文件最后提交记录最后更新时间
[Docs] Modify current repository URLs to relative paths Co-authored-by: AZe_404<wangze62@h-partners.com> # message auto-generated for no-merge-commit merge: !2360 merge chg_branch_2600 into 26.0.0 [Docs] Modify current repository URLs to relative paths Created-by: AZe_404 Commit-by: AZe_404 Merged-by: ascend-robot Description: ## What this PR does / why we need it? 1. 拉取代码修改为拉取26.0.0分支,包括之前未指定版本的MindSpeed Core 2. 将MM仓库内的链接修改为相对路径访问 ## Does this PR introduce any user-facing change? Please describe whether the PR will result in any user-facing usage changes. If there is related documentation, please specify its path. ## How was this patch tested? Please explain how to verify the correctness and effectiveness of this feature, as well as its usage constraints and limitations. See merge request: Ascend/MindSpeed-MM!23601 个月前
!1012 【feature】 add Qwen2.5-Omni AudioEncoder Merge pull request !1012 from pjgao/master 11 个月前
!1012 【feature】 add Qwen2.5-Omni AudioEncoder Merge pull request !1012 from pjgao/master 11 个月前
!1012 【feature】 add Qwen2.5-Omni AudioEncoder Merge pull request !1012 from pjgao/master 11 个月前
!1012 【feature】 add Qwen2.5-Omni AudioEncoder Merge pull request !1012 from pjgao/master 11 个月前
!1012 【feature】 add Qwen2.5-Omni AudioEncoder Merge pull request !1012 from pjgao/master 11 个月前
!1012 【feature】 add Qwen2.5-Omni AudioEncoder Merge pull request !1012 from pjgao/master 11 个月前
!291 【特性】Qwen2VL-7B在线推理能力构建 Merge pull request !291 from 邱祥鑫/master 1 年前
!1401 [Bugfix] remove tp/pp size in qwen2/2.5vl 7b evaluation's model.json Merge pull request !1401 from wanghao/master 9 个月前
[Docs] Document corrections Co-authored-by: js1234567<jiangshuo9@h-partners.com> # message auto-generated for no-merge-commit merge: !2108 merge master into master [Docs] Document corrections Created-by: js1234567 Commit-by: js1234567 Merged-by: ascend-robot Description: ## Motivation Document corrections: 1. 添加2.3.0配套信息 2. 中英文标点问题 3. 链接版本更新 4. CANN8.5.0版本配置环境变量刷新, 涉及环境变量设置需全面排查修改 ## Modification Readme.md ## Self-test (Optional) If modifications to this PR may cause/fix function/accuracy/performance DTSs/issues, a self-inspection record needs to be attached. ## BC-breaking (Optional) If there are compatibility issues, such as dependencies on cann/torch_npu versions, they need to be explained in the PR. ## Checklist **Before PR**: - [x] The new code needs to comply with the Clean Code specification. - [x] The PR content is self-checked, and the expression can be clear and the writing standardized **After PR**: - [ ] CLA has been signed and all committers have signed the CLA in this PR. - [ ] The ci-pipeline is passed, Code Check is passed. See merge request: Ascend/MindSpeed-MM!21083 个月前
[Docs] Document corrections Co-authored-by: js1234567<jiangshuo9@h-partners.com> # message auto-generated for no-merge-commit merge: !2108 merge master into master [Docs] Document corrections Created-by: js1234567 Commit-by: js1234567 Merged-by: ascend-robot Description: ## Motivation Document corrections: 1. 添加2.3.0配套信息 2. 中英文标点问题 3. 链接版本更新 4. CANN8.5.0版本配置环境变量刷新, 涉及环境变量设置需全面排查修改 ## Modification Readme.md ## Self-test (Optional) If modifications to this PR may cause/fix function/accuracy/performance DTSs/issues, a self-inspection record needs to be attached. ## BC-breaking (Optional) If there are compatibility issues, such as dependencies on cann/torch_npu versions, they need to be explained in the PR. ## Checklist **Before PR**: - [x] The new code needs to comply with the Clean Code specification. - [x] The PR content is self-checked, and the expression can be clear and the writing standardized **After PR**: - [ ] CLA has been signed and all committers have signed the CLA in this PR. - [ ] The ci-pipeline is passed, Code Check is passed. See merge request: Ascend/MindSpeed-MM!21083 个月前
[Docs] Document corrections Co-authored-by: js1234567<jiangshuo9@h-partners.com> # message auto-generated for no-merge-commit merge: !2108 merge master into master [Docs] Document corrections Created-by: js1234567 Commit-by: js1234567 Merged-by: ascend-robot Description: ## Motivation Document corrections: 1. 添加2.3.0配套信息 2. 中英文标点问题 3. 链接版本更新 4. CANN8.5.0版本配置环境变量刷新, 涉及环境变量设置需全面排查修改 ## Modification Readme.md ## Self-test (Optional) If modifications to this PR may cause/fix function/accuracy/performance DTSs/issues, a self-inspection record needs to be attached. ## BC-breaking (Optional) If there are compatibility issues, such as dependencies on cann/torch_npu versions, they need to be explained in the PR. ## Checklist **Before PR**: - [x] The new code needs to comply with the Clean Code specification. - [x] The PR content is self-checked, and the expression can be clear and the writing standardized **After PR**: - [ ] CLA has been signed and all committers have signed the CLA in this PR. - [ ] The ci-pipeline is passed, Code Check is passed. See merge request: Ascend/MindSpeed-MM!21083 个月前
[Docs] Document corrections Co-authored-by: js1234567<jiangshuo9@h-partners.com> # message auto-generated for no-merge-commit merge: !2108 merge master into master [Docs] Document corrections Created-by: js1234567 Commit-by: js1234567 Merged-by: ascend-robot Description: ## Motivation Document corrections: 1. 添加2.3.0配套信息 2. 中英文标点问题 3. 链接版本更新 4. CANN8.5.0版本配置环境变量刷新, 涉及环境变量设置需全面排查修改 ## Modification Readme.md ## Self-test (Optional) If modifications to this PR may cause/fix function/accuracy/performance DTSs/issues, a self-inspection record needs to be attached. ## BC-breaking (Optional) If there are compatibility issues, such as dependencies on cann/torch_npu versions, they need to be explained in the PR. ## Checklist **Before PR**: - [x] The new code needs to comply with the Clean Code specification. - [x] The PR content is self-checked, and the expression can be clear and the writing standardized **After PR**: - [ ] CLA has been signed and all committers have signed the CLA in this PR. - [ ] The ci-pipeline is passed, Code Check is passed. See merge request: Ascend/MindSpeed-MM!21083 个月前
[Docs] Document corrections Co-authored-by: js1234567<jiangshuo9@h-partners.com> # message auto-generated for no-merge-commit merge: !2108 merge master into master [Docs] Document corrections Created-by: js1234567 Commit-by: js1234567 Merged-by: ascend-robot Description: ## Motivation Document corrections: 1. 添加2.3.0配套信息 2. 中英文标点问题 3. 链接版本更新 4. CANN8.5.0版本配置环境变量刷新, 涉及环境变量设置需全面排查修改 ## Modification Readme.md ## Self-test (Optional) If modifications to this PR may cause/fix function/accuracy/performance DTSs/issues, a self-inspection record needs to be attached. ## BC-breaking (Optional) If there are compatibility issues, such as dependencies on cann/torch_npu versions, they need to be explained in the PR. ## Checklist **Before PR**: - [x] The new code needs to comply with the Clean Code specification. - [x] The PR content is self-checked, and the expression can be clear and the writing standardized **After PR**: - [ ] CLA has been signed and all committers have signed the CLA in this PR. - [ ] The ci-pipeline is passed, Code Check is passed. See merge request: Ascend/MindSpeed-MM!21083 个月前
!1312 [Refactor] Add inference fa patch Merge pull request !1312 from 王泽/infer_fa_patch 10 个月前
[Docs] Document corrections Co-authored-by: js1234567<jiangshuo9@h-partners.com> # message auto-generated for no-merge-commit merge: !2108 merge master into master [Docs] Document corrections Created-by: js1234567 Commit-by: js1234567 Merged-by: ascend-robot Description: ## Motivation Document corrections: 1. 添加2.3.0配套信息 2. 中英文标点问题 3. 链接版本更新 4. CANN8.5.0版本配置环境变量刷新, 涉及环境变量设置需全面排查修改 ## Modification Readme.md ## Self-test (Optional) If modifications to this PR may cause/fix function/accuracy/performance DTSs/issues, a self-inspection record needs to be attached. ## BC-breaking (Optional) If there are compatibility issues, such as dependencies on cann/torch_npu versions, they need to be explained in the PR. ## Checklist **Before PR**: - [x] The new code needs to comply with the Clean Code specification. - [x] The PR content is self-checked, and the expression can be clear and the writing standardized **After PR**: - [ ] CLA has been signed and all committers have signed the CLA in this PR. - [ ] The ci-pipeline is passed, Code Check is passed. See merge request: Ascend/MindSpeed-MM!21083 个月前
!1312 [Refactor] Add inference fa patch Merge pull request !1312 from 王泽/infer_fa_patch 10 个月前
[Docs] Document corrections Co-authored-by: js1234567<jiangshuo9@h-partners.com> # message auto-generated for no-merge-commit merge: !2108 merge master into master [Docs] Document corrections Created-by: js1234567 Commit-by: js1234567 Merged-by: ascend-robot Description: ## Motivation Document corrections: 1. 添加2.3.0配套信息 2. 中英文标点问题 3. 链接版本更新 4. CANN8.5.0版本配置环境变量刷新, 涉及环境变量设置需全面排查修改 ## Modification Readme.md ## Self-test (Optional) If modifications to this PR may cause/fix function/accuracy/performance DTSs/issues, a self-inspection record needs to be attached. ## BC-breaking (Optional) If there are compatibility issues, such as dependencies on cann/torch_npu versions, they need to be explained in the PR. ## Checklist **Before PR**: - [x] The new code needs to comply with the Clean Code specification. - [x] The PR content is self-checked, and the expression can be clear and the writing standardized **After PR**: - [ ] CLA has been signed and all committers have signed the CLA in this PR. - [ ] The ci-pipeline is passed, Code Check is passed. See merge request: Ascend/MindSpeed-MM!21083 个月前
!1312 [Refactor] Add inference fa patch Merge pull request !1312 from 王泽/infer_fa_patch 10 个月前
[Docs] Document corrections Co-authored-by: js1234567<jiangshuo9@h-partners.com> # message auto-generated for no-merge-commit merge: !2108 merge master into master [Docs] Document corrections Created-by: js1234567 Commit-by: js1234567 Merged-by: ascend-robot Description: ## Motivation Document corrections: 1. 添加2.3.0配套信息 2. 中英文标点问题 3. 链接版本更新 4. CANN8.5.0版本配置环境变量刷新, 涉及环境变量设置需全面排查修改 ## Modification Readme.md ## Self-test (Optional) If modifications to this PR may cause/fix function/accuracy/performance DTSs/issues, a self-inspection record needs to be attached. ## BC-breaking (Optional) If there are compatibility issues, such as dependencies on cann/torch_npu versions, they need to be explained in the PR. ## Checklist **Before PR**: - [x] The new code needs to comply with the Clean Code specification. - [x] The PR content is self-checked, and the expression can be clear and the writing standardized **After PR**: - [ ] CLA has been signed and all committers have signed the CLA in this PR. - [ ] The ci-pipeline is passed, Code Check is passed. See merge request: Ascend/MindSpeed-MM!21083 个月前
!648 【bugfix】qwen2vl数据转换bug修复 Merge pull request !648 from pjgao/master 1 年前
!764 [bugfix]Lora权重合并加速&cleancode代码规范整改 Merge pull request !764 from 陆劲夫/master 1 年前
!1049 [Refactor]adapt mindspeed012 Merge pull request !1049 from 王泽/adapt012 9 个月前
!1049 [Refactor]adapt mindspeed012 Merge pull request !1049 from 王泽/adapt012 9 个月前
!1049 [Refactor]adapt mindspeed012 Merge pull request !1049 from 王泽/adapt012 9 个月前
!751 【特性】为Qwen2VL 72B添加DPO算法支持 Merge pull request !751 from zhangxubin/detached 1 年前
README.md

Qwen2_VL 使用指南

目录

版本说明

参考实现

url=https://github.com/hiyouga/LLaMA-Factory.git
commit_id=52f2565

变更记录

2024.10.21: 首次支持Qwen2-VL模型 2025.03.26: 同步开源仓数据template修改 2025.05.29:同步开源仓数据处理修改


环境安装

1. 环境准备

【模型开发时推荐使用配套的环境版本】

请参考安装指南,完成昇腾软件安装。

2. 环境搭建

git clone --branch 26.0.0 https://gitcode.com/Ascend/MindSpeed-MM.git
git clone https://github.com/NVIDIA/Megatron-LM.git
cd Megatron-LM
git checkout core_v0.12.1
cp -r megatron ../MindSpeed-MM/
cd ..
cd MindSpeed-MM
mkdir logs data ckpt
# 安装加速库
git clone https://gitcode.com/Ascend/MindSpeed.git
cd MindSpeed
# checkout commit from MindSpeed core_r0.12.1
git checkout 5176c6f5f133111e55a404d82bd2dc14a809a6ab
# 安装mindspeed及依赖
pip install -e .
cd ..
# 安装mindspeed mm及依赖
pip install -e .

权重下载及转换

1. 权重下载

从Hugging Face库下载对应的模型权重:

将下载的模型权重保存到本地的ckpt/hf_path/Qwen2-VL-*B-Instruct目录下。(*表示对应的尺寸)

2. 权重转换(hf2mm)

MindSpeed MM修改了部分原始网络的结构名称,使用mm-convert工具对原始预训练权重进行转换。该工具实现了huggingface权重和MindSpeed MM权重的互相转换以及PP(Pipeline Parallel)权重的重切分。参考权重转换工具

# 2b
mm-convert  Qwen2VLConverter hf_to_mm \
  --cfg.mm_dir "ckpt/mm_path/Qwen2-VL-2B-Instruct" \
  --cfg.hf_config.hf_dir "ckpt/hf_path/Qwen2-VL-2B-Instruct" \
  --cfg.parallel_config.llm_pp_layers [[28]] \
  --cfg.parallel_config.vit_pp_layers [[32]] \
  --cfg.parallel_config.tp_size 1

# 7b
mm-convert  Qwen2VLConverter hf_to_mm \
  --cfg.mm_dir "ckpt/mm_path/Qwen2-VL-7B-Instruct" \
  --cfg.hf_config.hf_dir "ckpt/hf_path/Qwen2-VL-7B-Instruct" \
  --cfg.parallel_config.llm_pp_layers [[1,10,10,7]] \
  --cfg.parallel_config.vit_pp_layers [[32,0,0,0]] \
  --cfg.parallel_config.tp_size 1

# 7b vpp
mm-convert  Qwen2VLConverter hf_to_mm \
  --cfg.mm_dir "ckpt/mm_path/Qwen2-VL-7B-Instruct-vpp" \
  --cfg.hf_config.hf_dir "ckpt/hf_path/Qwen2-VL-7B-Instruct" \
  --cfg.parallel_config.llm_pp_layers [[0,0,0,1],[4,4,4,4],[4,3,2,2]] \
  --cfg.parallel_config.vit_pp_layers [[10,10,10,2],[0,0,0,0],[0,0,0,0]] \
  --cfg.parallel_config.tp_size 1

# 72b
mm-convert  Qwen2VLConverter hf_to_mm \
  --cfg.mm_dir "ckpt/mm_path/Qwen2-VL-72B-Instruct" \
  --cfg.hf_config.hf_dir "ckpt/hf_path/Qwen2-VL-72B-Instruct" \
  --cfg.parallel_config.llm_pp_layers [[5,11,11,11,11,11,11,9]] \
  --cfg.parallel_config.vit_pp_layers [[32,0,0,0,0,0,0,0]] \
  --cfg.parallel_config.tp_size 2
# 其中:
# mm_dir: 转换后保存目录
# hf_dir: huggingface权重目录
# llm_pp_layers: llm在每个卡上切分的层数,注意要和model.json中配置的pipeline_num_layers一致
# vit_pp_layers: vit在每个卡上切分的层数,注意要和model.json中配置的pipeline_num_layers一致
# tp_size: tp并行数量,注意要和微调启动脚本中的配置一致

如果需要用转换后模型训练的话,同步修改examples/qwen2vl/finetune_qwen2vl_7b.sh中的LOAD_PATH参数,该路径为转换后或者切分后的权重,注意与原始权重 ckpt/hf_path/Qwen2-VL-7B-Instruct进行区分。

LOAD_PATH="ckpt/mm_path/Qwen2-VL-7B-Instruct"

3. 训练后权重转回huggingface格式

MindSpeed MM修改了部分原始网络的结构名称,在微调后,如果需要将权重转回huggingface格式,可使用mm-convert权重转换工具对微调后的权重进行转换,将权重名称修改为与原始网络一致。

mm-convert  Qwen2VLConverter mm_to_hf \
  --cfg.save_hf_dir "ckpt/mm_to_hf/Qwen2-VL-7B-Instruct" \
  --cfg.mm_dir "ckpt/mm_path/Qwen2-VL-7B-Instruct" \
  --cfg.hf_config.hf_dir "ckpt/hf_path/Qwen2-VL-7B-Instruct" \
  --cfg.parallel_config.llm_pp_layers [1,10,10,7] \
  --cfg.parallel_config.vit_pp_layers [32,0,0,0] \
  --cfg.parallel_config.tp_size 1
# 其中:
# save_hf_dir: mm微调后转换回hf模型格式的目录
# mm_dir: 微调后保存的权重目录
# hf_dir: huggingface权重目录
# llm_pp_layers: llm在每个卡上切分的层数,注意要和微调时model.json中配置的pipeline_num_layers一致
# vit_pp_layers: vit在每个卡上切分的层数,注意要和微调时model.json中配置的pipeline_num_layers一致
# tp_size: tp并行数量,注意要和微调启动脚本中的配置一致

4. 训练后重新切分权重

权重下载及转换部分会把权重进行pp切分和tp切分,在微调后,如果需要对权重重新进行切分,可使用mm-convert权重转换工具对微调后的权重进行切分。

mm-convert  Qwen2VLConverter resplit \
  --cfg.source_dir "ckpt/mm_path/Qwen2-VL-7B-Instruct" \
  --cfg.target_dir "ckpt/mm_resplit_pp/Qwen2-VL-7B-Instruct" \
  --cfg.source_parallel_config.llm_pp_layers [1,10,10,7] \
  --cfg.source_parallel_config.vit_pp_layers [32,0,0,0] \
  --cfg.source_parallel_config.tp_size 1 \
  --cfg.target_parallel_config.llm_pp_layers [4,24] \
  --cfg.target_parallel_config.vit_pp_layers [32,0] \
  --cfg.target_parallel_config.tp_size 1
# 其中
# source_dir: 微调后保存的权重目录
# target_dir: 希望重新pp切分后保存的目录
# source_parallel_config.llm_pp_layers: 微调时llm的pp配置
# source_parallel_config.vit_pp_layers: 微调时vit的pp配置
# source_parallel_config.tp_size: 微调时tp并行配置
# target_parallel_config.llm_pp_layers: 期望的重切分llm模块切分层数
# target_parallel_config.vit_pp_layers: 期望的重切分vit模块切分层数
# target_parallel_config.tp_size: 期望的tp并行配置(tp_size不能超过原仓config.json中的num_key_value_heads)

数据集准备及处理

1. 数据集下载(以COCO2017数据集为例)

(1)用户需要自行下载COCO2017数据集COCO2017,并解压到项目目录下的./data/COCO2017文件夹中

(2)获取图片数据集的描述文件(LLaVA-Instruct-150K),下载至./data/路径下;

(3)运行数据转换脚本python examples/qwen2vl/llava_instruct_2_mllm_demo_format.py;

$playground
├── data
    ├── COCO2017
        ├── train2017

    ├── llava_instruct_150k.json
    ├── mllm_format_llava_instruct_data.json
    ...

当前支持读取多个以,(注意不要加空格)分隔的数据集,配置方式为data.json中 dataset_param->basic_parameters->dataset 从"./data/mllm_format_llava_instruct_data.json"修改为"./data/mllm_format_llava_instruct_data.json,./data/mllm_format_llava_instruct_data2.json"

同时注意data.jsondataset_param->basic_parameters->max_samples的配置,会限制数据只读max_samples条,这样可以快速验证功能。如果正式训练时,可以把该参数去掉则读取全部的数据。

2.纯文本或有图无图混合训练数据(以LLaVA-Instruct-150K为例)

现在本框架已经支持纯文本/混合数据(有图像和无图像数据混合训练)。

在数据构造时,对于包含图片的数据,需要保留image这个键值。

{
  "id": your_id,
  "image": your_image_path,
  "conversations": [
      {"from": "human", "value": your_query},
      {"from": "gpt", "value": your_response},
  ],
}

在数据构造时,对于纯文本数据,可以去除image这个键值。

{
  "id": your_id,
  "conversations": [
      {"from": "human", "value": your_query},
      {"from": "gpt", "value": your_response},
  ],
}

微调

1. 准备工作

配置脚本前需要完成前置准备工作,包括:环境安装权重下载及转换数据集准备及处理,详情可查看对应章节。

2. 配置参数

【数据目录配置】

根据实际情况修改data.json中的数据集路径,包括model_name_or_pathdataset_dirdataset等字段。

实例:如果数据及其对应的json都在/home/user/data/目录下,其中json目录为/home/user/data/video_data_path.json,此时配置如下: dataset_dir配置为/home/user/data/; dataset配置为./data/video_data_path.json 注意此时dataset需要配置为相对路径

以Qwen2VL-7B为例,data.json进行以下修改,注意model_name_or_path的权重路径为转换前的权重路径。

注意cache_dir在多机上不要配置同一个挂载目录避免写入同一个文件导致冲突

{
    "dataset_param": {
        "dataset_type": "huggingface",
        "preprocess_parameters": {
            "model_name_or_path": "./ckpt/hf_path/Qwen2-VL-7B-Instruct",
            ...
        },
        "basic_parameters": {
            ...
            "dataset_dir": "./data",
            "dataset": "./data/mllm_format_llava_instruct_data.json",
            "cache_dir": "./data/cache_dir",
            ...
        },
        ...
    },
    ...
}

如果需要加载大批量数据,可使用流式加载,修改data.json中的sampler_type字段,增加streaming字段。(注意:使用流式加载后当前仅支持num_workers=0,单进程处理数据,会有性能波动,并且不支持断点续训功能。)

{
    "dataset_param": {
        ...
        "basic_parameters": {
            ...
            "streaming": true
            ...
        },
        ...
    },
    "dataloader_param": {
        ...
        "sampler_type": "stateful_distributed_sampler",
        ...
    }
}

如果需要计算validation loss,需要在shell脚本中修改eval-interval参数和eval-iters参数;需要在data.json中的basic_parameters内增加字段: 对于非流式数据有两种方式:①根据实际情况增加val_dataset验证集路径,②增加val_rate字段对训练集进行切分; 对于流式数据,仅支持增加val_dataset字段进行计算。

{
    "dataset_param": {
        ...
        "basic_parameters": {
            ...
            "val_dataset": "./data/val_dataset.json",
            "val_max_samples": null,
            "val_rate": 0.1,
            ...
        },
        ...
    },
   ...
    }
}

【模型保存加载及日志信息配置】

根据实际情况配置examples/qwen2vl/finetune_qwen2vl_7b.sh的参数,包括加载、保存路径以及保存间隔--save-interval(注意:分布式优化器保存文件较大耗时较长,请谨慎设置保存间隔)

...
# 加载路径
LOAD_PATH="ckpt/mm_path/Qwen2-VL-7B-Instruct"
# 保存路径
SAVE_PATH="save_dir"
...
GPT_ARGS="
    ...
    --no-load-optim \  # 不加载优化器状态,若需加载请移除
    --no-load-rng \  # 不加载随机数状态,若需加载请移除
    --no-save-optim \  # 不保存优化器状态,若需保存请移除
    --no-save-rng \  # 不保存随机数状态,若需保存请移除
    ...
"
...
OUTPUT_ARGS="
    --log-interval 1 \  # 日志间隔
    --save-interval 5000 \  # 保存间隔
    ...
    --log-tps \  # 增加此参数可使能在训练中打印每步语言模块的平均序列长度,并在训练结束后计算每秒吞吐tokens量。
"

若需要加载指定迭代次数的权重、优化器等状态,需将加载路径LOAD_PATH设置为保存文件夹路径LOAD_PATH="save_dir",并修改latest_checkpointed_iteration.txt文件内容为指定迭代次数 (此功能coming soon)

$save_dir
   ├── latest_checkpointed_iteration.txt
   ├── ...

【单机运行配置】

配置examples/qwen2vl/finetune_qwen2vl_7b.sh参数如下

# 根据实际情况修改 ascend-toolkit 路径
source /usr/local/Ascend/cann/set_env.sh
NPUS_PER_NODE=8
MASTER_ADDR=localhost
MASTER_PORT=29501
NNODES=1
NODE_RANK=0
WORLD_SIZE=$(($NPUS_PER_NODE * $NNODES))

注意,当开启PP时,model.json中配置的vision_encodertext_decoderpipeline_num_layer参数控制了各自的PP切分策略。对于流水线并行,要先处理vision_encoder再处理text_decoder。 比如7b默认的值[32,0,0,0][1,10,10,7],其含义为PP域内第一张卡先放32层vision_encoder再放1层text_decoder、第二张卡放text_decoder接着的10层、第三张卡放text_decoder接着的10层、第四张卡放text_decoder接着的7层,vision_encoder没有放完时不能先放text_decoder(比如[30,2,0,0][1,10,10,7]的配置是错的)

同时注意,如果某张卡上的参数全部冻结时会导致没有梯度(比如vision_encoder冻结时PP配置[30,2,0,0][0,11,10,7]),需要在finetune_qwen2vl_7b.shGPT_ARGS参数中增加--enable-dummy-optimizer,参考dummy_optimizer特性文档

3. 启动微调

以Qwen2VL-7B为例,启动微调训练任务。
loss计算方式差异会对训练效果造成不同的影响,在启动训练任务之前,请查看关于loss计算的文档,选择合适的loss计算方式vlm_model_loss_calculate_type.md

bash examples/qwen2vl/finetune_qwen2vl_7b.sh

推理

1、准备工作(以微调环境为基础,包括环境安装、权重下载及转换-目前支持PP切分的推理)

