[Misc] Bump up transformers to v4.39.0 & Remove StarCoder2Config (#3551)

Co-authored-by: Roy <jasonailu87@gmail.com>
Co-authored-by: Roger Meier <r.meier@siemens.com>

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 @@ from typing import List, Optional, Tuple
 
 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 @@ from vllm.model_executor.weight_utils import (default_weight_loader,
                                               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 @@ _CONFIG_REGISTRY = {
     "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 @@ from vllm.transformers_utils.configs.mpt import MPTConfig
 # 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",
 ]

vllm/transformers_utils/configs/starcoder2.py

@@ -1,55 +0,0 @@
-from transformers import PretrainedConfig
-
-
-class Starcoder2Config(PretrainedConfig):
-    model_type = "starcoder2"
-    keys_to_ignore_at_inference = ["past_key_values"]
-
-    def __init__(
-        self,
-        vocab_size=49152,
-        hidden_size=3072,
-        intermediate_size=12288,
-        num_hidden_layers=30,
-        num_attention_heads=24,
-        num_key_value_heads=2,
-        hidden_act="gelu_pytorch_tanh",
-        max_position_embeddings=4096,
-        initializer_range=0.018042,
-        norm_epsilon=1e-5,
-        use_cache=True,
-        bos_token_id=50256,
-        eos_token_id=50256,
-        rope_theta=10000.0,
-        sliding_window=None,
-        attention_dropout=0.0,
-        residual_dropout=0.0,
-        embedding_dropout=0.0,
-        use_bias=True,
-        **kwargs,
-    ):
-        self.vocab_size = vocab_size
-        self.max_position_embeddings = max_position_embeddings
-        self.hidden_size = hidden_size
-        self.intermediate_size = intermediate_size
-        self.num_hidden_layers = num_hidden_layers
-        self.num_attention_heads = num_attention_heads
-        self.sliding_window = sliding_window
-        self.use_bias = use_bias
-        self.num_key_value_heads = num_key_value_heads
-        self.hidden_act = hidden_act
-        self.initializer_range = initializer_range
-        self.norm_epsilon = norm_epsilon
-        self.use_cache = use_cache
-        self.rope_theta = rope_theta
-        self.attention_dropout = attention_dropout
-        self.residual_dropout = residual_dropout
-        self.embedding_dropout = embedding_dropout
-
-        super().__init__(
-            bos_token_id=bos_token_id,
-            eos_token_id=eos_token_id,
-            **kwargs,
-        )
-        if self.architectures is None:
-            self.architectures = ['Starcoder2ForCausalLM']