Allow AsyncLLMEngine.generate to target a specific DP rank (#19102)

Signed-off-by: Jon Swenson <jmswen@gmail.com>

examples/online_serving/multi_instance_data_parallel.py

@@ -0,0 +1,58 @@
+# SPDX-License-Identifier: Apache-2.0
+import asyncio
+from typing import Optional
+
+from vllm.engine.arg_utils import AsyncEngineArgs
+from vllm.engine.async_llm_engine import AsyncLLMEngine
+from vllm.outputs import RequestOutput
+from vllm.sampling_params import SamplingParams
+
+"""
+To run this example, run the following commands simultaneously with
+different CUDA_VISIBLE_DEVICES:
+    python examples/online_serving/multi_instance_data_parallel.py
+
+    vllm serve ibm-research/PowerMoE-3b -dp 2 -dpr 1 \
+        --data-parallel-address 127.0.0.1 --data-parallel-rpc-port 62300 \
+        --data-parallel-size-local 1 --enforce-eager --headless
+
+Once both instances have completed the handshake, this example will
+send a request to the instance with DP rank 1.
+"""
+
+
+async def main():
+    engine_args = AsyncEngineArgs(
+        model="ibm-research/PowerMoE-3b",
+        data_parallel_size=2,
+        dtype="auto",
+        max_model_len=2048,
+        data_parallel_address="127.0.0.1",
+        data_parallel_rpc_port=62300,
+        data_parallel_size_local=1,
+        enforce_eager=True,
+    )
+
+    engine_client = AsyncLLMEngine.from_engine_args(engine_args)
+
+    sampling_params = SamplingParams(
+        temperature=0.7,
+        top_p=0.9,
+        max_tokens=100,
+    )
+
+    prompt = "Who won the 2004 World Series?"
+    final_output: Optional[RequestOutput] = None
+    async for output in engine_client.generate(
+        prompt=prompt,
+        sampling_params=sampling_params,
+        request_id="abcdef",
+        data_parallel_rank=1,
+    ):
+        final_output = output
+    if final_output:
+        print(final_output.outputs[0].text)
+
+
+if __name__ == "__main__":
+    asyncio.run(main())

tests/tokenization/test_detokenize.py

@@ -70,7 +70,8 @@ def _run_incremental_decode(tokenizer,
                                 None,
                                 0.0,
                                 None,
-                                cache_salt=None)
+                                cache_salt=None,
+                                data_parallel_rank=None)
 
     if fast is None:
         detokenizer = IncrementalDetokenizer.from_new_request(

tests/v1/engine/test_engine_core.py

@@ -42,6 +42,7 @@ def make_request() -> EngineCoreRequest:
         arrival_time=time.time(),
         lora_request=None,
         cache_salt=None,
+        data_parallel_rank=None,
     )
 
 

tests/v1/engine/test_engine_core_client.py

@@ -56,6 +56,7 @@ def make_request(
         arrival_time=time.time(),
         lora_request=None,
         cache_salt=None,
+        data_parallel_rank=None,
     )
 
 

tests/v1/engine/test_output_processor.py

@@ -59,6 +59,7 @@ def test_incremental_detokenization(request_output_kind: RequestOutputKind,
                           eos_token_id=None,
                           lora_request=None,
                           cache_salt=None,
+                          data_parallel_rank=None,
                           sampling_params=SamplingParams(
                               skip_special_tokens=False,
                               spaces_between_special_tokens=False,
@@ -406,6 +407,7 @@ def test_logprobs_processor(request_output_kind: RequestOutputKind,
                           eos_token_id=None,
                           lora_request=None,
                           cache_salt=None,
+                          data_parallel_rank=None,
                           sampling_params=SamplingParams(
                               skip_special_tokens=False,
                               spaces_between_special_tokens=False,
@@ -569,6 +571,7 @@ def test_stop_token(include_stop_str_in_output: bool,
         eos_token_id=eos_token_id,
         lora_request=None,
         cache_salt=None,
+        data_parallel_rank=None,
         sampling_params=SamplingParams(
             skip_special_tokens=False,
             spaces_between_special_tokens=False,
@@ -666,6 +669,7 @@ def test_stop_string(include_stop_str_in_output: bool,
             eos_token_id=None,
             lora_request=None,
             cache_salt=None,
+            data_parallel_rank=None,
             sampling_params=SamplingParams(
                 skip_special_tokens=False,
                 spaces_between_special_tokens=False,
@@ -780,6 +784,7 @@ def test_iteration_stats(dummy_test_vectors):
             eos_token_id=None,
             lora_request=None,
             cache_salt=None,
+            data_parallel_rank=None,
             sampling_params=SamplingParams(),
         ) for idx, prompt_tokens in enumerate(dummy_test_vectors.prompt_tokens)
     ]

