diff --git a/auto_round/compressors/base.py b/auto_round/compressors/base.py
index d9f1e6054..ccb12cf42 100644
--- a/auto_round/compressors/base.py
+++ b/auto_round/compressors/base.py
@@ -26,7 +26,7 @@
 import accelerate
 import torch
 from accelerate.big_modeling import dispatch_model, infer_auto_device_map
-from accelerate.utils import get_max_memory
+from accelerate.utils import get_balanced_memory, get_max_memory
 from torch import autocast
 from tqdm import tqdm
 from transformers import set_seed
@@ -1770,7 +1770,7 @@ def calib(self, nsamples, bs):
                     data_new[key] = data[key].to(self.model.device)
                 input_ids = data_new["input_ids"]
             elif isinstance(data, tuple) or isinstance(data, list):
-                data_new = to_device(data)
+                data_new = to_device(data, self.model.device)
                 input_ids = data_new[0]
             else:
                 data_new = {}
@@ -1904,6 +1904,7 @@ def try_cache_inter_data_gpucpu(self, block_names, nsamples, layer_names=None, l
                         if str(self.model.device) == "cpu" and (not self.device.startswith("hpu")):
                             no_split_modules = getattr(self.model, "_no_split_modules", [])
                             devices = parse_available_devices(self.device_map)
+
                             max_memory = get_max_memory()
                             new_max_memory = {}
                             if "cpu" not in devices:
@@ -1915,13 +1916,21 @@ def try_cache_inter_data_gpucpu(self, block_names, nsamples, layer_names=None, l
                                     device = "cpu"
                                 else:
                                     raise ValueError(f"Unsupported device {device} in device_map: {self.device_map}")
-                                new_max_memory[device] = max_memory[device]
+                                # Use 90% of the reported max memory to leave headroom for activations,
+                                # temporary tensors, other processes, and allocator fragmentation, reducing
+                                # the chance of runtime OOM while still utilizing most available memory.
+                                new_max_memory[device] = max_memory[device] * 0.9
+                            new_max_memory = get_balanced_memory(
+                                self.model,
+                                max_memory=new_max_memory,
+                                no_split_module_classes=no_split_modules,
+                            )
                             device_map = infer_auto_device_map(
                                 self.model, max_memory=new_max_memory, no_split_module_classes=no_split_modules
                             )
                             if len(devices) > 1 and "cpu" in device_map.values():
                                 logger.warning(
-                                    "Not enough vram cause the ram to be used, which may severely impact speed."
+                                    "Some layers are offloaded to cpu, which may severely impact calibration speed."
                                     " Please consider using more cards."
                                 )
 
diff --git a/auto_round/compressors/mllm/compressor.py b/auto_round/compressors/mllm/compressor.py
index 31f97cbe3..67180555a 100644
--- a/auto_round/compressors/mllm/compressor.py
+++ b/auto_round/compressors/mllm/compressor.py
@@ -345,7 +345,7 @@ def calib(self, nsamples, bs):
                     pbar.update(1)
                     continue
                 if isinstance(data, torch.Tensor):
-                    input_ids = data.to(self.device)
+                    input_ids = data.to(self.model.device)
                     data_new = input_ids
                 elif isinstance(data, str):
                     if self.tokenizer is None:
@@ -360,7 +360,7 @@ def calib(self, nsamples, bs):
                     )
                     data_new = {}
                     for key in data.keys():
-                        data_new[key] = data[key].to(self.device)
+                        data_new[key] = data[key].to(self.model.device)
                     input_ids = data_new["input_ids"]
                 elif isinstance(data, dict) and "text" in data.keys():
                     text = data["text"]
@@ -381,7 +381,7 @@ def calib(self, nsamples, bs):
                             data_new[key] = to_dtype(data_new[key], self.model.dtype)
                     input_ids = data_new["input_ids"]
                 elif isinstance(data, tuple) or isinstance(data, list):
-                    data_new = data
+                    data_new = to_device(data, self.model.device)
                     input_ids = data_new[0]
                 else:
                     data_new = {}
diff --git a/auto_round/modelling/qwen3_vl_moe.py b/auto_round/modelling/qwen3_vl_moe.py
index dbed42b1a..2e38bb740 100644
--- a/auto_round/modelling/qwen3_vl_moe.py
+++ b/auto_round/modelling/qwen3_vl_moe.py
@@ -67,8 +67,8 @@ def __init__(
         self.gate = original.gate
         self.calibrate_all_experts = calibrate_all_experts
         self.experts = SequentialQwen3VLMoeTextExperts(text_config, original.experts)
-        if not transformers_version <= version.parse(
-            "4.57.3"
+        if not transformers_version < version.parse(
+            "5.0"
         ):  # remove conversion_mapping for qwen3_vl_moe when transformers>=5.0
             from transformers.conversion_mapping import register_checkpoint_conversion_mapping