【peft】huggingface大模型加载多个LoRA并随时切换

import os
from transformers import AutoModel, LoRAModel
 
def load_lora_checkpoints(model_name_or_path, num_lora_parts):
    """
    加载多个LoRA分区模型。
    """
    lora_checkpoints = []
    for part in range(num_lora_parts):
        lora_path = f"{model_name_or_path}-lora-{part}"
        lora_model = LoRAModel.from_pretrained(lora_path)
        lora_checkpoints.append(lora_model)
    return lora_checkpoints
 
def switch_lora_part(lora_checkpoints, part_to_use):
    """
    切换到指定的LoRA分区。
    """
    # 确保分区索引在范围内
    assert 0 <= part_to_use < len(lora_checkpoints), "分区索引超出范围"
    # 将模型加载到CPU（如果当前在GPU上）并禁用梯度计算
    model = lora_checkpoints[part_to_use].cpu().eval()
    return model
 
# 示例使用
model_name_or_path = "distilbert-base-uncased"
num_lora_parts = 4
 
# 加载多个LoRA分区
lora_checkpoints = load_lora_checkpoints(model_name_or_path, num_lora_parts)
 
# 切换到第2个LoRA分区
part_to_use = 2
model = switch_lora_part(lora_checkpoints, part_to_use)
 
# 使用model进行推理或其他操作...

这段代码展示了如何加载多个LoRA分区模型并切换到指定的分区。首先定义了加载LoRA分区的函数，然后定义了切换到特定LoRA分区的函数。最后，给出了如何使用这些函数的示例。这个例子简洁明了，并且注重于教学和实用性。