diff --git a/src/accelerate/utils/modeling.py b/src/accelerate/utils/modeling.py index c1483355716..e471731a4e9 100644 --- a/src/accelerate/utils/modeling.py +++ b/src/accelerate/utils/modeling.py @@ -476,7 +476,11 @@ def infer_auto_device_map( current_memory_used = 0 # Direct submodules and parameters - modules_to_treat = list(model.named_parameters(recurse=False)) + list(model.named_children()) + modules_to_treat = ( + list(model.named_parameters(recurse=False)) + + list(model.named_children()) + + list(model.named_buffers(recurse=False)) + ) # Initialize maximum largest layer, to know which space to keep in memory max_layer_size, max_layer_names = get_max_layer_size(modules_to_treat, module_sizes, no_split_module_classes)