Fix bugs of pipeline on ascend. (#32737)

paddle/fluid/framework/device_worker.h

@@ -639,7 +639,7 @@ class PSGPUWorker : public HogwildWorker {
 #endif
 
 #if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL) || \
-    defined(WITH_ASCEND_CL)
+    defined(PADDLE_WITH_ASCEND_CL)
 class SectionWorker : public DeviceWorker {
  public:
   SectionWorker() {}

paddle/fluid/framework/device_worker_factory.cc

@@ -80,7 +80,7 @@ REGISTER_DEVICE_WORKER_CLASS(PSGPUWorker);
 #endif
 
 #if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL) || \
-    defined(WITH_ASCEND_CL)
+    defined(PADDLE_WITH_ASCEND_CL)
 REGISTER_DEVICE_WORKER_CLASS(SectionWorker);
 #endif
 }  // namespace framework

paddle/fluid/framework/pipeline_trainer.cc

@@ -13,7 +13,7 @@
 // limitations under the License.
 
 #if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL) || \
-    defined(WITH_ASCEND_CL)
+    defined(PADDLE_WITH_ASCEND_CL)
 #include "paddle/fluid/framework/data_feed_factory.h"
 #include "paddle/fluid/framework/device_worker_factory.h"
 #include "paddle/fluid/framework/trainer.h"
@@ -37,7 +37,7 @@ void PipelineTrainer::Initialize(const TrainerDesc& trainer_desc,
   int place_id = section_config.place_id();
 #if (defined PADDLE_WITH_NCCL)
   place_ = platform::CUDAPlace(place_id);
-#elif (defined WITH_ASCEND_CL)  // NOLINT
+#elif (defined PADDLE_WITH_ASCEND_CL)  // NOLINT
   place_ = platform::NPUPlace(place_id);
 #endif
   worker_ = DeviceWorkerFactory::CreateDeviceWorker(

paddle/fluid/framework/section_worker.cc

@@ -10,7 +10,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL) || \
-    defined(WITH_ASCEND_CL)
+    defined(PADDLE_WITH_ASCEND_CL)
 #include <float.h>
 #include "paddle/fluid/framework/device_worker.h"
 #include "paddle/fluid/framework/executor_gc_helper.h"

paddle/fluid/framework/trainer.h

@@ -332,7 +332,7 @@ class PSGPUTrainer : public TrainerBase {
 #endif
 
 #if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL) || \
-    defined(WITH_ASCEND_CL)
+    defined(PADDLE_WITH_ASCEND_CL)
 class PipelineTrainer : public TrainerBase {
  public:
   PipelineTrainer() {}

paddle/fluid/framework/trainer_factory.cc

@@ -76,7 +76,8 @@ REGISTER_TRAINER_CLASS(HeterBoxTrainer);
     (defined PADDLE_WITH_PSLIB)
 REGISTER_TRAINER_CLASS(PSGPUTrainer);
 #endif
-#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
+#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL) || \
+    defined(PADDLE_WITH_ASCEND_CL)
 REGISTER_TRAINER_CLASS(PipelineTrainer);
 #endif
 }  // namespace framework

paddle/fluid/operators/collective/c_allreduce_op.h

@@ -131,6 +131,7 @@ class CAllReduceOpASCENDKernel : public framework::OpKernel<T> {
     int64_t numel = in->numel();
 
     void* sendbuff = reinterpret_cast<void*>(const_cast<T*>(in->data<T>()));
+    out->mutable_data<T>(in->dims(), ctx.GetPlace());
     void* recvbuff = reinterpret_cast<void*>(out->data<T>());
 
     int ring_id = ctx.Attr<int>("ring_id");

python/paddle/fluid/framework.py

@@ -6124,9 +6124,9 @@ def device_guard(device=None):
         device, index = device.split(':')
         if device == 'cpu':
             raise ValueError("Should not set device id for cpu.")
-    if device not in ['cpu', 'gpu', '', None]:
+    if device not in ['cpu', 'gpu', 'npu', '', None]:
         raise ValueError(
-            "The Attr(device) should be 'cpu' or 'gpu', and it can also be empty string or None "
+            "The Attr(device) should be 'cpu' 'npu' or 'gpu', and it can also be empty string or None "
             "when there is no need to specify device. But received %s" % device)
     if index:
         device = ":".join([device, index])

python/paddle/fluid/optimizer.py

@@ -4116,7 +4116,7 @@ class PipelineOptimizer(object):
         device = op.attr(self._op_device_key) \
             if op.has_attr(self._op_device_key) else None
         if device:
-            assert device[0:3] == 'gpu', "Now, only gpu devices are " \
+            assert device[0:3] == 'gpu' or dev_type == 'npu', "Now, only gpu and npu devices are " \
                 "supported in pipeline parallemism."
         return device