[V1] Fix torch profiling for offline inference (#11125)

Signed-off-by: Roger Wang <ywang@roblox.com>

examples/offline_inference_with_profiler.py

@@ -1,4 +1,5 @@
 import os
+import time
 
 from vllm import LLM, SamplingParams
 
@@ -15,19 +16,25 @@ prompts = [
 # Create a sampling params object.
 sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
 
-# Create an LLM.
+if __name__ == "__main__":
-llm = LLM(model="facebook/opt-125m", tensor_parallel_size=1)
 
-llm.start_profile()
+    # Create an LLM.
+    llm = LLM(model="facebook/opt-125m", tensor_parallel_size=1)
 
-# Generate texts from the prompts. The output is a list of RequestOutput objects
+    llm.start_profile()
-# that contain the prompt, generated text, and other information.
-outputs = llm.generate(prompts, sampling_params)
 
-llm.stop_profile()
+    # Generate texts from the prompts. The output is a list of RequestOutput
+    # objects that contain the prompt, generated text, and other information.
+    outputs = llm.generate(prompts, sampling_params)
 
-# Print the outputs.
+    llm.stop_profile()
-for output in outputs:
+
+    # Print the outputs.
+    for output in outputs:
         prompt = output.prompt
         generated_text = output.outputs[0].text
         print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+
+    # Add a buffer to wait for profiler in the background process
+    # (in case MP is on) to finish writing profiling output.
+    time.sleep(10)

vllm/v1/engine/core_client.py

@@ -105,7 +105,7 @@ class InprocClient(EngineCoreClient):
     def __del__(self):
         self.shutdown()
 
-    async def profile(self, is_start=True) -> None:
+    def profile(self, is_start=True) -> None:
         self.engine_core.profile(is_start)
 
 
@@ -212,7 +212,7 @@ class SyncMPClient(MPClient):
     def abort_requests(self, request_ids: List[str]) -> None:
         self._send_input(EngineCoreRequestType.ABORT, request_ids)
 
-    async def profile(self, is_start=True) -> None:
+    def profile(self, is_start=True) -> None:
         self._send_input(EngineCoreRequestType.PROFILE,
                          EngineCoreProfile(is_start))