diff --git a/tests/entrypoints/openai/test_dp_supervisor.py b/tests/entrypoints/openai/test_dp_supervisor.py index 1dd6537f20..abe0cac890 100644 --- a/tests/entrypoints/openai/test_dp_supervisor.py +++ b/tests/entrypoints/openai/test_dp_supervisor.py @@ -176,6 +176,49 @@ def test_build_multi_port_external_lb_child_args_sets_external_rank_server(): assert child_args.api_server_count == 1 +def test_run_vllm_dp_server_uses_python_server_by_default(monkeypatch): + calls: list[str] = [] + + monkeypatch.setattr(dp_sup.os, "setpgrp", lambda: None) + monkeypatch.setattr(dp_sup, "set_process_title", lambda *_args: None) + monkeypatch.setattr(dp_sup, "decorate_logs", lambda *_args: None) + monkeypatch.setattr(dp_sup.envs, "VLLM_RUST_FRONTEND_PATH", None, raising=False) + monkeypatch.setattr( + dp_sup, "_run_python_vllm_dp_server", lambda _args: calls.append("python") + ) + monkeypatch.setattr( + dp_sup, "_run_rust_vllm_dp_server", lambda _args: calls.append("rust") + ) + + dp_sup._run_vllm_dp_server(_make_unit_args(data_parallel_rank=4)) + + assert calls == ["python"] + + +def test_run_vllm_dp_server_uses_rust_frontend_when_enabled(monkeypatch): + calls: list[str] = [] + + monkeypatch.setattr(dp_sup.os, "setpgrp", lambda: None) + monkeypatch.setattr(dp_sup, "set_process_title", lambda *_args: None) + monkeypatch.setattr(dp_sup, "decorate_logs", lambda *_args: None) + monkeypatch.setattr( + dp_sup.envs, + "VLLM_RUST_FRONTEND_PATH", + "/tmp/vllm-rs", + raising=False, + ) + monkeypatch.setattr( + dp_sup, "_run_python_vllm_dp_server", lambda _args: calls.append("python") + ) + monkeypatch.setattr( + dp_sup, "_run_rust_vllm_dp_server", lambda _args: calls.append("rust") + ) + + dp_sup._run_vllm_dp_server(_make_unit_args(data_parallel_rank=4)) + + assert calls == ["rust"] + + def test_validate_multi_port_external_lb_args_allows_ssl(): args = _make_unit_args( ssl_keyfile="/tmp/server.key", diff --git a/vllm/entrypoints/openai/dp_supervisor.py b/vllm/entrypoints/openai/dp_supervisor.py index 73b10a04ea..f55f11f81d 100644 --- a/vllm/entrypoints/openai/dp_supervisor.py +++ b/vllm/entrypoints/openai/dp_supervisor.py @@ -22,6 +22,7 @@ import uvicorn import uvloop from fastapi import FastAPI, Response +import vllm.envs as envs from vllm.logger import init_logger from vllm.utils.system_utils import ( decorate_logs, @@ -233,12 +234,22 @@ def _build_dp_supervisor_app(supervisor: DPSupervisor) -> FastAPI: return app +def _run_python_vllm_dp_server(child_args: argparse.Namespace) -> None: + from vllm.entrypoints.openai.api_server import run_server + + uvloop.run(run_server(child_args)) + + +def _run_rust_vllm_dp_server(child_args: argparse.Namespace) -> None: + from vllm.entrypoints.cli.serve import run_multi_api_server + + run_multi_api_server(child_args) + + def _run_vllm_dp_server(child_args: argparse.Namespace) -> None: """ Entrypoint function for the vLLM DP Server. """ - from vllm.entrypoints.openai.api_server import run_server - # Create a fresh process group for the vLLM DP Server, # so that CTRL-C is propagated cleanly. os.setpgrp() @@ -246,7 +257,10 @@ def _run_vllm_dp_server(child_args: argparse.Namespace) -> None: name = f"APIServer_DP{child_args.data_parallel_rank}" set_process_title(name) decorate_logs(name) - uvloop.run(run_server(child_args)) + if envs.VLLM_RUST_FRONTEND_PATH: + _run_rust_vllm_dp_server(child_args) + else: + _run_python_vllm_dp_server(child_args) class DPSupervisor: