pre-commit fix.

.vscode/settings.json

@@ -16,4 +16,4 @@
     "gotoSymbolStack.currentStackPosition": 0,
     "gotoSymbolStack.maxStackPosition": 0,
     "gotoSymbolStack.filePositionInfo": []
-}
+}

benchmarks/cpp/flashattention/Makefile

@@ -13,4 +13,4 @@ build:
 	@cd $(BUILD_DIR) && cmake .. && make -j$(proc)
 
 clean:
-	@rm -rf $(BUILD_DIR)
+	@rm -rf $(BUILD_DIR)

benchmarks/cpp/flashattention/convert.cuh

@@ -65,4 +65,4 @@ DEVICE auto convert_layout_scores_copyview(LayoutType layout_s) {
 }
 
 }  // namespace cutlass_wrapper
-}  // namespace benchmarks
+}  // namespace benchmarks

benchmarks/cpp/flashattention/main.cu

@@ -138,4 +138,4 @@ void run(bool check = true) {
     cudaDeviceSynchronize();
 }
 
-int main() { run(); }
+int main() { run(); }

benchmarks/cpp/flashattention/reduce.cuh

@@ -137,4 +137,4 @@ DEVICE void reduce_sumabs(cute::Tensor<Engine0, Layout0> const& tensor,
 }
 
 }  // namespace cutlass_wrapper
-}  // namespace benchmarks
+}  // namespace benchmarks

benchmarks/cpp/flashattention/util.hpp

@@ -29,4 +29,4 @@ bool check_results(const __half* values1, const __half* values2, int numel) {
         }
     }
     return passed;
-}
+}