追加安装:

pip install qwen_vl_utils

注:如果使用huggingface下载的原始权重,需要权重转换,权重转换步骤中,根据具体需求设置PP切分的参数。

注:如果使用的MindSpeed-MM中保存的权重则无需进行转换,可直接加载(需要保证与训练的切分一致)。

2、配置参数

根据实际情况修改examples/qwen2vl/inference_qwen2vl_7b.json和examples/qwen2vl/inference_qwen2vl_7b.sh中的路径配置,包括tokenizer的加载路径from_pretrained、以及图片处理器的路径image_processer_path。需注意

(1)tokenizer/from_pretrained配置的路径为从huggingface下载的原始Qwen2-VL-7B-Instruct路径。

(2)shell文件中的LOAD_PATH的路径为经过权重转换后的模型路径(可PP切分)。

3、启动推理

bash examples/qwen2vl/inference_qwen2vl_7b.sh

注:单卡推理需打开FA,否则可能会显存不足报错,开关--use-flash-attn 默认已开,确保FA步骤完成即可。如果使用多卡推理则需要调整相应的PP参数和NPU使用数量的NPUS_PER_NODE参数。以PP4为例,shell修改参数如下:

NPUS_PER_NODE=4 # 可用几张卡 要大于 PP*TP*CP
PP=4 #PP并行参数

Qwen2vl支持视频理解

1、加载视频数据集

数据集中的视频数据集取自llamafactory,https://github.com/hiyouga/LLaMA-Factory/tree/main/data

视频取自mllm_demo_data,使用时需要将该数据放到自己的data文件夹中去,同时将llamafactory上的mllm_video_demo.json也放到自己的data文件中

以data_72b.json为例加载数据集:参照data_72b_video.json

2、修改模型配置

以72b为例,需要修改model_72b.json:

"img_context_token_id": 151656

3、Qwen2vl支持视频推理

配置修改 以7b模型推理为例,修改inference_qwen2vl_7b.json

"img_context_token_id": 151656

修改prompts内容中添加对视频的描述

"prompts": "Describe this video and keep it within 100 words."

支持视频的推理将image_path修改为video_path,原来加载的图片的路径改为视频路径 视频数据样例: https://github.com/hiyouga/LLaMA-Factory/blob/main/data/mllm_demo_data/1.mp4

暂不支持image_path与video_path同时存在,不支持img和video混合推理


Qwen2VL支持DPO算法

当前仅支持72B Lora场景。

环境安装、权重下载、权重转换同微调章节。

1.数据集准备以及处理(以RLHF-V为例)

  • 下载数据集:RLHF-V

  • 处理数据集:在examples/qwen2vl/rlhfv_2_sharegpt_demo_format.py文件中,修改下方所述的三个路径、然后运行脚本。

    # 将其设置为图片保存的路径
    IMAGE_FOLDER = Path("./data/rlhf_v_images/res")
    # 将其设置为处理好的json路径
    OUTPUT_JSON_PATH = "./data/rlhf-v.json"
    # 将其设置为从huggingface下载的数据集路径
    DATASET_NAME = "./data/datasets/rlhf-v"
    

2.配置参数

  • data_72b_dpo.json

    参数含义同微调章节。

    根据实际情况修改data.json中的数据集路径,包括model_name_or_pathdataset_dirdataset等字段。

    例如:将下载好的权重放在./ckpt/hf_path/Qwen2-VL-72B-Instruct, 处理好的数据集放在./data/rlhf-v.json

    则data_72b_dpo.json里的参数设置如下:

        ......
     "dataset_param": {
            "dataset_type": "huggingface",
            "preprocess_parameters": {
                "model_name_or_path": "./ckpt/hf_path/Qwen2-VL-72B-Instruct",
                ......
            },
            "basic_parameters": {
                "template": "qwen2vl",
                "dataset_dir": "./data",
                "dataset": "./data/rlhf-v.json",
                ......
            },
          ......
    ......
    
  • model_72b.json

    参数含义同微调章节。

    以单机8卡为例,需要将model_72b.json里面的vision_encodertext_decoderpipeline_num_layers参数调整为:

    {
    ...
        "image_encoder": {
            "vision_encoder": {
                "model_id": "qwen2vit",
                "num_layers": 32,
    
                ...
    
                "pipeline_num_layers": [32, 0, 0, 0],
    
                ...
            },
     ...
        },
        "text_decoder": {
            "model_id": "qwen2lm",
            "kv_channels": 128,
            "num_layers": 80,
            "pipeline_num_layers": [17, 21, 22, 20],
            ...
    }
    ...
    
