Bump up transformers version & Remove MistralConfig (#1254)

Author: WoosukKwon

requirements.txt

@@ -6,8 +6,8 @@ pyarrow  # Required for Ray data.
 sentencepiece  # Required for LLaMA tokenizer.
 numpy
 torch == 2.0.1
-transformers >= 4.33.1  # Required for Code Llama.
-xformers == 0.0.22
+transformers >= 4.34.0  # Required for Mistral.
+xformers == 0.0.22  # Required for Mistral.
 fastapi
 uvicorn[standard]
 pydantic < 2  # Required for OpenAI server.

vllm/model_executor/models/__init__.py

@@ -9,10 +9,10 @@
 from vllm.model_executor.models.gpt_neox import GPTNeoXForCausalLM
 from vllm.model_executor.models.internlm import InternLMForCausalLM
 from vllm.model_executor.models.llama import LlamaForCausalLM
+from vllm.model_executor.models.mistral import MistralForCausalLM
 from vllm.model_executor.models.mpt import MPTForCausalLM
 from vllm.model_executor.models.opt import OPTForCausalLM
 from vllm.model_executor.models.qwen import QWenLMHeadModel
-from vllm.model_executor.models.mistral import MistralForCausalLM
 
 __all__ = [
     "AquilaForCausalLM",

vllm/model_executor/models/mistral.py

@@ -29,6 +29,7 @@
 
 import torch
 from torch import nn
+from transformers import MistralConfig
 
 from vllm.model_executor.input_metadata import InputMetadata
 from vllm.model_executor.layers.activation import SiluAndMul
@@ -44,7 +45,6 @@
     convert_pyslice_to_tensor, hf_model_weights_iterator,
     load_tensor_parallel_weights, load_padded_tensor_parallel_vocab)
 from vllm.sequence import SamplerOutput
-from vllm.transformers_utils.configs.mistral import MistralConfig
 
 KVCache = Tuple[torch.Tensor, torch.Tensor]
 

vllm/transformers_utils/config.py

@@ -17,15 +17,6 @@
 def get_config(model: str,
                trust_remote_code: bool,
                revision: Optional[str] = None) -> PretrainedConfig:
-    # NOTE: Because the Mistral model in HF hub does not have
-    # `configuration_mistral.py`, we cannot use `AutoConfig` to load the
-    # config. Instead, we use `MistralConfig` directly.
-    # NOTE: This is a hack. This does not work for local models.
-    # FIXME: Remove this once the Mistral model is available in the stable
-    # version of HF transformers.
-    if "mistral" in model.lower():
-        return MistralConfig.from_pretrained(model, revision=revision)
-
     try:
         config = AutoConfig.from_pretrained(
             model, trust_remote_code=trust_remote_code, revision=revision)

vllm/transformers_utils/configs/__init__.py

@@ -6,13 +6,11 @@
 # 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.mistral import MistralConfig
 
 __all__ = [
     "MPTConfig",
     "BaiChuanConfig",
     "AquilaConfig",
     "QWenConfig",
     "RWConfig",
-    "MistralConfig",
 ]