[V1] AsyncLLM data parallel

.buildkite/test-pipeline.yaml

@@ -135,12 +135,14 @@ steps:
   - examples/offline_inference/rlhf.py
   - examples/offline_inference/rlhf_colocate.py
   - tests/examples/offline_inference/data_parallel.py
+  - tests/v1/test_async_llm_dp.py
   commands:
   # test with tp=2 and external_dp=2
   - VLLM_USE_V1=0 torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
   - torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
   # test with internal dp
   - python3 ../examples/offline_inference/data_parallel.py
+  - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
   - pytest -v -s distributed/test_utils.py
   - pytest -v -s compile/test_basic_correctness.py
   - pytest -v -s distributed/test_pynccl.py
@@ -514,7 +516,10 @@ steps:
   - vllm/worker/worker.py
   - vllm/worker/model_runner.py
   - entrypoints/llm/test_collective_rpc.py
+  - tests/v1/test_async_llm_dp.py
+  - vllm/v1/engine/
   commands:
+  - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
   - VLLM_ENABLE_V1_MULTIPROCESSING=0 pytest -v -s entrypoints/llm/test_collective_rpc.py
   - pytest -v -s ./compile/test_basic_correctness.py
   - pytest -v -s ./compile/test_wrapper.py

examples/offline_inference/data_parallel.py

@@ -4,6 +4,7 @@
 # we need to have a launcher to create multiple data parallel
 # ranks. And each rank will create a vLLM instance to process its own prompts.
 import os
+from time import sleep
 
 from vllm import LLM, SamplingParams
 from vllm.utils import get_open_port
@@ -17,10 +18,9 @@ def main(dp_size, dp_rank, dp_master_ip, dp_master_port, GPUs_per_dp_rank):
     os.environ["VLLM_DP_SIZE"] = str(dp_size)
     os.environ["VLLM_DP_MASTER_IP"] = dp_master_ip
     os.environ["VLLM_DP_MASTER_PORT"] = str(dp_master_port)
-    # set devices for each dp_rank
-    os.environ["CUDA_VISIBLE_DEVICES"] = ",".join(
-        str(i) for i in range(dp_rank * GPUs_per_dp_rank, (dp_rank + 1) *
-                              GPUs_per_dp_rank))
+
+    # CUDA_VISIBLE_DEVICES for each DP rank is set automatically inside the
+    # engine processes.
 
     # Sample prompts.
     prompts = [
@@ -64,6 +64,9 @@ def main(dp_size, dp_rank, dp_master_ip, dp_master_port, GPUs_per_dp_rank):
         print(f"DP rank {dp_rank}, Prompt: {prompt!r}, "
               f"Generated text: {generated_text!r}")
 
+    # Give engines time to pause their processing loops before exiting.
+    sleep(1)
+
 
 if __name__ == "__main__":
     from multiprocessing import Process
@@ -78,8 +81,13 @@ def main(dp_size, dp_rank, dp_master_ip, dp_master_port, GPUs_per_dp_rank):
         procs.append(proc)
     exit_code = 0
     for proc in procs:
-        proc.join()
-        if proc.exitcode:
+        proc.join(timeout=300)
+        if proc.exitcode is None:
+            print(f"Killing process {proc.pid} that "
+                  f"didn't stop within 5 minutes.")
+            proc.kill()
+            exit_code = 1
+        elif proc.exitcode:
             exit_code = proc.exitcode
 
     exit(exit_code)

tests/v1/engine/test_engine_core_client.py

@@ -167,11 +167,11 @@ def test_engine_core_client(monkeypatch: pytest.MonkeyPatch,
 
             core_client: SyncMPClient = client
 
-            result = core_client._call_utility("echo", "testarg")
+            result = core_client.call_utility("echo", "testarg")
             assert result == "testarg"
 
             with pytest.raises(Exception) as e_info:
-                core_client._call_utility("echo", None, "help!")
+                core_client.call_utility("echo", None, "help!")
 
             assert str(e_info.value) == "Call to echo method failed: help!"
 
