chore(dependencies): update transformers to v4.46.3

pyproject.toml

@@ -42,7 +42,7 @@ keywords = [
 ]
 
 dependencies = [
-    "transformers == 4.41.1",
+    "transformers == 4.46.3",
     "torch == 2.5.1",
     "torch-xla[tpu] == 2.5.1",
     'typer == 0.6.1',

text-generation-inference/docker/Dockerfile

@@ -101,7 +101,7 @@ RUN apt-get update -y \
 RUN pip install --upgrade pip
 
 # Install HuggingFace packages
-ARG TRANSFORMERS_VERSION='4.41.1'
+ARG TRANSFORMERS_VERSION='4.46.3'
 ARG ACCELERATE_VERSION='1.1.1'
 ARG SAFETENSORS_VERSION='0.4.2'
 

text-generation-inference/server/pyproject.toml

@@ -15,7 +15,7 @@ dependencies = [
     'grpc-interceptor == 0.15.2',
     'typer == 0.6.1',
     'safetensors == 0.4.2',
-    'transformers == 4.41.1',
+    'transformers == 4.46.3',
     'loguru == 0.6.0',
     "sentencepiece == 0.2.0",
     "numpy<2.0",

text-generation-inference/server/text_generation_server/generator.py

@@ -314,6 +314,9 @@ def __init__(
         tokenizer.truncation_side = "left"
         self.tokenizer = tokenizer
         self.special_tokens = self.tokenizer.all_special_ids
+        # The token selector will use the model's generation mixin internal variables to select the next token, and it
+        # expects special tokens to be initialized in the model.
+        model._prepare_special_tokens(generation_config=model.generation_config, device=model.device)
         # Slots are empty to begin with, they will be populated as new batches arrive
         self.slots = []
         self.batch_id = 0

text-generation-inference/server/text_generation_server/jetstream_pt_support/generator.py

@@ -262,6 +262,10 @@ def __init__(
         tokenizer.truncation_side = "left"
         self.tokenizer = tokenizer
         self.special_tokens = self.tokenizer.all_special_ids
+        # The token selector will use the model's generation mixin internal variables to select the next token, and it
+        # expects special tokens to be initialized in the model.
+        model = self.engine.pt_model
+        model._prepare_special_tokens(generation_config=model.generation_config, device='cpu')
         # Slots number is static, it cannot grow over the size of the batch
         self.slots = [Slot(i, tokenizer) for i in range(self.model.config.batch_size)]
         self.batch_id = 0