FIX Bug when dequantizing 4bit bnb weights (#2847)

Fixes some failing GPU tests in CI.

A bug was introduced in #2797 where state.SCB was accessed while
dequantizing 4bit bnb weights even though state is None. This would
occur, for instance, when using DoRA, which needs to dequantize the
weight. The attribute access is now restricted to 8bit bnb weights.

Author: BenjaminBossan

src/peft/utils/integrations.py

@@ -88,16 +88,23 @@ def dequantize_bnb_weight(weight: torch.nn.Parameter, state=None):
     """Helper function to dequantize 4bit or 8bit bnb weights."""
     import bitsandbytes as bnb
 
-    if state.SCB is None:
-        state.SCB = weight.SCB
-
     device = weight.device
 
     cls_name = weight.__class__.__name__
     if cls_name == "Params4bit":
         dequantized = bnb.functional.dequantize_4bit(weight.data, weight.quant_state)
         return dequantized
 
+    # 8bit case
+    if state is None:
+        raise ValueError(
+            "No `state` was passed for bnb 8bit quantized weights. Please open an issue on the PEFT repository and "
+            "report the error: https://github.com/huggingface/peft/issues"
+        )
+
+    if state.SCB is None:
+        state.SCB = weight.SCB
+
     if hasattr(bnb.functional, "int8_vectorwise_dequant"):
         # Use bitsandbytes API if available (requires v0.45.0+)
         dequantized = bnb.functional.int8_vectorwise_dequant(weight.data, state.SCB)