@@ -238,10 +238,10 @@ async def test_engine_core_client_asyncio(monkeypatch: pytest.MonkeyPatch):
 
         core_client: AsyncMPClient = client
 
-        result = await core_client._call_utility_async("echo", "testarg")
+        result = await core_client.call_utility_async("echo", "testarg")
         assert result == "testarg"
 
         with pytest.raises(Exception) as e_info:
-            await core_client._call_utility_async("echo", None, "help!")
+            await core_client.call_utility_async("echo", None, "help!")
 
         assert str(e_info.value) == "Call to echo method failed: help!"

tests/v1/test_async_llm_dp.py

@@ -0,0 +1,108 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import asyncio
+import os
+from contextlib import ExitStack
+from typing import Optional
+
+import pytest
+
+from vllm import SamplingParams
+from vllm.engine.arg_utils import AsyncEngineArgs
+from vllm.inputs import PromptType
+from vllm.platforms import current_platform
+from vllm.sampling_params import RequestOutputKind
+from vllm.v1.engine.async_llm import AsyncLLM
+from vllm.v1.engine.core_client import DPAsyncMPClient
+
+if not current_platform.is_cuda():
+    pytest.skip(reason="V1 currently only supported on CUDA.",
+                allow_module_level=True)
+
+
+async def generate(engine: AsyncLLM,
+                   request_id: str,
+                   prompt: PromptType,
+                   output_kind: RequestOutputKind,
+                   max_tokens: int,
+                   prompt_logprobs: Optional[int] = None) -> tuple[int, str]:
+    # Ensure generate doesn't complete too fast for cancellation test.
+    await asyncio.sleep(0.2)
+
+    count = 0
+    sampling_params = SamplingParams(max_tokens=max_tokens,
+                                     ignore_eos=True,
+                                     output_kind=output_kind,
+                                     temperature=0,
+                                     prompt_logprobs=prompt_logprobs)
+    async for out in engine.generate(request_id=request_id,
+                                     prompt=prompt,
+                                     sampling_params=sampling_params):
+
+        num_tokens = len(out.outputs[0].token_ids)
+        if output_kind == RequestOutputKind.DELTA:
+            count += num_tokens
+        else:
+            count = num_tokens
+
+        await asyncio.sleep(0.)
+
+    return count, request_id
+
+
+@pytest.mark.parametrize(
+    "output_kind", [RequestOutputKind.DELTA, RequestOutputKind.FINAL_ONLY])
+@pytest.mark.asyncio
+async def test_load(monkeypatch, output_kind: RequestOutputKind):
+    with monkeypatch.context() as m, ExitStack() as after:
+        m.setenv("VLLM_USE_V1", "1")
+
+        engine_args = AsyncEngineArgs(
+            model="ibm-research/PowerMoE-3b",
+            enforce_eager=True,
+            disable_log_requests=True,
+            tensor_parallel_size=int(os.getenv("TP_SIZE", 1)),
+            data_parallel_size=int(os.getenv("DP_SIZE", 2)),
+        )
+
+        prompt = "This is a test of data parallel"
+
+        engine = AsyncLLM.from_engine_args(engine_args)
+        after.callback(engine.shutdown)
+
+        NUM_REQUESTS = 100
+        NUM_EXPECTED_TOKENS = 10
+
+        request_ids = [f"request-{i}" for i in range(NUM_REQUESTS)]
+
+        # Create concurrent requests.
+        tasks = []
+        for request_id in request_ids:
+            tasks.append(
+                asyncio.create_task(
+                    generate(engine, request_id, prompt, output_kind,
+                             NUM_EXPECTED_TOKENS)))
+
+        # Confirm that we got all the EXPECTED tokens from the requests.
+        done, pending = await asyncio.wait(tasks,
+                                           return_when=asyncio.FIRST_EXCEPTION)
+        for task in pending:
+            task.cancel()
+        for task in done:
+            num_generated_tokens, request_id = await task
+            assert num_generated_tokens == NUM_EXPECTED_TOKENS, (
+                f"{request_id} generated {num_generated_tokens} but "
+                f"expected {NUM_EXPECTED_TOKENS}")
+
+        assert not engine.output_processor.has_unfinished_requests()
+
+        # testing internals here which may break
+        core_client: DPAsyncMPClient = engine.engine_core
+        # the engines only synchronize stopping every N steps so
+        # allow a small amount of time here.
+        for _ in range(10):
+            if core_client.num_engines_running == 0:
+                break
+            await asyncio.sleep(0.5)
+        assert core_client.num_engines_running == 0
+        assert not core_client.reqs_in_flight

