Auto-enable Vulkan in builds, and update docs for Vulkan backend

docs/source/backends-overview.md

@@ -24,7 +24,7 @@ Backends are the bridge between your exported model and the hardware it runs on.
 | [CUDA](backends/cuda/cuda-overview.md)                       | Linux/Windows | GPU           | NVIDIA GPU acceleration         |
 | [Core ML](backends/coreml/coreml-overview.md)                | iOS, macOS    | NPU/GPU/CPU   | Apple devices, high performance |
 | [Metal Performance Shaders](backends/mps/mps-overview.md)    | iOS, macOS    | GPU           | Apple GPU acceleration          |
-| [Vulkan](backends/vulkan/vulkan-overview.md)                 | Android       | GPU           | Android GPU acceleration        |
+| [Vulkan](backends/vulkan/vulkan-overview.md)                 | Android, Linux, Windows | GPU  | Android devices (mature); Desktops (experimental) |
 | [Qualcomm](backends-qualcomm)                                | Android     | NPU           | Qualcomm SoCs                   |
 | [MediaTek](backends-mediatek)                                | Android     | NPU           | MediaTek SoCs                   |
 | [Arm Ethos-U](backends/arm-ethos-u/arm-ethos-u-overview.md)  | Embedded    | NPU           | Arm MCUs                        |

docs/source/backends/vulkan/vulkan-overview.md

@@ -3,7 +3,8 @@
 The ExecuTorch Vulkan (ET-VK) backend enables ExecuTorch models to execute on
 GPUs via the cross-platform [Vulkan API](https://www.vulkan.org/). Although the
 Vulkan API support is almost ubiquitous among modern GPUs, the ExecuTorch Vulkan
-backend is currently developed with a specific focus for **Android GPUs**.
+backend is developed with a focus on **Android GPUs**, and support for desktop
+platforms is experimental.
 
 ## Features
 

install_utils.py

@@ -6,6 +6,7 @@
 # LICENSE file in the root directory of this source tree.
 
 import functools
+import os
 import platform
 import re
 import subprocess
@@ -56,6 +57,44 @@ def is_cuda_available() -> bool:
         return False
 
 
+def is_vulkan_available() -> bool:
+    """
+    Check if the Vulkan shader compiler (glslc) is available on the system.
+
+    glslc is the only build-time dependency for the Vulkan backend; the Vulkan
+    loader itself is dlopen()ed at runtime via volk. Restricted to Linux and
+    Windows, the desktop GPU platforms the backend supports (macOS would require
+    MoltenVK).
+
+    glslc is looked up on PATH and, failing that, under $VULKAN_SDK/{bin,Bin} to
+    match the find_program() HINTS the build uses (see pybind.cmake and
+    ShaderLibrary.cmake): the Windows Vulkan SDK sets VULKAN_SDK but does not add
+    its bin directory to PATH, so a PATH-only probe would miss it there.
+
+    Returns:
+        True if glslc is available on a supported platform, False otherwise.
+    """
+    if sys.platform not in ("linux", "win32"):
+        return False
+    candidates = ["glslc"]
+    vulkan_sdk = os.environ.get("VULKAN_SDK")
+    if vulkan_sdk:
+        glslc_name = "glslc.exe" if sys.platform == "win32" else "glslc"
+        candidates += [
+            os.path.join(vulkan_sdk, "bin", glslc_name),
+            os.path.join(vulkan_sdk, "Bin", glslc_name),
+        ]
+    for glslc in candidates:
+        try:
+            subprocess.run(
+                [glslc, "--version"], capture_output=True, text=True, check=True
+            )
+            return True
+        except Exception:
+            continue
+    return False
+
+
 @functools.lru_cache(maxsize=1)
 def _get_cuda_version():
     """

setup.py

@@ -889,6 +889,27 @@ def run(self):  # noqa C901
         ):
             cmake_configuration_args += ["-DEXECUTORCH_BUILD_CUDA=ON"]
 
+        # Auto-enable the Vulkan backend when its only build-time dependency,
+        # the glslc shader compiler, is available (mirrors the CUDA detection
+        # above). Unlike CUDA, the backend also needs third-party submodules
+        # that are not in the default checkout set, so gate on their presence
+        # too: an incidental glslc must not turn a working source build into a
+        # hard CMake failure. An explicit -DEXECUTORCH_BUILD_VULKAN=ON still
+        # flows through and surfaces the backend's own missing-submodule error.
+        vulkan_third_party = os.path.join(
+            os.path.dirname(os.path.abspath(__file__)),
+            "backends/vulkan/third-party/volk/volk.c",
+        )
+        if (
+            not minimal_build
+            and os.path.exists(vulkan_third_party)
+            and install_utils.is_vulkan_available()
+            and install_utils.is_cmake_option_on(
+                cmake_configuration_args, "EXECUTORCH_BUILD_VULKAN", default=True
+            )
+        ):
+            cmake_configuration_args += ["-DEXECUTORCH_BUILD_VULKAN=ON"]
+
         # Check if QNN SDK is available (via QNN_SDK_ROOT env var), and if so,
         # enable building the Qualcomm backend by default.
         qnn_sdk_root = os.environ.get("QNN_SDK_ROOT", "").strip()