[Relax][PyTorch] improve the check for no bias situation (#18374)

* [#18373] improve the check for no bias situation

* [#18373] refactor the _normalize_python_tuple

Author: cchung100m

python/tvm/relax/block_builder.py

@@ -299,6 +299,10 @@ def _normalize_python_tuple(self, expr: Union[Expr, Sequence[Expr]]):
         """
         if isinstance(expr, (list, tuple)):
             return Tuple([self._normalize_python_tuple(element) for element in expr])
+        elif expr is None:
+            from . import op
+
+            return op.null_value()
         else:
             return expr
 

python/tvm/relax/frontend/torch/base_fx_graph_translator.py

@@ -88,6 +88,22 @@ def shape_of(tensor):
             return tensor.shape
         raise ValueError("Unsupported type: {}".format(type(tensor)))
 
+    @staticmethod
+    def _is_no_bias(bias):
+        """Check if bias represents 'no bias' condition.
+
+        This handles both Python None and relax.op.null_value() expressions
+        that might be used to represent missing bias parameters.
+        """
+        if bias is None:
+            return True
+
+        # Check if this is a null_value expression
+        if isinstance(bias, relax.Call) and bias.op.name == "relax.null_value":
+            return True
+
+        return False
+
     def retrieve_args(self, node: fx.Node):
         return self._retrieve_args(node.args)
 
@@ -103,7 +119,7 @@ def _retrieve_args(self, node):
         elif isinstance(node, dict):
             return {self._retrieve_args(k): self._retrieve_args(v) for k, v in node.items()}
         elif node is None:
-            return relax.op.null_value()
+            return None
         else:
             return node
 
@@ -758,7 +774,7 @@ def _conv_transpose1d_impl(
             )
         )
 
-        if bias is None:
+        if self._is_no_bias(bias):
             return conv1d_transpose
 
         assert len(self.shape_of(bias)) == 1
@@ -812,7 +828,7 @@ def _conv_transpose2d_impl(
             )
         )
 
-        if bias is None:
+        if self._is_no_bias(bias):
             return conv2d_transpose
 
         assert len(self.shape_of(bias)) == 1
@@ -864,7 +880,7 @@ def _conv1d_impl(
             )
         )
 
-        if bias is None:
+        if self._is_no_bias(bias):
             return conv1d
         assert len(self.shape_of(bias)) == 1
         bias = relax.op.reshape(bias, (1, -1, 1))
@@ -913,7 +929,7 @@ def _conv2d_impl(
             )
         )
 
-        if bias is None:
+        if self._is_no_bias(bias):
             return conv2d
         assert len(self.shape_of(bias)) == 1
         bias = relax.op.reshape(bias, (1, -1, 1, 1))
@@ -962,7 +978,7 @@ def _conv3d_impl(
             )
         )
 
-        if bias is None:
+        if self._is_no_bias(bias):
             return conv3d
         assert len(self.shape_of(bias)) == 1
         bias = relax.op.reshape(bias, (1, -1, 1, 1, 1))