diff --git a/tests/compile/fullgraph/test_full_graph.py b/tests/compile/fullgraph/test_full_graph.py
index ffec14d36fd17..c5baa66cbeb07 100644
--- a/tests/compile/fullgraph/test_full_graph.py
+++ b/tests/compile/fullgraph/test_full_graph.py
@@ -13,9 +13,7 @@ from vllm import LLM, SamplingParams
 from vllm.attention.backends.registry import AttentionBackendEnum
 from vllm.config import CompilationConfig, CompilationMode, CUDAGraphMode, PassConfig
 from vllm.platforms import current_platform
-from vllm.utils.torch_utils import (
-    is_torch_equal_or_newer,
-)
+from vllm.utils.torch_utils import is_torch_equal_or_newer
 
 from ...utils import create_new_process_for_each_test
 
diff --git a/vllm/env_override.py b/vllm/env_override.py
index e091867fd0ca8..474ac69919eb1 100644
--- a/vllm/env_override.py
+++ b/vllm/env_override.py
@@ -377,9 +377,12 @@ def _patch_get_raw_stream_if_needed():
     if is_torch_equal("2.9.0") or is_torch_equal("2.9.1"):
         import builtins
 
-        from torch._C import _cuda_getCurrentRawStream as _get_raw_stream
+        # Check if CUDA functionality is available without initializing CUDA
+        # _cuda_getCurrentRawStream only exists in CUDA builds of PyTorch
+        if hasattr(torch._C, "_cuda_getCurrentRawStream"):
+            from torch._C import _cuda_getCurrentRawStream as _get_raw_stream
 
-        builtins.get_raw_stream = _get_raw_stream
+            builtins.get_raw_stream = _get_raw_stream
 
 
 _patch_get_raw_stream_if_needed()