Inconsistency in PreTrainedModel.resize_token_embeddings When ZeRO3 I…

…s Enabled (#25394)

* Inconsistency in PreTrainedModel.resize_token_embeddings

This PR addresses #25241.

In previous implementation when ZeRO stage 3 was enbaled, resize_token_embeddings would create independent PyTorch weights on each device. Here we ensure that new embeddings are created with DeepSpeed init, and are properly partitioned accros devices.

* formatting with black

* adding the removed comments back in

---------

Co-authored-by: Sina Moeini <smoeini@amazon.com>

src/transformers/modeling_utils.py

@@ -1502,24 +1502,40 @@ def _get_resized_embeddings(
                 f" {nn.Embedding}."
             )
 
-        # Build new embeddings
-        new_embeddings = nn.Embedding(new_num_tokens, old_embedding_dim)
-        new_embeddings.to(old_embeddings.weight.device, dtype=old_embeddings.weight.dtype)
-
-        # initialize all new embeddings (in particular added tokens)
-        self._init_weights(new_embeddings)
-
-        # Copy token embeddings from the previous weights
-
         # numbers of tokens to copy
         n = min(old_num_tokens, new_num_tokens)
         if is_deepspeed_zero3_enabled():
             import deepspeed
 
-            with deepspeed.zero.GatheredParameters(old_embeddings.weight, modifier_rank=0):
-                if torch.distributed.get_rank() == 0:
-                    new_embeddings.weight.data[:n, :] = old_embeddings.weight.data[:n, :]
+            with deepspeed.zero.Init(config_dict_or_path=deepspeed_config()):
+                # Build new embeddings
+                new_embeddings = nn.Embedding(
+                    new_num_tokens,
+                    old_embedding_dim,
+                    device=old_embeddings.weight.device,
+                    dtype=old_embeddings.weight.dtype,
+                )
+
+            params = [old_embeddings.weight, new_embeddings.weight]
+            with deepspeed.zero.GatheredParameters(params, modifier_rank=0):
+                # initialize all new embeddings (in particular added tokens)
+                self._init_weights(new_embeddings)
+
+                # Copy token embeddings from the previous weights
+                new_embeddings.weight.data[:n, :] = old_embeddings.weight.data[:n, :]
         else:
+            # Build new embeddings
+            new_embeddings = nn.Embedding(
+                new_num_tokens,
+                old_embedding_dim,
+                device=old_embeddings.weight.device,
+                dtype=old_embeddings.weight.dtype,
+            )
+
+            # initialize all new embeddings (in particular added tokens)
+            self._init_weights(new_embeddings)
+
+            # Copy token embeddings from the previous weights
             new_embeddings.weight.data[:n, :] = old_embeddings.weight.data[:n, :]
 
         return new_embeddings
@@ -1575,35 +1591,40 @@ def _get_resized_lm_head(
         # Build new lm head
         new_lm_head_shape = (old_lm_head_dim, new_num_tokens) if not transposed else (new_num_tokens, old_lm_head_dim)
         has_new_lm_head_bias = old_lm_head.bias is not None
-        new_lm_head = nn.Linear(*new_lm_head_shape, bias=has_new_lm_head_bias)
-        new_lm_head = new_lm_head.to(old_lm_head.weight.device, dtype=old_lm_head.weight.dtype)
-
-        # initialize new lm head (in particular added tokens)
-        self._init_weights(new_lm_head)
 
         num_tokens_to_copy = min(old_num_tokens, new_num_tokens)
 
         # XXX: put the long block of code in a wrapper
         if is_deepspeed_zero3_enabled():
             import deepspeed
 
+            with deepspeed.zero.Init(config_dict_or_path=deepspeed_config()):
+                new_lm_head = nn.Linear(
+                    *new_lm_head_shape,
+                    bias=has_new_lm_head_bias,
+                    device=old_lm_head.weight.device,
+                    dtype=old_lm_head.weight.dtype,
+                )
             params = [old_lm_head.weight, old_lm_head.bias, new_lm_head.weight, new_lm_head.bias]
             with deepspeed.zero.GatheredParameters(params, modifier_rank=0):
-                if torch.distributed.get_rank() == 0:
-                    # Copy old lm head weights to new lm head
-                    if not transposed:
-                        new_lm_head.weight.data[:num_tokens_to_copy, :] = old_lm_head.weight.data[
-                            :num_tokens_to_copy, :
-                        ]
-                    else:
-                        new_lm_head.weight.data[:, :num_tokens_to_copy] = old_lm_head.weight.data[
-                            :, :num_tokens_to_copy
-                        ]
+                self._init_weights(new_lm_head)
+                # Copy old lm head weights to new lm head
+                if not transposed:
+                    new_lm_head.weight.data[:num_tokens_to_copy, :] = old_lm_head.weight.data[:num_tokens_to_copy, :]
+                else:
+                    new_lm_head.weight.data[:, :num_tokens_to_copy] = old_lm_head.weight.data[:, :num_tokens_to_copy]
 
-                    # Copy bias weights to new lm head
-                    if has_new_lm_head_bias:
-                        new_lm_head.bias.data[:num_tokens_to_copy] = old_lm_head.bias.data[:num_tokens_to_copy]
+                # Copy bias weights to new lm head
+                if has_new_lm_head_bias:
+                    new_lm_head.bias.data[:num_tokens_to_copy] = old_lm_head.bias.data[:num_tokens_to_copy]
         else:
+            new_lm_head = nn.Linear(
+                *new_lm_head_shape,
+                bias=has_new_lm_head_bias,
+                device=old_lm_head.weight.device,
+                dtype=old_lm_head.weight.dtype,
+            )
+            self._init_weights(new_lm_head)
             # Copy old lm head weights to new lm head
             if not transposed:
                 new_lm_head.weight.data[:num_tokens_to_copy, :] = old_lm_head.weight.data[:num_tokens_to_copy, :]