xpu add bf16 to two kernels (#65653)

FeixLiu · web-flow · commit 9ee4943da8bd · 2024-07-03T17:03:27.000+08:00
diff --git a/paddle/fluid/pybind/process_group_utils.h b/paddle/fluid/pybind/process_group_utils.h
@@ -170,6 +170,10 @@ void ConcatDenseTensorWithType(const phi::XPUContext &dev_ctx,
       ConcatDenseTensor<phi::XPUContext, phi::dtype::float16>()(
           dev_ctx, t_list, p_out);
       break;
+    case phi::DataType::BFLOAT16:
+      ConcatDenseTensor<phi::XPUContext, phi::dtype::bfloat16>()(
+          dev_ctx, t_list, p_out);
+      break;
     case phi::DataType::FLOAT32:
       ConcatDenseTensor<phi::XPUContext, float>()(dev_ctx, t_list, p_out);
       break;
@@ -241,6 +245,10 @@ void SplitDenseTensorWithType(const phi::XPUContext &dev_ctx,
       SplitDenseTensor<phi::XPUContext, phi::dtype::float16>()(
           dev_ctx, t_in, p_list);
       break;
+    case phi::DataType::BFLOAT16:
+      SplitDenseTensor<phi::XPUContext, phi::dtype::bfloat16>()(
+          dev_ctx, t_in, p_list);
+      break;
     case phi::DataType::FLOAT32:
       SplitDenseTensor<phi::XPUContext, float>()(dev_ctx, t_in, p_list);
       break;
