[PHI] Fix paddle.trunc for big tensor (#73198)

LCStayingdullCircuit · web-flow · commit db144c5eb3b0 · 2025-06-11T17:11:17.000+08:00
diff --git a/paddle/phi/kernels/gpu/trunc_grad_kernel.cu b/paddle/phi/kernels/gpu/trunc_grad_kernel.cu
@@ -25,7 +25,7 @@ using phi::PADDLE_CUDA_NUM_THREADS;
 
 template <typename T>
 __global__ void TruncGrad(T* dx, int64_t N) {
-  CUDA_KERNEL_LOOP(index, N) { dx[index] = static_cast<T>(0.0); }
+  CUDA_KERNEL_LOOP_TYPE(index, N, int64_t) { dx[index] = static_cast<T>(0.0); }
 }
 
 template <typename T, typename Context>

Original file line number	Diff line number	Diff line change
`@@ -25,7 +25,7 @@ using phi::PADDLE_CUDA_NUM_THREADS;`
`25`	`25`
`26`	`26`	`template <typename T>`
`27`	`27`	`__global__ void TruncGrad(T* dx, int64_t N) {`
`28`		`- CUDA_KERNEL_LOOP(index, N) { dx[index] = static_cast<T>(0.0); }`
	`28`	`+ CUDA_KERNEL_LOOP_TYPE(index, N, int64_t) { dx[index] = static_cast<T>(0.0); }`
`29`	`29`	`}`
`30`	`30`
`31`	`31`	`template <typename T, typename Context>`