Fix some minor issues:

- learning rate for IA³ EmbConv1D needs to be increased for changes in
  the output to become detectable
- transpose now works correctly with nn.Parameter

Note: IA³ + Conv1D tests are still commented out because they fail for
different reasons (shape mismatch).

src/peft/utils/other.py

@@ -331,7 +331,12 @@ def lambda_policy_fn(module):
 
 
 def transpose(weight, fan_in_fan_out):
-    return weight.T if fan_in_fan_out else weight
+    if not fan_in_fan_out:
+        return weight
+
+    if isinstance(weight, torch.nn.Parameter):
+        return torch.nn.Parameter(weight.T)
+    return weight.T
 
 
 def _is_valid_match(key: str, target_key: str):

tests/test_custom_models.py

@@ -493,7 +493,7 @@ def test_disable_adapters(self, test_name, model_id, config_cls, config_kwargs):
 
         model.train()
         # EmbConv1D is slow to learn for some reason
-        lr = 0.01 if model_id != "EmbConv1D" else 0.1
+        lr = 0.01 if model_id != "EmbConv1D" else 100.0
         optimizer = torch.optim.SGD(model.parameters(), lr=lr)
 
         # train at least 3 steps for all parameters to be updated (probably this is required because of symmetry
@@ -534,7 +534,7 @@ def test_disable_adapters_with_merging(self, test_name, model_id, config_cls, co
 
         model.train()
         # EmbConv1D is slow to learn for some reason
-        lr = 0.01 if model_id != "EmbConv1D" else 0.1
+        lr = 0.01 if model_id != "EmbConv1D" else 100.0
         optimizer = torch.optim.SGD(model.parameters(), lr=lr)
 
         # train at least 3 steps for all parameters to be updated (probably this is required because of symmetry