Apply suggestions from code review

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

Author: sayakpaul

src/diffusers/loaders/lora_pipeline.py

@@ -2357,7 +2357,10 @@ def _maybe_expand_transformer_param_shape_or_error_(
                         expanded_module = torch.nn.Linear(
                             in_features, out_features, bias=bias, dtype=module_weight.dtype
                         )
-                    # Only weights are expanded and biases are not.
+                    # Only weights are expanded and biases are not. This is because only the input dimensions
+                    # are changed while the output dimensions remain the same. The shape of the weight tensor
+                    # is (out_features, in_features), while the shape of bias tensor is (out_features,), which
+                    # explains the reason why only weights are expanded.
                     new_weight = torch.zeros_like(
                         expanded_module.weight.data, device=module_weight.device, dtype=module_weight.dtype
                     )

tests/lora/test_lora_layers_flux.py

@@ -583,7 +583,7 @@ def test_fuse_expanded_lora_with_regular_lora(self):
         self.assertTrue(np.allclose(lora_output_3, lora_output_4, atol=1e-3, rtol=1e-3))
 
     def test_load_regular_lora(self):
-        # This test checks if a regular lora (think of one trained Flux.1 Dev for example) can be loaded
+        # This test checks if a regular lora (think of one trained on Flux.1 Dev for example) can be loaded
         # into the transformer with more input channels than Flux.1 Dev, for example. Some examples of those
         # transformers include Flux Fill, Flux Control, etc.
         components, _, _ = self.get_dummy_components(FlowMatchEulerDiscreteScheduler)