Fix Params4bit tensor subclass handling

bitsandbytes/nn/modules.py

@@ -356,6 +356,46 @@ class Params4bit(torch.nn.Parameter):
 
             return new_param
 
+    @classmethod
+    def __torch_function__(cls, func, types, args=(), kwargs=None):
+        if kwargs is None:
+            kwargs = {}
+    
+        if func in [torch.chunk, torch.split]:
+            tensor = args[0]
+            
+            result = super().__torch_function__(func, types, args, kwargs)
+    
+            if isinstance(result, tuple):
+                return tuple(
+                    cls(
+                        data=chunk,
+                        requires_grad=tensor.requires_grad,
+                        quant_state=tensor.quant_state,
+                        blocksize=tensor.blocksize,
+                        compress_statistics=tensor.compress_statistics,
+                        quant_type=tensor.quant_type,
+                        quant_storage=tensor.quant_storage,
+                        module=tensor.module,
+                        bnb_quantized=tensor.bnb_quantized,
+                    )
+                    for chunk in result
+                )
+            else:
+                return cls(
+                    data=result,
+                    requires_grad=tensor.requires_grad,
+                    quant_state=tensor.quant_state,
+                    blocksize=tensor.blocksize,
+                    compress_statistics=tensor.compress_statistics,
+                    quant_type=tensor.quant_type,
+                    quant_storage=tensor.quant_storage,
+                    module=tensor.module,
+                    bnb_quantized=tensor.bnb_quantized,
+                )
+        
+        return super().__torch_function__(func, types, args, kwargs)
+        
 
 def fix_4bit_weight_quant_state_from_module(module: Union["Embedding4bit", "Linear4bit"]):
     if getattr(module.weight, "quant_state", None) is not None: