Remove torch.compile for double_quant

bitsandbytes/__init__.py

@@ -25,7 +25,6 @@
     from .optim import adam
 
     register_backend("cuda", CUDABackend())
-
 __pdoc__ = {
     "libbitsandbytes": False,
     "optim.optimizer.Optimizer8bit": False,

bitsandbytes/backends/cpu_xpu_common.py

@@ -22,7 +22,7 @@ def _maybe_torch_compile(func):
     return func
 
 
-@_maybe_torch_compile
+# Don't use torch.compile for now due to PyTorch issue https://github.com/pytorch/pytorch/issues/124382
 def double_quant_impl(
     A, col_stats=None, row_stats=None, out_col=None, out_row=None, threshold=0.0
 ):

bitsandbytes/nn/modules.py

@@ -611,11 +611,10 @@ def to(self: T, tensor: Tensor, non_blocking: bool = ...) -> T: ...
     def to(self, *args, **kwargs):
         device, dtype, non_blocking, convert_to_format = torch._C._nn._parse_to(*args, **kwargs)
 
-        if device is not None and device.type == "cuda" and self.data.device.type == "cpu":
+        if device.type == "cuda" and self.data.device.type == "cpu":
             return self.cuda(device)
         elif (
-            device is not None
-            and device.type == "cpu"
+            device.type == "cpu"
             and self.data.dtype != torch.int8
         ):
             return self.cpu()