[PIR]Fix value numel method (#67754)

Author: 0x45f

paddle/fluid/pybind/pir.cc

@@ -1322,7 +1322,6 @@ void BindValue(py::module *m) {
           "get_defining_op",
           [](Value self) -> pir::Operation * { return self.defining_op(); },
           return_value_policy::reference)
-      .def("numel", [](Value self) { return phi::product(GetValueDims(self)); })
       .def("type", &Value::type)
       .def("index",
            [](Value self) -> uint32_t {

paddle/phi/ops/yaml/ops.yaml

@@ -3476,6 +3476,7 @@
   data_transform:
     skip_transform : x
   no_need_buffer : x
+  traits : paddle::dialect::ForwardOnlyTrait
   interfaces : paddle::dialect::InferSymbolicShapeInterface
 
 - op : one_hot

python/paddle/decomposition/recompute.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 from __future__ import annotations
 
+import functools
 import math
 import os
 from typing import TYPE_CHECKING
@@ -797,7 +798,10 @@ def cal_value_node_size(value_node):
     # todo(wanghao107) hack for dynamic shape
     if is_dynamic_value_node(value_node):
         return 1
-    return value_node.numel() * _PADDLE_DTYPE_2_NBYTES[value_node.dtype]
+    return (
+        functools.reduce(lambda x, y: x * y, value_node.shape, 1)
+        * _PADDLE_DTYPE_2_NBYTES[value_node.dtype]
+    )
 
 
 def cal_value_nodes_dist_to_backward(all_ops, required_fw_value_nodes):

python/paddle/static/nn/common.py

@@ -3492,7 +3492,7 @@ def spectral_norm(weight, dim=0, power_iters=1, eps=1e-12, name=None):
 
     # create input and parameters
     input_shape = weight.shape
-    assert weight.numel() > 0, "Any dimension of input cannot be equal to 0."
+    assert 0 not in input_shape, "Any dimension of input cannot be equal to 0."
 
     if dim not in [0, 1]:
         raise ValueError(

python/paddle/tensor/manipulation.py

@@ -14,6 +14,7 @@
 
 from __future__ import annotations
 
+import functools
 import math
 from typing import TYPE_CHECKING, Any, Literal, overload
 
@@ -4459,8 +4460,9 @@ def check_input(x, repeat_times):
         else:
             for elem in repeat_times:
                 if isinstance(elem, (Variable, paddle.pir.Value)):
+                    numel = functools.reduce(lambda x, y: x * y, elem.shape, 1)
                     assert (
-                        elem.numel() == 1
+                        numel == 1
                     ), 'Elements in repeat_times must be Tensor with one element or integers.'
                 else:
                     type_tuple = (int, np.int32, np.int64)