fix cuda generator hard-coded offset step (#27027)

paddle/fluid/operators/gaussian_random_op.cu

@@ -71,11 +71,7 @@ class GPUGaussianRandomKernel : public framework::OpKernel<T> {
 
     if (gen_cuda->GetIsInitPy() && seed_flag) {
       auto seed_offset = gen_cuda->IncrementOffset(1);
-      int offset_step = 100;
+      int gen_offset = size * seed_offset.second;
-      // NOTE(xuefeng): Currently, we let offset step fixed to avoid
-      // unexpected results which may cause ut fail.
-      // we will fix this in future.
-      int gen_offset = offset_step * seed_offset.second;
       thrust::transform(
           index_sequence_begin, index_sequence_begin + size,
           thrust::device_ptr<T>(data),
@@ -112,11 +108,7 @@ class GPUGaussianRandomBatchSizeLikeKernel : public framework::OpKernel<T> {
 
     if (gen_cuda->GetIsInitPy() && seed_flag) {
       auto seed_offset = gen_cuda->IncrementOffset(1);
-      int offset_step = 100;
+      int gen_offset = size * seed_offset.second;
-      // NOTE(xuefeng): Currently, we let offset step fixed to avoid
-      // unexpected results which may cause ut fail.
-      // we will fix this in future.
-      int gen_offset = offset_step * seed_offset.second;
       thrust::transform(index_sequence_begin, index_sequence_begin + size,
                         thrust::device_ptr<T>(data),
                         GaussianGenerator<T>(mean, std, seed_offset.first,

paddle/fluid/operators/truncated_gaussian_random_op.cu

@@ -103,11 +103,7 @@ class GPUTruncatedGaussianRandomKernel : public framework::OpKernel<T> {
 
     if (gen_cuda->GetIsInitPy() && seed_flag) {
       auto seed_offset = gen_cuda->IncrementOffset(1);
-      int offset_step = 100;
+      int gen_offset = size * seed_offset.second;
-      // NOTE(xuefeng): Currently, we let offset step fixed to avoid
-      // unexpected results which may cause ut fail.
-      // we will fix this in future.
-      int gen_offset = offset_step * seed_offset.second;
       thrust::transform(
           index_sequence_begin, index_sequence_begin + size,
           thrust::device_ptr<T>(data),

paddle/fluid/operators/uniform_random_op.cu

@@ -143,11 +143,7 @@ class GPUUniformRandomKernel : public framework::OpKernel<T> {
     auto gen_cuda = framework::GetDefaultCUDAGenerator(device_id);
     if (gen_cuda->GetIsInitPy() && seed_flag) {
       auto seed_offset = gen_cuda->IncrementOffset(1);
-      int offset_step = 100;
+      int gen_offset = size * seed_offset.second;
-      // NOTE(xuefeng): Currently, we let offset step fixed to avoid
-      // unexpected results which may cause ut fail.
-      // we will fix this in future.
-      int gen_offset = offset_step * seed_offset.second;
       thrust::transform(
           index_sequence_begin, index_sequence_begin + size,
           thrust::device_ptr<T>(data),