3.prune.py

import torch
from transformers import Qwen2VLForConditionalGeneration, Qwen2_5_VLForConditionalGeneration, AutoProcessor
import torch_pruning as tp
from qwen_vl_utils import process_vision_info
from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VLPatchMerger
from torch import nn
from typing import Sequence
import os

def prune_model(model, processor, pruning_ratio):
    """同步剪枝LM和视觉模块，确保维度对齐"""

    num_heads = {}
    for name, module in model.named_modules():
        if name.endswith("self_attn"):
            num_heads[module.q_proj] = model.config.num_attention_heads
            num_heads[module.k_proj] = model.config.num_key_value_heads
            num_heads[module.v_proj] = model.config.num_key_value_heads

    importance = tp.importance.GroupNormImportance(p=2, group_reduction='mean') #tp.importance.ActivationImportance(p=2, target_types=[torch.nn.Linear])
    
    # 处理未封装的参数
    unwrapped_parameters = []

    # 忽略最后的lm_head和LM部分的embedding 
    ignored_layers = []
    for m in model.modules():
        if isinstance(m, torch.nn.Linear) and m.out_features == 151936:
            ignored_layers.append(m)
        if isinstance(m, torch.nn.Embedding):
            ignored_layers.append(m)
    print("ignored_layers", ignored_layers)

    # 构建输入
    # example_inputs = torch.randint(0, 100000, (3, 56, 56), dtype=torch.long, device='cuda:1')
    
    text = "描述这张图片。"
    example_inputs = torch.tensor(processor.tokenizer.encode(text)).unsqueeze(0).to(model.device)
    print(example_inputs.shape)

    # 创建剪枝器
    model.config.use_cache = False
    pruner = tp.pruner.MetaPruner(
        model,
        example_inputs=example_inputs,
        importance=importance,
        global_pruning=False,
        pruning_ratio=pruning_ratio,
        ignored_layers=ignored_layers,
        num_heads=num_heads,
        prune_num_heads=False,
        prune_head_dims=False,
        head_pruning_ratio=pruning_ratio,
        round_to=4,
        unwrapped_parameters=unwrapped_parameters,
    )
    
    # 执行剪枝
    for g in pruner.step(interactive=True):
        # print(g)
        g.prune()

    model.config.hidden_size = model.lm_head.in_features
    for name, m in model.model.named_modules():
        if name.endswith("self_attn"):
            print(name)
            m.hidden_size = m.q_proj.out_features
            m.num_heads = m.hidden_size // m.head_dim
            model.config.num_attention_heads = m.num_heads
            m.num_key_value_groups = m.num_heads // m.num_key_value_heads
        elif name.endswith("mlp"):
            if hasattr(m, "gate_proj"):
                print(name)
                m.hidden_size = m.gate_proj.in_features
                model.config.intermediate_size = m.gate_proj.out_features

    return model
            
def main():
    model_path = "/home/rzhong/project/unsloth/model_pretrain_sft_20250303_125849"
    # model_path = "/home/rzhong/project/FSTSPrune/Qwen2.5-VL-3B-Instruct-LatexOCR"
    model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
        model_path,
        torch_dtype=torch.bfloat16,
        device_map="cuda:1"
    )
    processor = AutoProcessor.from_pretrained(model_path)

    print("========= Before Pruning =========")
    print(model)
    ori_size = tp.utils.count_params(model)

    print("Starting pruning process...")
    pruned_model = prune_model(model, processor, pruning_ratio=0.5)
    print("========= After Pruning =========")
    print(pruned_model)

    print("  Params: %.2f M => %.2f M" %
          (ori_size / 1e6, tp.utils.count_params(pruned_model) / 1e6))
    
    # pruned_model.zero_grad()
    # save_path = "/home/rzhong/project/FSTSPrune/model_pretrain_sft_20250303_125849-Pruned"
    # os.makedirs(save_path, exist_ok=True)
    # # pruned_model.save_pretrained(save_path)
    # torch.save(pruned_model, os.path.join(save_path, "pytorch_model.bin"))
    # processor.save_pretrained(save_path)
    # pruned_model.config.save_pretrained(save_path)

    # pruned_model.zero_grad()
    # save_path = "/home/rzhong/project/FSTSPrune/model_pretrain_sft_20250303_125849-Pruned-hf"
    # os.makedirs(save_path, exist_ok=True)
    # pruned_model.save_pretrained(save_path)
    # # torch.save(pruned_model, os.path.join(save_path, "pytorch_model.bin"))
    # processor.save_pretrained(save_path)
    # # pruned_model.config.save_pretrained(save_path)

    # load_test_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    #     save_path,
    #     device_map="cpu"
    # )
    # print("load test pass!")

if __name__ == "__main__":
    main()