[CherryPick][Auto Parallel] fix loss scale in xpu (#71698) (#71765)

Author: Waynezee

python/paddle/distributed/auto_parallel/api.py

@@ -1933,6 +1933,17 @@ def unscale_method(self, optimizer):
                     temp_param_grads_half,
                     temp_scale,
                 )
+
+                # AllReduce for "bool" is not supported on XPU
+                if "xpu" in paddle.device.get_device():
+                    temp_param_grads_half = paddle.cast(
+                        temp_param_grads_half, "int32"
+                    )
+                    temp_param_grads_half = paddle.sum(temp_param_grads_half)
+                    temp_param_grads_half = paddle.cast(
+                        temp_param_grads_half, "bool"
+                    )
+
                 temp_found_inf = _C_ops.bitwise_or(
                     temp_found_inf, temp_found_inf_half
                 )
@@ -1941,6 +1952,17 @@ def unscale_method(self, optimizer):
                     temp_param_grads_fp32,
                     temp_scale,
                 )
+
+                # AllReduce for "bool" is not supported on XPU
+                if "xpu" in paddle.device.get_device():
+                    temp_found_inf_fp32 = paddle.cast(
+                        temp_found_inf_fp32, "int32"
+                    )
+                    temp_found_inf_fp32 = paddle.sum(temp_found_inf_fp32)
+                    temp_found_inf_fp32 = paddle.cast(
+                        temp_found_inf_fp32, "bool"
+                    )
+
                 temp_found_inf = _C_ops.bitwise_or(
                     temp_found_inf, temp_found_inf_fp32
                 )

python/paddle/optimizer/optimizer.py

@@ -1336,6 +1336,12 @@ def _create_optimization_pass(
 
             if framework.in_dygraph_mode():
                 found_inf = self._get_auxiliary_var('found_inf')
+                if (
+                    "xpu" in paddle.device.get_device()
+                    and found_inf is not None
+                    and found_inf.is_dist()
+                ):
+                    found_inf = found_inf._local_value()
                 if found_inf:
                     if isinstance(found_inf, core.eager.Tensor):
                         self._set_auxiliary_var('found_inf', True)