[distributed][misc] use fork by default for mp (#5669)

.buildkite/test-pipeline.yaml

@@ -37,6 +37,9 @@ steps:
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   commands:
+  # FIXIT: find out which code initialize cuda before running the test
+  # before the fix, we need to use spawn to test it
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
   - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
@@ -55,6 +58,9 @@ steps:
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
   commands:
+  # FIXIT: find out which code initialize cuda before running the test
+  # before the fix, we need to use spawn to test it
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s distributed/test_pynccl.py
   # We want to test that models which use 2 GPUs work with 4 GPUs, which is why we duplicate them here.
   # See https://github.com/vllm-project/vllm/pull/5473#issuecomment-2166601837 for context.
@@ -145,6 +151,9 @@ steps:
   num_gpus: 4
   # This test runs llama 13B, so it is required to run on 4 GPUs.
   commands:
+    # FIXIT: find out which code initialize cuda before running the test
+    # before the fix, we need to use spawn to test it
+    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
     - pytest -v -s -x lora/test_long_context.py
 
 - label: Tensorizer Test

vllm/distributed/device_communicators/custom_all_reduce_utils.py

@@ -1,6 +1,9 @@
 import ctypes
 import json
 import os
+import pickle
+import subprocess
+import sys
 from itertools import product
 from typing import Dict, List, Optional, Sequence
 
@@ -198,7 +201,25 @@ def gpu_p2p_access_check(src: int, tgt: int) -> bool:
         ids = list(range(num_dev))
         # batch of all pairs of GPUs
         batch_src, batch_tgt = zip(*list(product(ids, ids)))
-        result = can_actually_p2p(batch_src, batch_tgt)
+        # NOTE: we use `subprocess` rather than `multiprocessing` here
+        # because the caller might not have `if __name__ == "__main__":`,
+        # in that case we cannot use spawn method in multiprocessing.
+        # However, `can_actually_p2p` requires spawn method.
+        # The fix is, we use `subprocess` to call the function,
+        # where we have `if __name__ == "__main__":` in this file.
+        input_bytes = pickle.dumps((batch_src, batch_tgt))
+        returned = subprocess.run([sys.executable, __file__],
+                                  input=input_bytes,
+                                  capture_output=True)
+        # check if the subprocess is successful
+        try:
+            returned.check_returncode()
+        except Exception as e:
+            # wrap raised exception to provide more information
+            raise RuntimeError(
+                f"Error happened when batch testing "
+                f"peer-to-peer access from {batch_src} to {batch_tgt}") from e
+        result = pickle.loads(returned.stdout)
         for _i, _j, r in zip(batch_src, batch_tgt, result):
             cache[f"{_i}->{_j}"] = r
         with open(path, "w") as f:
@@ -213,3 +234,8 @@ def gpu_p2p_access_check(src: int, tgt: int) -> bool:
 
 
 __all__ = ["gpu_p2p_access_check"]
+
+if __name__ == "__main__":
+    batch_src, batch_tgt = pickle.loads(sys.stdin.buffer.read())
+    result = can_actually_p2p(batch_src, batch_tgt)
+    sys.stdout.buffer.write(pickle.dumps(result))

vllm/envs.py

@@ -29,7 +29,7 @@ if TYPE_CHECKING:
     VLLM_CPU_KVCACHE_SPACE: int = 0
     VLLM_XLA_CACHE_PATH: str = "~/.vllm/xla_cache/"
     VLLM_USE_RAY_COMPILED_DAG: bool = False
-    VLLM_WORKER_MULTIPROC_METHOD: str = "spawn"
+    VLLM_WORKER_MULTIPROC_METHOD: str = "fork"
     VLLM_IMAGE_FETCH_TIMEOUT: int = 5
     VLLM_TARGET_DEVICE: str = "cuda"
     MAX_JOBS: Optional[str] = None
@@ -212,7 +212,7 @@ environment_variables: Dict[str, Callable[[], Any]] = {
     # Use dedicated multiprocess context for workers.
     # Both spawn and fork work
     "VLLM_WORKER_MULTIPROC_METHOD":
-    lambda: os.getenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn"),
+    lambda: os.getenv("VLLM_WORKER_MULTIPROC_METHOD", "fork"),
 
     # Timeout for fetching images when serving multimodal models
     # Default is 5 seconds