target() -> None:
-    exit(0)

optimum_benchmark/launchers/torchrun/launcher.py

@@ -25,12 +25,6 @@ def __init__(self, config: TorchrunConfig):
         if get_start_method(allow_none=True) != self.config.start_method:
             self.logger.info(f"\t+ Setting multiprocessing start method to {self.config.start_method}")
             set_start_method(self.config.start_method, force=True)
-            self.logger.info("\t+ Warming up multiprocessing context")
-            # creates the resource tracker with default executable
-            dummy_process = Process(target=dummy_target, daemon=False)
-            dummy_process.start()
-            dummy_process.join()
-            dummy_process.close()
 
         self.launch_config = LaunchConfig(
             min_nodes=self.config.min_nodes,
@@ -68,7 +62,6 @@ def launch(self, worker: Callable[..., BenchmarkReport], worker_args: List[Any])
             else:
                 raise RuntimeError("Could not synchronize with isolated process")
 
-        with ExitStack() as stack:
             if self.config.device_isolation:
                 stack.enter_context(self.device_isolation(isolated_process.pid))
 
@@ -152,19 +145,12 @@ def target(
 
 
 def sync_with_parent(child_connection: Connection) -> None:
-    if child_connection.poll():
-        response = child_connection.recv()
-    else:
-        raise RuntimeError("Received no response from main process")
-
-    if response == "SYNC":
-        return
-    else:
-        raise RuntimeError(f"Received an unexpected response from main process: {response}")
+    child_connection.recv()
+    child_connection.send(0)
 
 
 def sync_with_child(parent_connection: Connection) -> None:
-    parent_connection.send("SYNC")
+    parent_connection.send(0)
     parent_connection.recv()
 
 
@@ -200,7 +186,3 @@ def entrypoint(worker: Callable[..., BenchmarkReport], worker_args: List[Any], l
         torch.distributed.destroy_process_group()
         logger.info("\t+ Exiting rank process")
         return output
-
-
-def dummy_target() -> None:
-    exit(0)

tests/test_cli.py

@@ -72,9 +72,6 @@ def test_cli_exit_code_0(launcher):
 
 @pytest.mark.parametrize("launcher", ["inline", "process", "torchrun"])
 def test_cli_exit_code_1(launcher):
-    if launcher == "torchrun" and sys.platform != "linux":
-        pytest.skip("torchrun is only supported on Linux")
-
     args_1 = [
         "optimum-benchmark",
         "--config-dir",
@@ -83,7 +80,7 @@ def test_cli_exit_code_1(launcher):
         "_base_",
         "name=test",
         f"launcher={launcher}",
-        # incompatible task and model to trigger error
+        # incompatible task and model to trigger an error
         "backend.task=image-classification",
         "backend.model=bert-base-uncased",
         "backend.device=cpu",