[Core] Introduce SPMD worker execution using Ray accelerated DAG

.buildkite/test-pipeline.yaml

@@ -84,6 +84,8 @@ steps:
   - 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
+  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray VLLM_USE_RAY_SPMD_WORKER=1 VLLM_USE_RAY_COMPILED_DAG=1 pytest -v -s distributed/test_basic_distributed_correctness.py
+  - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray VLLM_USE_RAY_SPMD_WORKER=1 VLLM_USE_RAY_COMPILED_DAG=1 pytest -v -s distributed/test_basic_distributed_correctness.py
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_chunked_prefill_distributed.py
   - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_chunked_prefill_distributed.py
   - TEST_DIST_MODEL=llava-hf/llava-1.5-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_multimodal_broadcast.py
@@ -108,6 +110,7 @@ steps:
   # 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.
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
+  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray VLLM_USE_RAY_SPMD_WORKER=1 VLLM_USE_RAY_COMPILED_DAG=1 pytest -v -s distributed/test_basic_distributed_correctness.py
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_basic_distributed_correctness.py
   - pytest -v -s spec_decode/e2e/test_integration_dist_tp4.py
 

vllm/engine/llm_engine.py

@@ -6,6 +6,7 @@
 
 from transformers import PreTrainedTokenizer
 
+import vllm.envs as envs
 from vllm.config import (CacheConfig, DecodingConfig, DeviceConfig, LoadConfig,
                          LoRAConfig, ModelConfig, MultiModalConfig,
                          ObservabilityConfig, ParallelConfig,
@@ -414,6 +415,9 @@ def from_engine_args(
         elif distributed_executor_backend == "mp":
             from vllm.executor.multiproc_gpu_executor import (
                 MultiprocessingGPUExecutor)
+            assert not envs.VLLM_USE_RAY_SPMD_WORKER, (
+                "multiprocessing distributed executor backend does not "
+                "support VLLM_USE_RAY_SPMD_WORKER=1")
             executor_class = MultiprocessingGPUExecutor
         else:
             from vllm.executor.gpu_executor import GPUExecutor
@@ -426,6 +430,7 @@ def from_engine_args(
             usage_context=usage_context,
             stat_loggers=stat_loggers,
         )
+
         return engine
 
     def __reduce__(self):

vllm/envs.py

@@ -34,6 +34,7 @@
     VLLM_OPENVINO_ENABLE_QUANTIZED_WEIGHTS: bool = False
     VLLM_XLA_CACHE_PATH: str = os.path.join(VLLM_CACHE_ROOT, "xla_cache")
     VLLM_FUSED_MOE_CHUNK_SIZE: int = 64 * 1024
+    VLLM_USE_RAY_SPMD_WORKER: bool = False
     VLLM_USE_RAY_COMPILED_DAG: bool = False
     VLLM_WORKER_MULTIPROC_METHOD: str = "fork"
     VLLM_ASSETS_CACHE: str = os.path.join(VLLM_CACHE_ROOT, "assets")
@@ -261,6 +262,13 @@ def get_default_config_root():
     "VLLM_OPENVINO_ENABLE_QUANTIZED_WEIGHTS":
     lambda: bool(os.getenv("VLLM_OPENVINO_ENABLE_QUANTIZED_WEIGHTS", False)),
 
+    # If the env var is set, then all workers will execute as separate
+    # processes from the engine, and we use the same mechanism to trigger
+    # execution on all workers.
+    # Run vLLM with VLLM_USE_RAY_SPMD_WORKER=1 to enable it.
+    "VLLM_USE_RAY_SPMD_WORKER":
+    lambda: bool(os.getenv("VLLM_USE_RAY_SPMD_WORKER", 0)),
+
     # If the env var is set, it uses the Ray's compiled DAG API
     # which optimizes the control plane overhead.
     # Run vLLM with VLLM_USE_RAY_COMPILED_DAG=1 to enable it.

vllm/executor/distributed_gpu_executor.py

@@ -64,16 +64,18 @@ def initialize_cache(self, num_gpu_blocks: int,
                           num_cpu_blocks=num_cpu_blocks)
 
     def execute_model(
-        self, execute_model_req: ExecuteModelRequest
-    ) -> Optional[List[SamplerOutput]]:
+            self,
+            execute_model_req: ExecuteModelRequest) -> List[SamplerOutput]:
         if self.parallel_worker_tasks is None:
             self.parallel_worker_tasks = self._run_workers(
                 "start_worker_execution_loop",
                 async_run_tensor_parallel_workers_only=True,
                 **self.extra_execute_model_run_workers_kwargs)
 
         # Only the driver worker returns the sampling results.
-        return self._driver_execute_model(execute_model_req)
+        driver_outputs = self._driver_execute_model(execute_model_req)
+        assert driver_outputs is not None
+        return driver_outputs
 
     def stop_remote_worker_execution_loop(self) -> None:
         if self.parallel_worker_tasks is None: