[Misc] Bump up transformers to v4.39.0 & Remove StarCoder2Config

requirements-rocm.txt

@@ -7,7 +7,7 @@ ray >= 2.9
 sentencepiece  # Required for LLaMA tokenizer.
 numpy
 tokenizers>=0.15.0
-transformers >= 4.38.0  # Required for Gemma.
+transformers >= 4.39.0  # Required for StarCoder2.
 fastapi
 uvicorn[standard]
 pydantic >= 2.0  # Required for OpenAI server.

requirements.txt

@@ -5,7 +5,7 @@ ray >= 2.9
 sentencepiece  # Required for LLaMA tokenizer.
 numpy
 torch == 2.1.2
-transformers >= 4.38.0  # Required for Gemma.
+transformers >= 4.39.0  # Required for StarCoder2.
 xformers == 0.0.23.post1  # Required for CUDA 12.1.
 fastapi
 uvicorn[standard]

vllm/model_executor/models/starcoder2.py

@@ -22,6 +22,7 @@
 
 import torch
 from torch import nn
+from transformers import Starcoder2Config
 
 from vllm.model_executor.input_metadata import InputMetadata
 from vllm.model_executor.sampling_metadata import SamplingMetadata
@@ -42,13 +43,6 @@
                                               hf_model_weights_iterator)
 from vllm.sequence import SamplerOutput
 
-try:
-    from transformers import Starcoder2Config
-except ImportError:
-    # fallback to PretrainedConfig
-    # NOTE: Please install transformers from source or use transformers>=4.39.0
-    from transformers import PretrainedConfig as Starcoder2Config
-
 KVCache = Tuple[torch.Tensor, torch.Tensor]
 
 

vllm/transformers_utils/config.py

@@ -9,7 +9,6 @@
     "mpt": MPTConfig,
     "RefinedWeb": RWConfig,  # For tiiuae/falcon-40b(-instruct)
     "RefinedWebModel": RWConfig,  # For tiiuae/falcon-7b(-instruct)
-    "starcoder2": Starcoder2Config,
     "jais": JAISConfig,
 }
 
@@ -18,15 +17,6 @@ def get_config(model: str,
                trust_remote_code: bool,
                revision: Optional[str] = None,
                code_revision: Optional[str] = None) -> PretrainedConfig:
-    # FIXME(woosuk): This is a temporary fix for StarCoder2.
-    # Remove this when the model is supported by HuggingFace transformers.
-    if "bigcode" in model and "starcoder2" in model:
-        config_class = _CONFIG_REGISTRY["starcoder2"]
-        config = config_class.from_pretrained(model,
-                                              revision=revision,
-                                              code_revision=code_revision)
-        return config
-
     try:
         config = AutoConfig.from_pretrained(
             model,

vllm/transformers_utils/configs/__init__.py

@@ -4,13 +4,11 @@
 # tiiuae/falcon-7b(-instruct) models. Newer Falcon models will use the
 # `FalconConfig` class from the official HuggingFace transformers library.
 from vllm.transformers_utils.configs.falcon import RWConfig
-from vllm.transformers_utils.configs.starcoder2 import Starcoder2Config
 from vllm.transformers_utils.configs.jais import JAISConfig
 
 __all__ = [
     "ChatGLMConfig",
     "MPTConfig",
     "RWConfig",
-    "Starcoder2Config",
     "JAISConfig",
 ]