vllm/engine/async_llm_engine.py

@@ -442,6 +442,7 @@ class _AsyncLLMEngine(LLMEngine):
         trace_headers: Optional[Mapping[str, str]] = None,
         prompt_adapter_request: Optional[PromptAdapterRequest] = None,
         priority: int = 0,
+        data_parallel_rank: Optional[int] = None,
     ) -> None:
         ...
 
@@ -456,6 +457,7 @@ class _AsyncLLMEngine(LLMEngine):
         trace_headers: Optional[Mapping[str, str]] = None,
         prompt_adapter_request: Optional[PromptAdapterRequest] = None,
         priority: int = 0,
+        data_parallel_rank: Optional[int] = None,
     ) -> None:
         ...
 
@@ -473,6 +475,7 @@ class _AsyncLLMEngine(LLMEngine):
             trace_headers: Optional[Mapping[str, str]] = None,
             prompt_adapter_request: Optional[PromptAdapterRequest] = None,
             priority: int = 0,
+            data_parallel_rank: Optional[int] = None,
             *,
             inputs: Optional[PromptType] = None,  # DEPRECATED
     ) -> None:
@@ -902,6 +905,7 @@ class AsyncLLMEngine(EngineClient):
         trace_headers: Optional[Mapping[str, str]] = None,
         prompt_adapter_request: Optional[PromptAdapterRequest] = None,
         priority: int = 0,
+        data_parallel_rank: Optional[int] = None,
     ) -> Coroutine[None, None, AsyncGenerator[Union[
             RequestOutput, PoolingRequestOutput], None]]:
         ...
@@ -917,6 +921,7 @@ class AsyncLLMEngine(EngineClient):
         trace_headers: Optional[Mapping[str, str]] = None,
         prompt_adapter_request: Optional[PromptAdapterRequest] = None,
         priority: int = 0,
+        data_parallel_rank: Optional[int] = None,
     ) -> Coroutine[None, None, AsyncGenerator[Union[
             RequestOutput, PoolingRequestOutput], None]]:
         ...
@@ -935,6 +940,7 @@ class AsyncLLMEngine(EngineClient):
         trace_headers: Optional[Mapping[str, str]] = None,
         prompt_adapter_request: Optional[PromptAdapterRequest] = None,
         priority: int = 0,
+        data_parallel_rank: Optional[int] = None,
         *,
         inputs: Optional[PromptType] = None,  # DEPRECATED
     ) -> AsyncGenerator[Union[RequestOutput, PoolingRequestOutput], None]:
@@ -967,6 +973,7 @@ class AsyncLLMEngine(EngineClient):
             trace_headers=trace_headers,
             prompt_adapter_request=prompt_adapter_request,
             priority=priority,
+            data_parallel_rank=data_parallel_rank,
         )
 
         return stream.generator()
@@ -980,6 +987,7 @@ class AsyncLLMEngine(EngineClient):
         trace_headers: Optional[Mapping[str, str]] = None,
         prompt_adapter_request: Optional[PromptAdapterRequest] = None,
         priority: int = 0,
+        data_parallel_rank: Optional[int] = None,
     ) -> AsyncGenerator[RequestOutput, None]:
         """Generate outputs for a request.
 
