Fix rtn tuning_device issue (#893)

Signed-off-by: Kaihui-intel <kaihui.tang@intel.com>

Author: Kaihui-intel

auto_round/compressors/base.py

@@ -1422,7 +1422,7 @@ def _quantize_layer_via_rtn(self, name: str) -> None:
             m.zp = None
         else:
             try:
-                m = m.to(self.device)
+                m = m.to(m.tuning_device if hasattr(m, "tuning_device") else self.device)
                 m = WrapperLinear(
                     m,
                     enable_minmax_tuning=False,

test/test_cuda/test_multiple_card.py

@@ -242,6 +242,19 @@ def test_device_map_dict(self):
         )
         autoround.quantize()
 
+        # test rtn
+        autoround = AutoRound(
+            model_name,
+            tokenizer,
+            bits=bits,
+            group_size=group_size,
+            sym=sym,
+            iters=0,
+            seqlen=2,
+            device_map=device_map,
+        )
+        autoround.quantize()
+
     @multi_card
     @require_greater_than_050
     def test_device_map_for_triton(self):