diff --git a/amdsmi-shim/strixhalo_vram_fix.py b/amdsmi-shim/strixhalo_vram_fix.py
index 4002289..e296d7c 100644
--- a/amdsmi-shim/strixhalo_vram_fix.py
+++ b/amdsmi-shim/strixhalo_vram_fix.py
@@ -55,8 +55,33 @@ def _get_real_vram() -> tuple[int, int] | None:
     return None
 
 
+def _should_skip() -> bool:
+    """Check if we should skip the patch (lightweight/init containers)."""
+    # Check cgroup memory limit — if under 512Mi, skip the expensive
+    # torch/ROCm import. KubeRay's wait-gcs-ready init container has
+    # only 256Mi and importing torch+ROCm would OOMKill it.
+    for cgroup_mem_path in (
+        "/sys/fs/cgroup/memory.max",           # cgroup v2
+        "/sys/fs/cgroup/memory/memory.limit_in_bytes",  # cgroup v1
+    ):
+        try:
+            with open(cgroup_mem_path) as f:
+                val = f.read().strip()
+                if val != "max" and int(val) < 512 * 1024 * 1024:
+                    return True
+        except (OSError, ValueError):
+            continue
+    return False
+
+
 def _apply_patch() -> None:
     """Patch torch.cuda.mem_get_info if we detect unified memory mis-reporting."""
+    if _should_skip():
+        return
+
+    if _get_real_vram() is None:
+        return
+
     try:
         import torch
         if not hasattr(torch, "cuda") or not torch.cuda.is_available():