Fix cuda/cpu check on NoneType (Unit test) (#88970)

test/test_transformers.py

@@ -970,6 +970,15 @@ def make_tensor(*size, device=device, dtype=dtype):
             self.assertRaises(RuntimeError, lambda: torch.nn.functional._scaled_dot_product_attention(
                 q, k, v, torch.ones_like(q), 0.0, False, False))
 
+    # Test failing MHA when bias was NoneType
+    def test_bias_is_none(self):
+        x = torch.rand((1, 5, 10))
+        model = torch.nn.modules.activation.MultiheadAttention(10, 1, bias=False, batch_first=True)
+        model.eval()
+        model(x, x, x)
+        # completes without error
+
+
 # TODO: Replace this with instantiate_device_type_tests() to take advantage of test framework support for
 # cross device / dtype testing.
 instantiate_parametrized_tests(TestTransformers)

torch/nn/modules/activation.py

@@ -1118,7 +1118,7 @@ def forward(self, query: Tensor, key: Tensor, value: Tensor, key_padding_mask: O
                 why_not_fast_path = "some Tensor argument has_torch_function"
             elif not all([(x.is_cuda or 'cpu' in str(x.device)) for x in tensor_args]):
                 why_not_fast_path = "some Tensor argument is neither CUDA nor CPU"
-            elif torch.is_grad_enabled() and any([x.requires_grad for x in tensor_args]):
+            elif torch.is_grad_enabled() and any([x is not None and x.requires_grad for x in tensor_args]):
                 why_not_fast_path = ("grad is enabled and at least one of query or the "
                                      "input/output projection weights or biases requires_grad")
             if not why_not_fast_path: