fix bug of recompute in hybridparallel (#35588)

paddle/fluid/operators/flatten_op.cu.cc

@@ -15,6 +15,7 @@ limitations under the License. */
 #include "paddle/fluid/operators/flatten_op.h"
 
 namespace ops = paddle::operators;
+namespace plat = paddle::platform;
 
 REGISTER_OP_CUDA_KERNEL(
     flatten, ops::FlattenKernel<paddle::platform::CUDADeviceContext, float>,
@@ -50,6 +51,8 @@ REGISTER_OP_CUDA_KERNEL(
     flatten_contiguous_range,
     ops::FlattenContiguousRangeKernel<paddle::platform::CUDADeviceContext,
                                       float>,
+    ops::FlattenContiguousRangeKernel<paddle::platform::CUDADeviceContext,
+                                      plat::float16>,
     ops::FlattenContiguousRangeKernel<paddle::platform::CUDADeviceContext,
                                       double>,
     ops::FlattenContiguousRangeKernel<paddle::platform::CUDADeviceContext,
@@ -63,6 +66,8 @@ REGISTER_OP_CUDA_KERNEL(
     flatten_contiguous_range_grad,
     ops::FlattenContiguousRangeGradKernel<paddle::platform::CUDADeviceContext,
                                           float>,
+    ops::FlattenContiguousRangeGradKernel<paddle::platform::CUDADeviceContext,
+                                          plat::float16>,
     ops::FlattenContiguousRangeGradKernel<paddle::platform::CUDADeviceContext,
                                           double>,
     ops::FlattenContiguousRangeGradKernel<paddle::platform::CUDADeviceContext,

python/paddle/distributed/fleet/meta_parallel/pp_utils/utils.py

@@ -133,6 +133,7 @@ def _split_activation(tensor):
 
     # use inplace operation to save memory
     data = tensor.flatten_()
+
     part_size = tensor_numel // mp_degree
     start = part_size * mp_rank
     end = start + part_size

python/paddle/fluid/contrib/mixed_precision/fp16_lists.py

@@ -94,6 +94,7 @@ black_list = {
     'softmax',
     'softmax_with_cross_entropy',
     'sigmoid_cross_entropy_with_logits',
+    'c_softmax_with_cross_entropy',
     'cross_entropy',
     'cross_entropy2',
     # fp16 is slower than fp32, though fp16 is supported.

python/paddle/fluid/dygraph/amp/auto_cast.py

@@ -45,6 +45,7 @@ BLACK_LIST = {
     'softmax',
     'softmax_with_cross_entropy',
     'sigmoid_cross_entropy_with_logits',
+    'c_softmax_with_cross_entropy',
     'cross_entropy',
     'cross_entropy2',
     # default fp32 can avoid return inf when the sum value large than 65504