Default vLLM mode to native instead of docker sidecar

lib/marin/src/marin/inference/vllm_server.py

@@ -236,7 +236,11 @@ def stop(self, handle: VllmServerHandle) -> None:
 
 
 def resolve_vllm_mode(mode: Literal["native", "docker"] | None) -> Literal["native", "docker"]:
-    mode_str = (mode if mode is not None else os.environ.get("MARIN_VLLM_MODE", "docker")).lower()
+    # Default to native vLLM. The Docker sidecar path requires a mounted
+    # /var/run/docker.sock (docker-alongside-docker), which Iris workers do not
+    # provide. Set MARIN_VLLM_MODE=docker to opt in for Ray-era flows that still
+    # need the sidecar.
+    mode_str = (mode if mode is not None else os.environ.get("MARIN_VLLM_MODE", "native")).lower()
     if mode_str not in ("native", "docker"):
         raise ValueError(f"Unknown MARIN_VLLM_MODE={mode_str!r}; expected 'native' or 'docker'.")
     return mode_str  # type: ignore[return-value]

lib/marin/src/marin/inference/vllm_smoke_test.py

@@ -137,7 +137,7 @@ def main(argv: list[str] | None = None) -> int:
         "--mode",
         choices=["docker", "native"],
         default=None,
-        help="Override MARIN_VLLM_MODE (default: use env; docker if unset).",
+        help="Override MARIN_VLLM_MODE (default: use env; native if unset).",
     )
     parser.add_argument(
         "--docker-image",

tests/vllm/test_vllm_server.py

@@ -0,0 +1,30 @@
+# Copyright The Marin Authors
+# SPDX-License-Identifier: Apache-2.0
+
+"""Tests for marin.inference.vllm_server helpers (do not require vllm)."""
+
+import pytest
+
+from marin.inference.vllm_server import resolve_vllm_mode
+
+
+def test_default_mode_is_native(monkeypatch):
+    """Default mode must be native so Iris workers without /var/run/docker.sock work out of the box."""
+    monkeypatch.delenv("MARIN_VLLM_MODE", raising=False)
+    assert resolve_vllm_mode(None) == "native"
+
+
+def test_env_var_can_select_docker(monkeypatch):
+    monkeypatch.setenv("MARIN_VLLM_MODE", "docker")
+    assert resolve_vllm_mode(None) == "docker"
+
+
+def test_explicit_argument_overrides_env(monkeypatch):
+    monkeypatch.setenv("MARIN_VLLM_MODE", "docker")
+    assert resolve_vllm_mode("native") == "native"
+
+
+def test_unknown_mode_raises(monkeypatch):
+    monkeypatch.setenv("MARIN_VLLM_MODE", "kubernetes")
+    with pytest.raises(ValueError, match="MARIN_VLLM_MODE"):
+        resolve_vllm_mode(None)