diff --git a/unsloth/device_type.py b/unsloth/device_type.py
index 547750019a..e04a3c55b8 100644
--- a/unsloth/device_type.py
+++ b/unsloth/device_type.py
@@ -19,10 +19,12 @@ __all__ = [
     "DEVICE_TYPE_TORCH",
     "DEVICE_COUNT",
     "ALLOW_PREQUANTIZED_MODELS",
+    "ALLOW_BITSANDBYTES",
 ]
 
 import torch
 import functools
+from unsloth_zoo.utils import Version
 
 @functools.cache
 def is_hip():
@@ -70,11 +72,15 @@ DEVICE_COUNT : int = get_device_count()
 # Check blocksize for 4bit -> 64 for CUDA, 128 for AMD
 # If AMD, we cannot load pre-quantized models for now :(
 ALLOW_PREQUANTIZED_MODELS : bool = True
+# HSA_STATUS_ERROR_EXCEPTION checks - sometimes AMD fails for BnB
+ALLOW_BITSANDBYTES : bool = True
 if DEVICE_TYPE == "hip":
     try:
         from bitsandbytes.nn.modules import Params4bit
         if "blocksize = 64 if not HIP_ENVIRONMENT else 128" in inspect.getsource(Params4bit):
             ALLOW_PREQUANTIZED_MODELS = False
+        import bitsandbytes
+        ALLOW_BITSANDBYTES = Version(bitsandbytes.__version__) > Version("0.48.2.dev0")
     except:
         pass
 pass
diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index 5d1896fd5b..d8548ded24 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -52,6 +52,7 @@ from ..device_type import (
     DEVICE_TYPE_TORCH,
     DEVICE_COUNT,
     ALLOW_PREQUANTIZED_MODELS,
+    ALLOW_BITSANDBYTES,
 )
 
 # https://github.com/huggingface/transformers/pull/26037 allows 4 bit loading!
@@ -199,6 +200,10 @@ class FastLanguageModel(FastLlamaModel):
                 )
             pass
         pass
+        # Check if 4bit is allowed specifically for AMD
+        if not ALLOW_BITSANDBYTES:
+            print("Unsloth: AMD currently is not stable with 4bit bitsandbytes. Disabling for now.")
+            load_in_4bit = False
 
         old_model_name = model_name
         if not use_exact_model_name:
@@ -638,6 +643,11 @@ class FastModel(FastBaseModel):
         old_model_name = model_name
         if not use_exact_model_name:
             model_name = get_model_name(model_name, load_in_4bit)
+
+        # Check if 4bit is allowed specifically for AMD
+        if not ALLOW_BITSANDBYTES:
+            print("Unsloth: AMD currently is not stable with 4bit bitsandbytes. Disabling for now.")
+            load_in_4bit = False
         # Check if pre-quantized models are allowed
         # For eg AMD GPUs need blocksize = 128, but our pre-quants are blocksize = 64
         if not ALLOW_PREQUANTIZED_MODELS and model_name.endswith(("-unsloth-bnb-4bit", "-bnb-4bit")):