Fix autocast_adapter_dtype=False for quantized models

src/peft/tuners/lora/layer.py

@@ -127,6 +127,24 @@ def __init__(self, base_layer: nn.Module, ephemeral_gpu_offload: bool = False, *
         self.in_features = in_features
         self.out_features = out_features
 
+    def _get_weight_dtype(self) -> torch.dtype:
+        """
+        Get the dtype for initializing LoRA parameters.
+        Handles both regular and quantized layers.
+        """
+        base_layer = self.get_base_layer()
+
+        # For quantized layers (BitsAndBytes) - check compute_dtype first
+        if hasattr(base_layer, 'compute_dtype'):
+            return base_layer.compute_dtype
+
+        # For regular layers, use weight dtype
+        if hasattr(base_layer, 'weight'):
+            return base_layer.weight.dtype
+
+        # Fallback to float32
+        return torch.float32
+
     def resolve_lora_variant(self, *, use_dora: bool, **kwargs) -> Optional[LoraVariant]:
         """Return a matching LoRA variant for this layer type.
 
@@ -193,8 +211,9 @@ def update_layer(
         self.lora_dropout.update(nn.ModuleDict({adapter_name: lora_dropout_layer}))
 
         # Actual trainable parameters
-        self.lora_A[adapter_name] = nn.Linear(self.in_features, r, bias=False)
-        self.lora_B[adapter_name] = nn.Linear(r, self.out_features, bias=lora_bias)
+        weight_dtype = self._get_weight_dtype()
+        self.lora_A[adapter_name] = nn.Linear(self.in_features, r, bias=False, dtype=weight_dtype)
+        self.lora_B[adapter_name] = nn.Linear(r, self.out_features, bias=lora_bias, dtype=weight_dtype)
         self.lora_bias[adapter_name] = lora_bias
 
         if use_rslora: