FIX for ConvNd layers using the groups argument.

src/peft/tuners/lora/layer.py

@@ -1061,7 +1061,9 @@ def update_layer(
         conv_layer = type(base_layer)
         out_kernel = out_stride = (1,) * (self._kernel_dim - 2)
         self.lora_A[adapter_name] = conv_layer(self.in_features, r, kernel_size, stride, padding, bias=False)
-        self.lora_B[adapter_name] = conv_layer(r, self.out_features, out_kernel, out_stride, bias=lora_bias)
+        self.lora_B[adapter_name] = conv_layer(
+            r, self.out_features // base_layer.groups, out_kernel, out_stride, bias=lora_bias
+        )
         self.lora_bias[adapter_name] = lora_bias
 
         if use_rslora:
@@ -1243,13 +1245,12 @@ def get_delta_weight(self, adapter) -> torch.Tensor:
                 3
             ) * self.scaling[adapter]
         else:
-            output_tensor = (
-                self.conv_fn(
-                    weight_A.transpose(0, 1),
-                    weight_B,
-                ).transpose(0, 1)
-                * self.scaling[adapter]
-            )
+            output_tensor = self.conv_fn(weight_A.transpose(0, 1), weight_B)
+
+            if self.get_base_layer().groups > 1:
+                output_tensor = output_tensor * self.scaling[adapter]
+            else:
+                output_tensor = output_tensor.transpose(0, 1) * self.scaling[adapter]
 
         if cast_to_fp32:
             output_tensor = output_tensor.to(dtype=dtype)

tests/test_custom_models.py

@@ -115,6 +115,8 @@
     ("Conv2d 2 LoRA", "Conv2d", LoraConfig, {"target_modules": ["conv2d", "lin0"]}),
     ("Conv2d 1 LoRA with DoRA", "Conv2d", LoraConfig, {"target_modules": ["conv2d"], "use_dora": True}),
     ("Conv2d 2 LoRA with DoRA", "Conv2d", LoraConfig, {"target_modules": ["conv2d", "lin0"], "use_dora": True}),
+    ("Conv2d Groups LoRA", "Conv2dGroups", LoraConfig, {"target_modules": ["conv2d"]}),
+    ("Conv2d Groups LoRA with DoRA", "Conv2dGroups", LoraConfig, {"target_modules": ["conv2d"], "use_dora": True}),
     ("Conv3d 1 LoRA", "Conv3d", LoraConfig, {"target_modules": ["conv3d"]}),
     ("Conv3d 2 LoRA", "Conv3d", LoraConfig, {"target_modules": ["conv3d", "lin0"]}),
     ("Conv3d 1 LoRA with DoRA", "Conv3d", LoraConfig, {"target_modules": ["conv3d"], "use_dora": True}),
@@ -903,6 +905,25 @@ def forward(self, X):
         return X
 
 
+class ModelConv2DGroups(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.conv2d = nn.Conv2d(5, 5, 3, groups=5)
+        self.relu = nn.ReLU()
+        self.flat = nn.Flatten()
+        self.lin0 = nn.Linear(5, 2)
+        self.sm = nn.LogSoftmax(dim=-1)
+
+    def forward(self, X):
+        X = X.float().reshape(-1, 5, 3, 3)
+        X = self.conv2d(X)
+        X = self.relu(X)
+        X = self.flat(X)
+        X = self.lin0(X)
+        X = self.sm(X)
+        return X
+
+
 class ModelConv3D(nn.Module):
     def __init__(self):
         super().__init__()
@@ -967,6 +988,9 @@ def from_pretrained(cls, model_id, torch_dtype=None):
         if model_id == "Conv2d":
             return ModelConv2D().to(torch_dtype)
 
+        if model_id == "Conv2dGroups":
+            return ModelConv2DGroups().to(torch_dtype)
+
         if model_id == "Conv3d":
             return ModelConv3D().to(torch_dtype)