diff --git a/bitsandbytes/autograd/_functions.py b/bitsandbytes/autograd/_functions.py
index b34f4a943..0258768e2 100644
--- a/bitsandbytes/autograd/_functions.py
+++ b/bitsandbytes/autograd/_functions.py
@@ -378,7 +378,8 @@ def matmul_4bit(
     out: Optional[torch.Tensor] = None,
     bias: Optional[torch.Tensor] = None,
 ):
-    assert quant_state is not None
+    if quant_state is None:
+        raise ValueError("quant_state cannot be None")
     # Change dtype to input dtype on CPU
     if A.device.type == "cpu":
         quant_state.dtype = A.dtype
diff --git a/bitsandbytes/functional.py b/bitsandbytes/functional.py
index 6ce846277..22109c3d2 100644
--- a/bitsandbytes/functional.py
+++ b/bitsandbytes/functional.py
@@ -707,7 +707,8 @@ def dequantize_blockwise(
             The dequantized tensor. The datatype is indicated by `quant_state.dtype` and defaults to `torch.float32`.
     """
 
-    assert quant_state is not None or absmax is not None
+    if quant_state is None and absmax is None:
+        raise ValueError("Either quant_state or absmax must be provided")
     if code is None and quant_state is None:
         if "dynamic" not in name2qmap:
             name2qmap["dynamic"] = create_dynamic_map().to(A.device)