@@ -999,7 +1007,8 @@ class AsyncLLMEngine(EngineClient):
                                             for generation, if any.
             priority: The priority of the request.
                 Only applicable with priority scheduling.
-
+            data_parallel_rank: The (global) data parallel rank that must
+                handle this request. Only applicable if DP is enabled.
         Yields:
             The output `RequestOutput` objects from the LLMEngine
             for the request.
@@ -1057,6 +1066,7 @@ class AsyncLLMEngine(EngineClient):
                     trace_headers=trace_headers,
                     prompt_adapter_request=prompt_adapter_request,
                     priority=priority,
+                    data_parallel_rank=data_parallel_rank,
             ):
                 yield LLMEngine.validate_output(output, RequestOutput)
         except asyncio.CancelledError:

vllm/v1/engine/__init__.py

@@ -55,6 +55,7 @@ class EngineCoreRequest(
     arrival_time: float
     lora_request: Optional[LoRARequest]
     cache_salt: Optional[str]
+    data_parallel_rank: Optional[int]
 
     # Index of the client, used to ensure outputs are sent back to the same
     # client for this request when scaling out the front-end.

vllm/v1/engine/async_llm.py

@@ -229,6 +229,7 @@ class AsyncLLM(EngineClient):
         trace_headers: Optional[Mapping[str, str]] = None,
         prompt_adapter_request: Optional[PromptAdapterRequest] = None,
         priority: int = 0,
+        data_parallel_rank: Optional[int] = None,
     ) -> RequestOutputCollector:
         """Add new request to the AsyncLLM."""
 
@@ -245,7 +246,7 @@ class AsyncLLM(EngineClient):
         prompt_str, request = self.processor.process_inputs(
             request_id, prompt, params, arrival_time, lora_request,
             tokenization_kwargs, trace_headers, prompt_adapter_request,
-            priority)
+            priority, data_parallel_rank)
 
         if params.n == 1:
             await self._add_request(request, prompt_str, None, 0, queue)
@@ -291,6 +292,7 @@ class AsyncLLM(EngineClient):
         trace_headers: Optional[Mapping[str, str]] = None,
         prompt_adapter_request: Optional[PromptAdapterRequest] = None,
         priority: int = 0,
+        data_parallel_rank: Optional[int] = None,
     ) -> AsyncGenerator[RequestOutput, None]:
         """
         Main function called by the API server to kick off a request
@@ -321,6 +323,7 @@ class AsyncLLM(EngineClient):
                 trace_headers=trace_headers,
                 prompt_adapter_request=prompt_adapter_request,
                 priority=priority,
+                data_parallel_rank=data_parallel_rank,
             )
 
             # The output_handler task pushes items into the queue.

vllm/v1/engine/core_client.py

@@ -982,7 +982,16 @@ class DPAsyncMPClient(AsyncMPClient):
         resources.stats_update_task = asyncio.create_task(
             run_engine_stats_update_task())
 
-    def get_core_engine_for_request(self) -> CoreEngine:
+    def get_core_engine_for_request(self,
+                                    dp_rank: Optional[int] = None
+                                    ) -> CoreEngine:
+        if dp_rank is not None:
+            # engines are already in rank order
+            if dp_rank < 0 or dp_rank >= len(self.core_engines):
+                raise ValueError(f"Requested DP rank {dp_rank} is out of "
+                                 f"range [0, {len(self.core_engines)})")
+            return self.core_engines[dp_rank]
+
         if not self.lb_engines:
             return self.core_engines[0]
         # TODO use P2C alg for larger DP sizes
@@ -1018,7 +1027,8 @@ class DPAsyncMPClient(AsyncMPClient):
         request.current_wave = self.current_wave
         request.client_index = self.client_index
 
-        chosen_engine = self.get_core_engine_for_request()
+        chosen_engine = self.get_core_engine_for_request(
+            request.data_parallel_rank)
         self.reqs_in_flight[request.request_id] = chosen_engine
 
         to_await = self._send_input(EngineCoreRequestType.ADD, request,

vllm/v1/engine/processor.py

@@ -212,6 +212,7 @@ class Processor:
         trace_headers: Optional[Mapping[str, str]] = None,
         prompt_adapter_request: Optional[PromptAdapterRequest] = None,
         priority: int = 0,
+        data_parallel_rank: Optional[int] = None,
     ) -> tuple[Optional[str], EngineCoreRequest]:
 
         # TODO(woosuk): Support pooling models.
@@ -328,6 +329,7 @@ class Processor:
             arrival_time=arrival_time,
             lora_request=lora_request,
             cache_salt=decoder_inputs.get("cache_salt"),
+            data_parallel_rank=data_parallel_rank,
         )
 
     def _validate_model_inputs(self,