jinaai/jina-reranker-v2-base-multilingual · OSError: Unable to load weights from pytorch checkpoint file

Jul 8

OSError: Unable to load weights from pytorch checkpoint file for '/home/ubuntu/Training/Models/jina-reranker-v2-base-multilingual/pytorch_model.bin' at '/home/ubuntu/Training/Models/jina-reranker-v2-base-multilingual/pytorch_model.bin'. If you tried to load a PyTorch model from a TF 2.0 checkpoint, please set from_tf=True.

最新的两个commit(2f34eeb3099209d80b9ebf42f8a62713bf44ec6a, 298e48cada4a9318650d7fbd795f63827f884087)都报这个错

环境：
GPU: 4090
CUDA: Driver Version: 535.146.02 CUDA Version: 12.2
python: 3.10.13
torch:2.1.2
transformers:4.41.0
sentence-transformers:3.0.1

以下两个示例都报一样的错：

from transformers import AutoModelForSequenceClassification

model = AutoModelForSequenceClassification.from_pretrained(
    '/home/ubuntu/Training/Models/jina-reranker-v2-base-multilingual',
    torch_dtype="auto",
    trust_remote_code=True,
)

from sentence_transformers import CrossEncoder

model = CrossEncoder(
    "/home/ubuntu/Training/Models/jina-reranker-v2-base-multilingual",
    automodel_args={"torch_dtype": "auto"},
    trust_remote_code=True,
)

模型已经下载到本地的：

Wilson-Wen

Jul 8

好消息：跑通了
坏消息：玄学，不知道怎么回事（可能是hf下载脚本的问题，下载不完整或是什么）
官方脚本：本地下载的模型跑不起来，用model id让脚本重新下载可以跑；最后重新手动下载模型，指定路径让脚本加载，也能跑了
xinference：把cache删除，重新下载，能加载，发请求报了另一个flash_attn的错，python3.10死活装不上，换3.11，然后现在都能跑了
（上周五已经下载过很多次，今天也是下载过很多次，现在是奇迹般的都行了）

Wilson-Wen changed discussion status to closed Jul 8