[ROCm] Derive device capability from GCN arch string without CUDA init (#35069)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>

vllm/utils/system_utils.py

@@ -16,6 +16,7 @@ import psutil
 
 import vllm.envs as envs
 from vllm.logger import init_logger
+from vllm.platforms import current_platform
 from vllm.platforms.interface import in_wsl
 from vllm.ray.lazy_utils import is_in_ray_actor
 
@@ -111,6 +112,17 @@ def unique_filepath(fn: Callable[[int], Path]) -> Path:
 # Process management utilities
 
 
+def _sync_visible_devices_env_vars():
+    """Sync HIP/CUDA visibility env vars before spawning (ROCm only)."""
+
+    if not current_platform.is_rocm():
+        return
+
+    from vllm.platforms.rocm import _sync_hip_cuda_env_vars
+
+    _sync_hip_cuda_env_vars()
+
+
 def _maybe_force_spawn():
     """Check if we need to force the use of the `spawn` multiprocessing start
     method.
@@ -156,6 +168,10 @@ def get_mp_context():
     VLLM_WORKER_MULTIPROC_METHOD.
     """
     _maybe_force_spawn()
+    # (ROCm): Sync GPU visibility env vars so spawned children inherit
+    # consistent values. Must run after _maybe_force_spawn and regardless
+    # of whether spawn was already set.
+    _sync_visible_devices_env_vars()
     mp_method = envs.VLLM_WORKER_MULTIPROC_METHOD
     return multiprocessing.get_context(mp_method)