vllm/config.py

@@ -40,7 +40,8 @@
 from vllm.transformers_utils.s3_utils import S3Model
 from vllm.transformers_utils.utils import is_s3
 from vllm.utils import (GiB_bytes, LayerBlockType, cuda_device_count_stateless,
-                        get_cpu_memory, random_uuid, resolve_obj_by_qualname)
+                        get_cpu_memory, get_open_port, random_uuid,
+                        resolve_obj_by_qualname)
 
 if TYPE_CHECKING:
     from ray.util.placement_group import PlacementGroup
@@ -1469,10 +1470,17 @@ def __post_init__(self) -> None:
         self.world_size = self.pipeline_parallel_size * \
             self.tensor_parallel_size
 
-        self.data_parallel_size = envs.VLLM_DP_SIZE
-        self.data_parallel_rank = envs.VLLM_DP_RANK
-        self.data_parallel_master_ip = envs.VLLM_DP_MASTER_IP
-        self.data_parallel_master_port = envs.VLLM_DP_MASTER_PORT
+        if self.data_parallel_size > 1:
+            # Data parallel was specified in the engine args.
+            self.data_parallel_master_port = get_open_port()
+            # TODO multi-node
+        else:
+            # Otherwise fall back to env vars (e.g. for offline SPMD case).
+            self.data_parallel_size = envs.VLLM_DP_SIZE
+            self.data_parallel_rank = envs.VLLM_DP_RANK
+            self.data_parallel_master_ip = envs.VLLM_DP_MASTER_IP
+            self.data_parallel_master_port = envs.VLLM_DP_MASTER_PORT
+
         self.world_size_across_dp = self.world_size * self.data_parallel_size
 
         if self.distributed_executor_backend == "external_launcher":

vllm/distributed/utils.py

@@ -15,6 +15,8 @@
 from torch.distributed import ProcessGroup, TCPStore
 from torch.distributed.distributed_c10d import (Backend, PrefixStore,
                                                 _get_default_timeout,
+                                                _shutdown_backend,
+                                                _unregister_process_group,
                                                 is_nccl_available)
 from torch.distributed.rendezvous import rendezvous
 
@@ -333,3 +335,13 @@ def stateless_init_torch_distributed_process_group(
     pg._register_backend(device, backend_type, backend_class)
 
     return pg
+
+
+def stateless_destroy_torch_distributed_process_group(
+        pg: ProcessGroup) -> None:
+    """
+    Destroy ProcessGroup returned by
+        stateless_init_torch_distributed_process_group().
+    """
+    _shutdown_backend(pg)
+    _unregister_process_group(pg.group_name)

