diff --git a/setup.py b/setup.py
index 5673a4a47a..bc766de4a7 100644
--- a/setup.py
+++ b/setup.py
@@ -62,6 +62,11 @@ def _read_requirements(filename: str) -> list[str]:
     install_requires=get_requirements(),
     ext_modules=ext_modules,
     extras_require={},
+    data_files=[
+        # Install a .pth file so the torch compat shim runs at Python startup,
+        # before ``import vllm`` triggers env_override.py.
+        (".", ["vllm_gaudi_torch_compat.pth"]),
+    ],
     entry_points={
         "vllm.platform_plugins": ["hpu = vllm_gaudi:register"],
         "vllm.general_plugins": [
diff --git a/tests/conftest.py b/tests/conftest.py
new file mode 100644
index 0000000000..17ef7add89
--- /dev/null
+++ b/tests/conftest.py
@@ -0,0 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Root-level conftest – ensures torch compatibility shims are applied
+before any ``import vllm`` happens during the test session.
+"""
+
+import vllm_gaudi._torch_compat  # noqa: F401  -- side-effect: patches GraphCaptureOutput alias
diff --git a/vllm_gaudi/_torch_compat.py b/vllm_gaudi/_torch_compat.py
new file mode 100644
index 0000000000..cc4233f807
--- /dev/null
+++ b/vllm_gaudi/_torch_compat.py
@@ -0,0 +1,42 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Torch compatibility shim for Gaudi's custom PyTorch builds.
+
+Gaudi's PyTorch build (2.9+hpu) cherry-picked the builtins fix from
+upstream PyTorch (pytorch/177558), which renamed ``GraphCaptureOutput``
+to ``CaptureOutput`` and removed the ``get_runtime_env`` method.
+
+vLLM's ``env_override.py`` (guarded by ``not is_torch_equal_or_newer("2.12.0")``)
+tries to import ``GraphCaptureOutput`` and patch its ``get_runtime_env``.
+On Gaudi's build this block must be skipped because the fix is already applied.
+
+We inject a stub ``GraphCaptureOutput`` class with a ``get_runtime_env``
+class-method so that ``env_override.py`` can import and "patch" it without
+error.  The patched method is never actually called because the underlying
+PyTorch code already contains the fix.
+
+This module is loaded:
+* In tests  – via ``tests/conftest.py`` (runs before any ``import vllm``).
+* At runtime – via a ``.pth`` file installed into site-packages so that
+  the shim is in place before *any* Python code imports ``vllm``.
+"""
+
+try:
+    import torch._dynamo.convert_frame as _cf
+
+    if not hasattr(_cf, "GraphCaptureOutput"):
+        # The Gaudi PyTorch build already has the builtins fix applied;
+        # create a stub so that env_override.py can import and monkey-patch
+        # it harmlessly.
+
+        class _GraphCaptureOutputStub:
+            """Stub standing in for the removed GraphCaptureOutput class."""
+
+            def get_runtime_env(self):  # type: ignore[override]
+                """No-op — the real fix is already in this PyTorch build."""
+                return None
+
+        _cf.GraphCaptureOutput = _GraphCaptureOutputStub  # type: ignore[attr-defined]
+except Exception:
+    # If torch._dynamo.convert_frame is unavailable, there is nothing
+    # to patch – silently continue.
+    pass
diff --git a/vllm_gaudi/models/deepseek_ocr.py b/vllm_gaudi/models/deepseek_ocr.py
index 13eecae107..1205167abd 100644
--- a/vllm_gaudi/models/deepseek_ocr.py
+++ b/vllm_gaudi/models/deepseek_ocr.py
@@ -10,7 +10,7 @@
 from vllm.config import VllmConfig
 from vllm.config.multimodal import BaseDummyOptions
 from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.multimodal.inputs import MultiModalDataDict
+from vllm.inputs import MultiModalDataDict
 from vllm.model_executor.models.deepseek_ocr import (
     DeepseekOCRForCausalLM,
     DeepseekOCRMultiModalProcessor,
diff --git a/vllm_gaudi_torch_compat.pth b/vllm_gaudi_torch_compat.pth
new file mode 100644
index 0000000000..b03d1f248a
--- /dev/null
+++ b/vllm_gaudi_torch_compat.pth
@@ -0,0 +1 @@
+import vllm_gaudi._torch_compat