Don't touch DS

src/transformers/modeling_utils.py

@@ -679,13 +679,45 @@ def _load_state_dict_into_model(model_to_load, state_dict, start_prefix):
 
     error_msgs = []
 
-    # Adjust and remove our `start_prefix` as we don't need it anymore
-    state_dict = {
-        key[len(start_prefix) :] if key.startswith(start_prefix) else key: value for key, value in state_dict.items()
-    }
-    # By passing in `assign=True`, we can be memory efficient by mapping the tensors directly, using only 1x
-    # the memory of the original state_dict instead of 2.
-    model_to_load.load_state_dict(state_dict, assign=True, strict=False)
+    # Note: for now this is only for DeepSpeed Zero3
+    # PyTorch's `_load_from_state_dict` does not copy parameters in a module's descendants
+    # so we need to apply the function recursively.
+    def load(module: nn.Module, state_dict, prefix=""):
+        local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})
+        args = (state_dict, prefix, local_metadata, True, [], [], error_msgs)
+        # Parameters of module and children will start with prefix. We can exit early if there are none in this
+        # state_dict
+        if len([key for key in state_dict if key.startswith(prefix)]) > 0:
+            if is_deepspeed_zero3_enabled():
+                import deepspeed
+
+                # In sharded models, each shard has only part of the full state_dict, so only gather
+                # parameters that are in the current state_dict.
+                named_parameters = dict(module.named_parameters(prefix=prefix[:-1], recurse=False))
+                params_to_gather = [named_parameters[k] for k in state_dict.keys() if k in named_parameters]
+                if len(params_to_gather) > 0:
+                    # because zero3 puts placeholders in model params, this context
+                    # manager gathers (unpartitions) the params of the current layer, then loads from
+                    # the state dict and then re-partitions them again
+                    with deepspeed.zero.GatheredParameters(params_to_gather, modifier_rank=0):
+                        if torch.distributed.get_rank() == 0:
+                            module._load_from_state_dict(*args)
+
+        for name, child in module._modules.items():
+            if child is not None:
+                load(child, state_dict, prefix + name + ".")
+
+    if is_deepspeed_zero3_enabled():
+        load(model_to_load, state_dict, prefix=start_prefix)
+    else:
+        # Adjust and remove our `start_prefix` as we don't need it anymore
+        state_dict = {
+            key[len(start_prefix) :] if key.startswith(start_prefix) else key: value
+            for key, value in state_dict.items()
+        }
+        # By passing in `assign=True`, we can be memory efficient by mapping the tensors directly, using only 1x
+        # the memory of the original state_dict instead of 2.
+        model_to_load.load_state_dict(state_dict, assign=True, strict=False)
     # Delete `state_dict` so it could be collected by GC earlier. Note that `state_dict` is a copy of the argument, so
     # it's safe to delete it.
     del state_dict