Add clang-format to pre-commit hooks

.clang-format

@@ -0,0 +1,39 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+# A clang-format style derived from Python's PEP 7 (4-space indent, Stroustrup braces,
+# return type on its own line), except that the column limit is 120 rather than 79.
+# Based on https://gist.github.com/pganssle/0e3a5f828b4d07d79447f6ced8e7e4db
+BasedOnStyle: Google
+BreakAfterReturnType: AllDefinitions
+AllowShortIfStatementsOnASingleLine: Never
+BreakBeforeBraces: Stroustrup
+ColumnLimit: 120
+DerivePointerAlignment: false
+IndentWidth: 4
+Language: Cpp
+PointerAlignment: Right
+ReflowComments: Always
+SpaceBeforeParens: ControlStatements
+SpacesInParens: Never
+TabWidth: 4
+UseTab: Never
+
+# Keep include order as written; reordering can break order-sensitive builds.
+SortIncludes: {Enabled: false}
+
+# Ruff-like wrapping: break after "(", indent arguments, put ")" on its own line.
+# This does not contradict PEP7, but extends it to the consistent style where
+# calls either fit on a single line or they are broken across multiple lines with
+# one argument per line (indented).
+AlignAfterOpenBracket: true
+BreakAfterOpenBracketBracedList: true
+BreakAfterOpenBracketFunction: true
+BreakAfterOpenBracketIf: true
+BreakBeforeCloseBracketBracedList: true
+BreakBeforeCloseBracketFunction: true
+BreakBeforeCloseBracketIf: true
+PackArguments: {BinPack: OnePerLine}
+PackParameters: {BinPack: OnePerLine}
+AllowAllArgumentsOnNextLine: false

.pre-commit-config.yaml

@@ -32,6 +32,13 @@ repos:
       - id: ruff-format
         exclude: .*\.pyi$
 
+  - repo: https://github.com/pre-commit/mirrors-clang-format
+    rev: 98ad4ddb9c057de2f5b9e93e44e6db959807d68c  # frozen: v23.1.3
+    hooks:
+      - id: clang-format
+        types_or: [c, c++, cuda]
+        exclude: ^cuda_core/(cuda/core/_include/(dlpack|aoti_shim)\.h|third_party/)
+
   - repo: local
     hooks:
       - id: check-precommit-installed

CONTRIBUTING.md

@@ -238,6 +238,15 @@ between commits, leaving stale headers or out-of-date stubs in the history.
 If the hook isn't installed, `pre-commit run` will print a visible warning
 reminding you to run `pre-commit install`.
 
+C, C++ and CUDA sources are formatted with `clang-format` using the
+[`.clang-format`](.clang-format) at the repository root, which follows PEP 7
+except for a 120-column limit. The
+pre-commit hook installs its own pinned `clang-format` (from PyPI) into the
+hook environment, so no system-wide installation is required. Vendored code
+(for example `dlpack.h` and `cuda_core/third_party/`) is excluded in
+`.pre-commit-config.yaml`. To format in your editor, install the same version
+with `pip install clang-format==23.1.3`.
+
 ## CodeRabbit
 
 This section explains how to use CodeRabbit for CUDA Python pull request review.

benchmarks/cuda_bindings/benchmarks/cpp/bench_ctx_device.cpp

@@ -9,18 +9,20 @@
 #include <cstdlib>
 #include <iostream>
 
-
-static void check_cu(CUresult status, const char* message) {
+static void
+check_cu(CUresult status, const char *message)
+{
     if (status != CUDA_SUCCESS) {
-        const char* error_name = nullptr;
+        const char *error_name = nullptr;
         cuGetErrorName(status, &error_name);
         std::cerr << message << ": " << (error_name ? error_name : "unknown") << '\n';
         std::exit(1);
     }
 }
 
-
-int main(int argc, char** argv) {
+int
+main(int argc, char **argv)
+{
     bench::Options options = bench::parse_args(argc, argv);
 
     // Setup: init CUDA and create a context
@@ -45,25 +47,19 @@ int main(int argc, char** argv) {
 
