fixes to properly shard FSDP across cpu and meta for cpu_efficient_lo…

…ading for prequantized 4bit (huggingface#32276)

src/transformers/modeling_utils.py

@@ -932,6 +932,8 @@ def _load_state_dict_into_meta_model(
                 )
             )
         ):
+            if is_fsdp_enabled():
+                param_device = "cpu" if is_local_dist_rank_0() else "meta"
             # For backward compatibility with older versions of `accelerate` and for non-quantized params
             set_module_tensor_to_device(model, param_name, param_device, **set_module_kwargs)
         else:
@@ -942,7 +944,10 @@ def _load_state_dict_into_meta_model(
             if is_fsdp_enabled() or is_deepspeed_zero3_enabled():
                 module, tensor_name = get_module_from_name(model, param_name)
                 value = getattr(module, tensor_name)
-                value = type(value)(value.data.to("cpu"), **value.__dict__)
+                param_to = "cpu"
+                if is_fsdp_enabled() and not is_local_dist_rank_0():
+                    param_to = "meta"
+                value = type(value)(value.data.to(param_to), **value.__dict__)
                 setattr(module, tensor_name, value)
             # TODO: consider removing used param_parts from state_dict before return
 

src/transformers/quantizers/quantizer_bnb_4bit.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import importlib
+import inspect
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Union
 
 from packaging import version
@@ -199,11 +200,16 @@ def create_quantized_param(
                     if unexpected_keys is not None and k in unexpected_keys:
                         unexpected_keys.remove(k)
 
+            param_kwargs = {}
+            sig = inspect.signature(bnb.nn.Params4bit.from_prequantized)
+            if "module" in sig.parameters:
+                param_kwargs["module"] = module
             new_value = bnb.nn.Params4bit.from_prequantized(
                 data=param_value,
                 quantized_stats=quantized_stats,
                 requires_grad=False,
                 device=target_device,
+                **param_kwargs,
             )
         else:
             new_value = param_value.to("cpu")