Patch release v4.43.4

* fix resize when deepspeed

* deepsped uses new embeds

* we needed this

setup.py

@@ -430,7 +430,7 @@ install_requires = [
 
 setup(
     name="transformers",
-    version="4.43.3",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
+    version="4.43.4",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
     author="The Hugging Face team (past and future) with the help of all our contributors (https://github.com/huggingface/transformers/graphs/contributors)",
     author_email="transformers@huggingface.co",
     description="State-of-the-art Machine Learning for JAX, PyTorch and TensorFlow",

src/transformers/__init__.py

@@ -18,7 +18,7 @@
 # to defer the actual importing for when the objects are requested. This way `import transformers` provides the names
 # in the namespace without actually importing anything (and especially none of the backends).
 
-__version__ = "4.43.3"
+__version__ = "4.43.4"
 
 from typing import TYPE_CHECKING
 

src/transformers/modeling_utils.py

@@ -1980,12 +1980,22 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
         if new_num_tokens is None and pad_to_multiple_of is None:
             return model_embeds
 
+        # Since we are basically resuing the same old embeddings with new weight values, gathering is required
+        is_quantized = hasattr(self, "hf_quantizer") and self.hf_quantizer is not None
+        if is_deepspeed_zero3_enabled() and not is_quantized:
+            import deepspeed
+
+            with deepspeed.zero.GatheredParameters(model_embeds.weight, modifier_rank=None):
+                vocab_size = model_embeds.weight.shape[0]
+        else:
+            vocab_size = model_embeds.weight.shape[0]
+
         # Update base model and current model config
         if hasattr(self.config, "text_config"):
-            self.config.text_config.vocab_size = model_embeds.weight.shape[0]
+            self.config.text_config.vocab_size = vocab_size
         else:
-            self.config.vocab_size = model_embeds.weight.shape[0]
-        self.vocab_size = model_embeds.weight.shape[0]
+            self.config.vocab_size = vocab_size
+        self.vocab_size = vocab_size
 
         # Tie weights again if needed
         self.tie_weights()
@@ -2139,7 +2149,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
 
             params = [old_embeddings.weight, new_embeddings.weight]
             with deepspeed.zero.GatheredParameters(params, modifier_rank=0):
-                old_embeddings.weight.data = new_embeddings.weight.data
+                old_embeddings.weight = new_embeddings.weight
                 old_embeddings.num_embeddings = new_embeddings.weight.data.shape[0]
 
                 # If the new number of tokens is smaller than the original `padding_idx`, the `padding_idx`