02

端侧能力与效率

modelscope download --model OpenBMB/MiniCPM-V-4.7-35B-A3B \  --local_dir ./MiniCPM-V-4.7-35B-A3B

pip install "transformers[torch]>=5.18.0" torchvision torchcodec

import torchfrom transformers import AutoModelForImageTextToText, AutoProcessor
model_id = "openbmb/MiniCPM-V-4.7-35B-A3B"
processor = AutoProcessor.from_pretrained(model_id)model = AutoModelForImageTextToText.from_pretrained(    model_id, torch_dtype="auto", device_map="auto")# 推荐使用 Flash Attention 2 以获得更好的加速与显存节省，# 尤其在多图和视频场景下效果显著。# model = AutoModelForImageTextToText.from_pretrained(#     model_id,#     torch_dtype=torch.bfloat16,#     attn_implementation="flash_attention_2",#     device_map="auto",# )

messages = [    {        "role": "user",        "content": [            {"type": "image", "url": "https://huggingface.co/datasets/openbmb/DemoCase/resolve/main/refract.png"},            {"type": "text", "text": "What causes this phenomenon?"},        ],    }]downsample_mode = "16x"  # Using `downsample_mode="4x"` for Finer Detailinputs = processor.apply_chat_template(    messages, tokenize=True, add_generation_prompt=True,    return_dict=True, return_tensors="pt",    downsample_mode=downsample_mode,    max_slice_nums=36,    enable_thinking=False,).to(model.device)generated_ids = model.generate(**inputs, downsample_mode=downsample_mode, max_new_tokens=512)generated_ids_trimmed = [    out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]output_text = processor.batch_decode(    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)print(output_text[0])

messages = [    {        "role": "user",        "content": [            {"type": "video", "url": "https://huggingface.co/datasets/openbmb/DemoCase/resolve/main/football.mp4"},            {"type": "text", "text": "Describe this video in detail. Follow the timeline and focus on on-screen text, interface changes, main actions, and scene changes."},        ],    }]downsample_mode = "16x"  # Using `downsample_mode="4x"` for Finer Detailinputs = processor.apply_chat_template(    messages, tokenize=True, add_generation_prompt=True,    return_dict=True, return_tensors="pt",    downsample_mode=downsample_mode,    max_num_frames=128,    stack_frames=1,    max_slice_nums=1,    use_image_id=False,    enable_thinking=False,).to(model.device)generated_ids = model.generate(**inputs, downsample_mode=downsample_mode, max_new_tokens=2048)generated_ids_trimmed = [    out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]output_text = processor.batch_decode(    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)print(output_text[0])

ms-swift 支持使用对 MiniCPM-V-4.7-1B / 35B-A3B 进行训练。ms-swift 是魔搭社区官方提供的大模型训练框架，开源地址：https://github.com/modelscope/ms-swift

环境准备：

pip install git+https://github.com/modelscope/ms-swift.gitpip install "transformers>=5.18.0" -U# 更多依赖安装参考: https://swift.readthedocs.io/zh-cn/latest/Instruction/安装.html

多模态 SFT 脚本如下

# 16GiBCUDA_VISIBLE_DEVICES=0 \swift sft \--model OpenBMB/MiniCPM-V-4.7-1B \--dataset 'AI-ModelScope/LaTeX_OCR:human_handwrite#2000' \--load_from_cache_file true \--tuner_type full \--torch_dtype bfloat16 \--attn_impl sdpa \--max_length 2048 \--num_train_epochs 2 \--per_device_train_batch_size 8 \--gradient_accumulation_steps 1 \--learning_rate 1e-5 \--lr_scheduler_type cosine \--split_dataset_ratio 0 \--logging_steps 1 \--save_steps 100 \--dataset_num_proc 4 \--dataloader_num_workers 4 \--output_dir output

训练后对验证集进行推理：

CUDA_VISIBLE_DEVICES=0 swift infer \--model output/vx-xxx/checkpoint-xxx \--val_dataset latex_test3.jsonl \--max_new_tokens 256 \--temperature 0 \--attn_impl sdpa \--stream false

推送微调后的模型到ModelScope：

swift export \      --adapters output/vx-xxx/checkpoint-xxx \  --push_to_hub true \  --hub_model_id '<your-model-id>' \  --hub_token '<your-sdk-token>'

点击阅读原文，即可跳转模型链接~

👇点击关注ModelScope公众号获取
更多技术信息~