vllm/engine/arg_utils.py

@@ -114,6 +114,7 @@ class EngineArgs:
     # number of P/D disaggregation (or other disaggregation) workers
     pipeline_parallel_size: int = 1
     tensor_parallel_size: int = 1
+    data_parallel_size: int = 1
     enable_expert_parallel: bool = False
     max_parallel_loading_workers: Optional[int] = None
     block_size: Optional[int] = None
@@ -441,6 +442,14 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
                             type=int,
                             default=EngineArgs.tensor_parallel_size,
                             help='Number of tensor parallel replicas.')
+        parser.add_argument('--data-parallel-size',
+                            '-dp',
+                            type=int,
+                            default=EngineArgs.data_parallel_size,
+                            help='Number of data parallel replicas. '
+                            'MoE layers will be sharded according to the '
+                            'product of the tensor-parallel-size and '
+                            'data-parallel-size.')
         parser.add_argument(
             '--enable-expert-parallel',
             action='store_true',
@@ -1347,6 +1356,7 @@ def create_engine_config(
         parallel_config = ParallelConfig(
             pipeline_parallel_size=self.pipeline_parallel_size,
             tensor_parallel_size=self.tensor_parallel_size,
+            data_parallel_size=self.data_parallel_size,
             enable_expert_parallel=self.enable_expert_parallel,
             max_parallel_loading_workers=self.max_parallel_loading_workers,
             disable_custom_all_reduce=self.disable_custom_all_reduce,

vllm/utils.py

@@ -527,7 +527,7 @@ def get_open_port() -> int:
         dp_port = envs.VLLM_DP_MASTER_PORT
         while True:
             port = _get_open_port()
-            if port >= dp_port and port < dp_port + 10:
+            if dp_port <= port < dp_port + 10:
                 continue
             return port
     return _get_open_port()
@@ -2122,19 +2122,23 @@ def make_zmq_socket(
     if socket_type == zmq.constants.PULL:
         socket.setsockopt(zmq.constants.RCVHWM, 0)
         socket.setsockopt(zmq.constants.RCVBUF, buf_size)
-        socket.connect(path)
+        socket.bind(path)
     elif socket_type == zmq.constants.PUSH:
         socket.setsockopt(zmq.constants.SNDHWM, 0)
         socket.setsockopt(zmq.constants.SNDBUF, buf_size)
-        socket.bind(path)
+        socket.connect(path)
     else:
         raise ValueError(f"Unknown Socket Type: {socket_type}")
 
     return socket
 
 
 @contextlib.contextmanager
-def zmq_socket_ctx(path: str, socket_type: Any) -> Iterator[zmq.Socket]:
+def zmq_socket_ctx(
+    path: str,
+    socket_type: Any,
+    linger: int = 0,
+) -> Iterator[zmq.Socket]:
     """Context manager for a ZMQ socket"""
 
     ctx = zmq.Context()  # type: ignore[attr-defined]
@@ -2145,7 +2149,7 @@ def zmq_socket_ctx(path: str, socket_type: Any) -> Iterator[zmq.Socket]:
         logger.debug("Got Keyboard Interrupt.")
 
     finally:
-        ctx.destroy(linger=0)
+        ctx.destroy(linger=linger)
 
 
 def is_in_ray_actor():

vllm/v1/core/sched/scheduler.py

@@ -36,13 +36,15 @@ def __init__(
         cache_config: CacheConfig,
         lora_config: Optional[LoRAConfig],
         speculative_config: Optional[SpeculativeConfig],
-        log_stats: bool,
         structured_output_manager: StructuredOutputManager,
+        include_finished_set: bool = False,
+        log_stats: bool = False,
     ) -> None:
         self.scheduler_config = scheduler_config
         self.cache_config = cache_config
         self.lora_config = lora_config
         self.speculative_config = speculative_config
+        self.include_finished_set = include_finished_set
         self.log_stats = log_stats
         self.structured_output_manager = structured_output_manager
 
@@ -647,10 +649,16 @@ def update_from_output(
                 new_running.append(request)
 
         self.running = new_running
-        return EngineCoreOutputs(
+        engine_core_outputs = EngineCoreOutputs(
             outputs=outputs,
             scheduler_stats=self.make_stats(),
         )
+        if self.include_finished_set:
+            #TODO currently sending duplicates here, improve this
+            engine_core_outputs.finished_requests = (
+                scheduler_output.finished_req_ids | self.finished_req_ids)
+
+        return engine_core_outputs
 
     def add_request(self, request: Request) -> None:
         self.waiting.append(request)