     // --- ctx_set_current ---
     {
-        suite.run("ctx_device.ctx_set_current", [&]() {
-            check_cu(cuCtxSetCurrent(ctx), "cuCtxSetCurrent failed");
-        });
+        suite.run("ctx_device.ctx_set_current", [&]() { check_cu(cuCtxSetCurrent(ctx), "cuCtxSetCurrent failed"); });
     }
 
     // --- ctx_get_device ---
     {
         CUdevice dev;
-        suite.run("ctx_device.ctx_get_device", [&]() {
-            check_cu(cuCtxGetDevice(&dev), "cuCtxGetDevice failed");
-        });
+        suite.run("ctx_device.ctx_get_device", [&]() { check_cu(cuCtxGetDevice(&dev), "cuCtxGetDevice failed"); });
     }
 
     // --- device_get ---
     {
         CUdevice dev;
-        suite.run("ctx_device.device_get", [&]() {
-            check_cu(cuDeviceGet(&dev, 0), "cuDeviceGet failed");
-        });
+        suite.run("ctx_device.device_get", [&]() { check_cu(cuDeviceGet(&dev, 0), "cuDeviceGet failed"); });
     }
 
     // --- device_get_attribute ---
@@ -80,21 +76,15 @@ int main(int argc, char** argv) {
     // --- device_primary_ctx_retain ---
     // Outer retain so the benchmarked retain/release pair just bumps the refcount.
     CUcontext primary_outer = nullptr;
-    check_cu(
-        cuDevicePrimaryCtxRetain(&primary_outer, device),
-        "cuDevicePrimaryCtxRetain (setup) failed"
-    );
+    check_cu(cuDevicePrimaryCtxRetain(&primary_outer, device), "cuDevicePrimaryCtxRetain (setup) failed");
     {
         CUcontext primary = nullptr;
         suite.run("ctx_device.device_primary_ctx_retain", [&]() {
             check_cu(cuDevicePrimaryCtxRetain(&primary, device), "cuDevicePrimaryCtxRetain failed");
             check_cu(cuDevicePrimaryCtxRelease(device), "cuDevicePrimaryCtxRelease failed");
         });
     }
-    check_cu(
-        cuDevicePrimaryCtxRelease(device),
-        "cuDevicePrimaryCtxRelease (teardown) failed"
-    );
+    check_cu(cuDevicePrimaryCtxRelease(device), "cuDevicePrimaryCtxRelease (teardown) failed");
 
     // Cleanup
     check_cu(cuCtxDestroy(ctx), "cuCtxDestroy failed");

benchmarks/cuda_bindings/benchmarks/cpp/bench_event.cpp

@@ -9,18 +9,20 @@
 #include <cstdlib>
 #include <iostream>
 
-
-static void check_cu(CUresult status, const char* message) {
+static void
+check_cu(CUresult status, const char *message)
+{
     if (status != CUDA_SUCCESS) {
-        const char* error_name = nullptr;
+        const char *error_name = nullptr;
         cuGetErrorName(status, &error_name);
         std::cerr << message << ": " << (error_name ? error_name : "unknown") << '\n';
         std::exit(1);
     }
 }
 
-
-int main(int argc, char** argv) {
+int
+main(int argc, char **argv)
+{
     bench::Options options = bench::parse_args(argc, argv);
 
     // Setup
@@ -47,9 +49,7 @@ int main(int argc, char** argv) {
     bench::BenchmarkSuite suite(options);
     // Drain the persistent stream after calibration so event_record (which
     // enqueues onto the stream) and event_synchronize start from a known state.
-    suite.set_post_calibrate([&]() {
-        check_cu(cuStreamSynchronize(stream), "post-calibrate sync failed");
-    });
+    suite.set_post_calibrate([&]() { check_cu(cuStreamSynchronize(stream), "post-calibrate sync failed"); });
 
     // --- event_create_destroy ---
     {
@@ -62,9 +62,7 @@ int main(int argc, char** argv) {
 
     // --- event_record ---
     {
-        suite.run("event.event_record", [&]() {
-            check_cu(cuEventRecord(event, stream), "cuEventRecord failed");
-        });
+        suite.run("event.event_record", [&]() { check_cu(cuEventRecord(event, stream), "cuEventRecord failed"); });
     }
 
     // Re-sync so event is in a known completed state after the record benchmark