rag_server/applications/clip_embedding/clip_model.py

import os
import torch

from transformers import AutoModel, AutoConfig, AutoTokenizer

MODEL_NAME = "BAAI/EVA-CLIP-8B-448"
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
DTYPE = torch.float16 if DEVICE == "cuda" else torch.float32
MAX_BATCH = int(os.getenv("MAX_BATCH", "32"))

print(f"[model_config] Loading {MODEL_NAME} on {DEVICE} dtype={DTYPE} ...")

config = AutoConfig.from_pretrained(MODEL_NAME, trust_remote_code=True)
model = AutoModel.from_pretrained(
    MODEL_NAME, config=config, trust_remote_code=True
).to(dtype=DTYPE, device=DEVICE).eval()


try:
    from transformers import CLIPImageProcessor
    image_processor = CLIPImageProcessor.from_pretrained(MODEL_NAME)
except Exception:
    print("[warning] EVA-CLIP 没有预处理配置，使用默认参数构造 ImageProcessor")
    from transformers import CLIPImageProcessor
    image_processor = CLIPImageProcessor(
        size={"shortest_edge": 448},  # 最短边 resize 到 448
        resample=3,
        crop_size={"height": 448, "width": 448},  # 中心裁剪 448×448
        image_mean=[0.48145466, 0.4578275, 0.40821073],
        image_std=[0.26862954, 0.26130258, 0.27577711]
    )

# 如果后续要做 text embedding，可以加 tokenizer
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)

def get_model():
    return model, image_processor, tokenizer, DEVICE, DTYPE, MAX_BATCH