Fix GradientClipByGlobalNorm dtype bug (#27437)

* fix dtype of gradientclipbyglobalnorm

* fix dtype bug of GradientClipbyGlobalnorm

python/paddle/fluid/clip.py

@@ -590,7 +590,7 @@ class GradientClipByGlobalNorm(GradientClipBase):
         global_norm_var = layers.reduce_sum(global_norm_var)
         global_norm_var = layers.sqrt(global_norm_var)
         max_global_norm = layers.fill_constant(
-            shape=[1], dtype='float32', value=self.clip_norm)
+            shape=[1], dtype=global_norm_var.dtype, value=self.clip_norm)
         clip_var = layers.elementwise_div(
             x=max_global_norm,
             y=layers.elementwise_max(
@@ -635,7 +635,9 @@ class GradientClipByGlobalNorm(GradientClipBase):
                 global_norm_var = layers.sums(sum_square_list)
                 global_norm_var = layers.sqrt(x=global_norm_var)
                 max_global_norm = layers.fill_constant(
-                    shape=[1], dtype="float32", value=self.clip_norm)
+                    shape=[1],
+                    dtype=global_norm_var.dtype,
+                    value=self.clip_norm)
                 scale_var = layers.elementwise_div(
                     x=max_global_norm,
                     y=layers.elementwise_max(
@@ -663,7 +665,7 @@ class GradientClipByGlobalNorm(GradientClipBase):
             context[self.group_name] = []
             context[self.group_name + "_clip_value"] = self.clip_norm
             context[self.group_name + "_clip"] = layers.fill_constant(
-                shape=[1], dtype="float32", value=self.clip_norm)
+                shape=[1], dtype=grad.dtype, value=self.clip_norm)
         else:
             if not self.clip_norm == context[self.group_name + "_clip_value"]:
                 raise ValueError(