[Frontend] Move CLI code into vllm.cmd package (#12971)

vllm/entrypoints/cli/main.py

@@ -0,0 +1,79 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# The CLI entrypoint to vLLM.
+import os
+import signal
+import sys
+
+import vllm.entrypoints.cli.openai
+import vllm.entrypoints.cli.serve
+import vllm.version
+from vllm.logger import init_logger
+from vllm.utils import FlexibleArgumentParser
+
+logger = init_logger(__name__)
+
+CMD_MODULES = [
+    vllm.entrypoints.cli.openai,
+    vllm.entrypoints.cli.serve,
+]
+
+
+def register_signal_handlers():
+
+    def signal_handler(sig, frame):
+        sys.exit(0)
+
+    signal.signal(signal.SIGINT, signal_handler)
+    signal.signal(signal.SIGTSTP, signal_handler)
+
+
+def env_setup():
+    # The safest multiprocessing method is `spawn`, as the default `fork` method
+    # is not compatible with some accelerators. The default method will be
+    # changing in future versions of Python, so we should use it explicitly when
+    # possible.
+    #
+    # We only set it here in the CLI entrypoint, because changing to `spawn`
+    # could break some existing code using vLLM as a library. `spawn` will cause
+    # unexpected behavior if the code is not protected by
+    # `if __name__ == "__main__":`.
+    #
+    # References:
+    # - https://docs.python.org/3/library/multiprocessing.html#contexts-and-start-methods
+    # - https://pytorch.org/docs/stable/notes/multiprocessing.html#cuda-in-multiprocessing
+    # - https://pytorch.org/docs/stable/multiprocessing.html#sharing-cuda-tensors
+    # - https://docs.habana.ai/en/latest/PyTorch/Getting_Started_with_PyTorch_and_Gaudi/Getting_Started_with_PyTorch.html?highlight=multiprocessing#torch-multiprocessing-for-dataloaders
+    if "VLLM_WORKER_MULTIPROC_METHOD" not in os.environ:
+        logger.debug("Setting VLLM_WORKER_MULTIPROC_METHOD to 'spawn'")
+        os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
+
+
+def main():
+    env_setup()
+
+    parser = FlexibleArgumentParser(description="vLLM CLI")
+    parser.add_argument('-v',
+                        '--version',
+                        action='version',
+                        version=vllm.version.__version__)
+    subparsers = parser.add_subparsers(required=False, dest="subparser")
+    cmds = {}
+    for cmd_module in CMD_MODULES:
+        new_cmds = cmd_module.cmd_init()
+        for cmd in new_cmds:
+            cmd.subparser_init(subparsers).set_defaults(
+                dispatch_function=cmd.cmd)
+            cmds[cmd.name] = cmd
+    args = parser.parse_args()
+    if args.subparser in cmds:
+        cmds[args.subparser].validate(args)
+
+    if hasattr(args, "dispatch_function"):
+        args.dispatch_function(args)
+    else:
+        parser.print_help()
+
+
+if __name__ == "__main__":
+    main()