Fix a bug of quantization (#36982) (#37381)

* fix a quantization bug

Co-authored-by: XGZhang <[email protected]>

Author: ceci3

python/paddle/fluid/contrib/slim/quantization/quantization_pass.py

@@ -1292,10 +1292,11 @@ def _insert_post_channel_dequant_op(self, graph, op_node, quant_axis):
             var_type=output_var_node.type(),
             shape=output_var_node.shape(),
             var_dtype=output_var_node.dtype())
+        x_num_col_dims = 1
+        if op_node.name() in ['matmul', 'matmul_v2', 'mul']:
+            x_num_col_dims = len(op_node.outputs[0].shape()) - 1
         if op_node.op().has_attr("x_num_col_dims"):
             x_num_col_dims = op_node.op().attr("x_num_col_dims")
-        else:
-            x_num_col_dims = 1
         dequant_op_node = graph.create_op_node(
             op_type='fake_channel_wise_dequantize_max_abs',
             attrs={