Enforce `max_memory` for device_map strategies (#22311)

Enforce  for device_map strategies

src/transformers/modeling_utils.py

@@ -2571,6 +2571,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                     model,
                     dtype=torch_dtype,
                     low_zero=(device_map == "balanced_low_0"),
+                    max_memory=max_memory,
                     **kwargs,
                 )
             kwargs["max_memory"] = max_memory