[Fix] Improve ROCm detection in WSL environments (#38434)

Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>

vllm/platforms/__init__.py

@@ -10,7 +10,7 @@ from vllm import envs
 from vllm.plugins import PLATFORM_PLUGINS_GROUP, load_plugins_by_group
 from vllm.utils.import_utils import resolve_obj_by_qualname
 
-from .interface import CpuArchEnum, Platform, PlatformEnum
+from .interface import CpuArchEnum, Platform, PlatformEnum, in_wsl
 
 logger = logging.getLogger(__name__)
 
@@ -125,6 +125,20 @@ def rocm_platform_plugin() -> str | None:
     except Exception as e:
         logger.debug("ROCm platform is not available because: %s", str(e))
 
+    if not is_rocm and in_wsl():
+        try:
+            import torch
+
+            if (
+                not vllm_version_matches_substr("cpu")
+                and getattr(torch.version, "hip", None)
+                and torch.accelerator.is_available()
+            ):
+                is_rocm = True
+                logger.debug("Confirmed ROCm platform is available in WSL via PyTorch.")
+        except Exception as e:
+            logger.debug("WSL ROCm fallback detection failed because: %s", str(e))
+
     return "vllm.platforms.rocm.RocmPlatform" if is_rocm else None