fix eb4 (#62032)

Author: YuanRisheng

python/paddle/distributed/fleet/layers/mpu/mp_layers.py

@@ -16,6 +16,7 @@
 
 import paddle
 from paddle.autograd import PyLayer
+from paddle.base import core
 from paddle.distributed import fleet
 from paddle.nn import functional as F
 
@@ -33,7 +34,7 @@
 
 
 def is_fused_matmul_bias_supported():
-    return hasattr(paddle._C_ops, 'fused_gemm_epilogue')
+    return hasattr(core.eager.ops.legacy, 'fused_gemm_epilogue')
 
 
 def is_fused_linear_param_grad_add_supported():
@@ -213,10 +214,7 @@ def forward(
         if not fuse_matmul_bias:
             return paddle._C_ops.linear(x, weight, bias)
         else:
-            result, _ = paddle._C_ops.fused_gemm_epilogue(
-                x, weight, bias, False, False, "none"
-            )
-            return result
+            return paddle._legacy_C_ops.fused_gemm_epilogue(x, weight, bias)
 
     @staticmethod
     def backward(ctx, dy):

python/paddle/distributed/fleet/utils/sequence_parallel_utils.py

@@ -17,6 +17,7 @@
 import paddle
 from paddle import distributed as dist
 from paddle.autograd import PyLayer
+from paddle.base import core
 from paddle.distributed import fleet
 from paddle.distributed.fleet.meta_parallel import get_rng_state_tracker
 from paddle.distributed.fleet.utils.hybrid_parallel_util import (
@@ -221,7 +222,7 @@ def is_fused_matmul_bias_supported():
         and not paddle.is_compiled_with_rocm()
         or paddle.is_compiled_with_xpu()
     ):
-        return hasattr(paddle._C_ops, "fused_gemm_epilogue")
+        return hasattr(core.eager.ops.legacy, "fused_gemm_epilogue")
     else:
         return False
 

python/paddle/incubate/nn/functional/fused_matmul_bias.py

@@ -12,9 +12,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from paddle import _C_ops
+from paddle import _C_ops, _legacy_C_ops
 from paddle.base.layer_helper import LayerHelper
-from paddle.framework import in_dynamic_or_pir_mode
+from paddle.framework import (
+    in_dynamic_mode,
+    in_pir_mode,
+)
 from paddle.tensor.linalg import matmul
 
 
@@ -56,7 +59,11 @@ def fused_matmul_bias(
     """
     if bias is None:
         return matmul(x, y, transpose_x, transpose_y, name)
-    if in_dynamic_or_pir_mode():
+    if in_dynamic_mode():
+        return _legacy_C_ops.fused_gemm_epilogue(
+            x, y, bias, 'trans_x', transpose_x, 'trans_y', transpose_y
+        )
+    if in_pir_mode():
         out, _ = _C_ops.fused_gemm_epilogue(
             x, y, bias, transpose_x, transpose_y, "none"
         )
@@ -146,7 +153,20 @@ def fused_linear_activation(
     if activation is None:
         activation = "none"
 
-    if in_dynamic_or_pir_mode():
+    if in_dynamic_mode():
+        return _legacy_C_ops.fused_gemm_epilogue(
+            x,
+            y,
+            bias,
+            'trans_x',
+            trans_x,
+            'trans_y',
+            trans_y,
+            'activation',
+            activation,
+        )
+
+    if in_pir_mode():
         out, _ = _C_ops.fused_gemm_epilogue(
             x,
             y,