[BugFix] Fix duplicate id tool-call race condition (#29355)

Signed-off-by: Nick Hill <nhill@redhat.com>

vllm/entrypoints/openai/serving_chat.py

@@ -273,6 +273,11 @@ class OpenAIServingChat(OpenAIServing):
         try:
             for i, engine_prompt in enumerate(engine_prompts):
                 prompt_text, _, _ = self._get_prompt_components(request_prompts[i])
+                # If we are creating sub requests for multiple prompts, ensure that they
+                # have unique request ids.
+                sub_request_id = (
+                    request_id if len(engine_prompts) == 1 else f"{request_id}_{i}"
+                )
 
                 if self.default_sampling_params is None:
                     self.default_sampling_params = {}
@@ -301,7 +306,7 @@ class OpenAIServingChat(OpenAIServing):
                     )
 
                 self._log_inputs(
-                    request_id,
+                    sub_request_id,
                     request_prompts[i],
                     params=sampling_params,
                     lora_request=lora_request,
@@ -316,14 +321,14 @@ class OpenAIServingChat(OpenAIServing):
                 if isinstance(sampling_params, BeamSearchParams):
                     generator = self.beam_search(
                         prompt=engine_prompt,
-                        request_id=request_id,
+                        request_id=sub_request_id,
                         params=sampling_params,
                         lora_request=lora_request,
                         trace_headers=trace_headers,
                     )
                 else:
                     engine_request, tokenization_kwargs = await self._process_inputs(
-                        request_id,
+                        sub_request_id,
                         engine_prompt,
                         sampling_params,
                         lora_request=lora_request,
@@ -334,7 +339,7 @@ class OpenAIServingChat(OpenAIServing):
                     generator = self.engine_client.generate(
                         engine_request,
                         sampling_params,
-                        request_id,
+                        sub_request_id,
                         lora_request=lora_request,
                         trace_headers=trace_headers,
                         priority=request.priority,

vllm/entrypoints/openai/serving_engine.py

@@ -1242,16 +1242,19 @@ class OpenAIServing:
     ):
         prompt_text, _, _ = self._get_prompt_components(request_prompt)
         orig_priority = priority
+        sub_request = 0
         while True:
+            # Ensure that each sub-request has a unique request id.
+            sub_request_id = f"{request_id}_{sub_request}"
             self._log_inputs(
-                request_id,
+                sub_request_id,
                 request_prompt,
                 params=sampling_params,
                 lora_request=lora_request,
             )
             trace_headers = kwargs.get("trace_headers")
             engine_request, tokenization_kwargs = await self._process_inputs(
-                request_id,
+                sub_request_id,
                 engine_prompt,
                 sampling_params,
                 lora_request=lora_request,
@@ -1262,7 +1265,7 @@ class OpenAIServing:
             generator = self.engine_client.generate(
                 engine_request,
                 sampling_params,
-                request_id,
+                sub_request_id,
                 lora_request=lora_request,
                 priority=priority,
                 prompt_text=prompt_text,
@@ -1295,6 +1298,7 @@ class OpenAIServing:
             sampling_params.max_tokens = self.max_model_len - len(prompt_token_ids)
             # OPTIMIZATION
             priority = orig_priority - 1
+            sub_request += 1
 
     def _get_prompt_components(
         self,