  • finetune_qwen2vl_72b_dpo.sh

    参数含义、配置项同微调章节。

    下面介绍DPO的参数含义:

    参数 含义
    dpo-beta 正则化参数,平衡奖励得分与KL散度,默认0.1
    dpo-loss-type 指定loss计算方法,目前支持:sigmoid(dpo原始方案),其他方法例如hinge、ipo因为未验证,所以不支持
    dpo-label-smoothing 考虑样本噪声,计算loss时的平滑参数,取值范围0到0.5,默认0.0
    pref-ftx dpo loss中加入sft loss时用的乘数,默认0.0
    ref-model 参考模型的权重路径。当前不支持断点续训。

3.启动DPO任务

bash examples/qwen2vl/finetune_qwen2vl_72b_dpo.sh

评测

数据集准备

当前模型支持AI2D(test)、ChartQA(test)、Docvqa(val)、MMMU(val)四种数据集的评测。 数据集参考下载链接:

参数配置

如果要进行评测需要将要评测的数据集名称和路径传到examples/qwen2vl/evaluate_qwen2vl_7b.json 需要更改的字段有

  • tokenizer中的from_pretrained为huggingface的Qwen2-VL的权重,参考readme上面链接自行下载传入
  • dataset_path为上述评测数据集的本地路径
  • evaluation_dataset为评测数据集的名称可选的名称有(ai2d_testmmmu_dev_valdocvqa_valchartqa_test), 注意:需要与上面的数据集路径相对应。
  • result_output_path为评测结果的输出路径,注意:每次评测前需要将之前保存在该路径下评测文件删除。
    "tokenizer": {
        "from_pretrained": "./Qwen2-VL-7B-Instruct",

    },
    "dataset_path": "./AI2D_TEST.tsv",
    "evaluation_dataset":"ai2d_test",
    "evaluation_model":"qwen2_vl_7b",
    "result_output_path":"./evaluation_outputs/"

examples/qwen2vl/evaluate_qwen2vl_7b.json改完后,需要将json文件的路径传入到examples/qwen2vl/evaluate_qwen2vl_7b.sh MM_MODEL字段中。

以及需要将上面提到的权重转换后模型传入examples/qwen2vl/evaluate_qwen2vl_7b.sh中的LOAD_PATH字段中。

MM_MODEL=examples/qwen2vl/evaluate_qwen2vl_7b.json
LOAD_PATH="./qwen_7b_pp1/Qwen2-VL-7B-Instruct"

评测支持多卡DP评测需要更改的配置,为NPU卡数量

NPUS_PER_NODE=1

启动评测

评测额外依赖一些python包,使用下面命令进行安装

pip install -e ".[evaluate]"

启动shell开始评测

bash examples/qwen2vl/evaluate_qwen2vl_7b.sh

评测结果会输出到result_output_path路径中,会输出结果文件:

  • *.xlsx文件,这个文件会输出每道题的预测结果和答案等详细信息。
  • *.csv文件,这个文件会输出统计准确率等数据。

特性使用介绍

lora微调

LoRA为框架通用能力,当前功能已支持,可参考LoRA特性文档

非均匀CP切分

非均匀CP的介绍和使能方式,可参考unaligned_ulysses_cp

非均匀SP切分

非均匀SP的介绍和使能方式,可参考unaligned_sequence_parallel

环境变量声明

环境变量 描述 取值说明
ASCEND_SLOG_PRINT_TO_STDOUT 是否开启日志打印 0: 关闭日志打屏
1: 开启日志打屏
ASCEND_GLOBAL_LOG_LEVEL 设置应用类日志的日志级别及各模块日志级别,仅支持调试日志 0: 对应DEBUG级别
1: 对应INFO级别
2: 对应WARNING级别
3: 对应ERROR级别
4: 对应NULL级别,不输出日志
TASK_QUEUE_ENABLE 用于控制开启task_queue算子下发队列优化的等级 0: 关闭
1: 开启Level 1优化
2: 开启Level 2优化
COMBINED_ENABLE 设置combined标志。设置为0表示关闭此功能;设置为1表示开启,用于优化非连续两个算子组合类场景 0: 关闭
1: 开启
CPU_AFFINITY_CONF 控制CPU端算子任务的处理器亲和性,即设定任务绑核 设置0或未设置: 表示不启用绑核功能
1: 表示开启粗粒度绑核
2: 表示开启细粒度绑核
HCCL_CONNECT_TIMEOUT 用于限制不同设备之间socket建链过程的超时等待时间 需要配置为整数,取值范围[120,7200],默认值为120,单位s
PYTORCH_NPU_ALLOC_CONF 控制缓存分配器行为 expandable_segments:<value>: 使能内存池扩展段功能,即虚拟内存特征
HCCL_EXEC_TIMEOUT 控制设备间执行时同步等待的时间,在该配置时间内各设备进程等待其他设备执行通信同步 需要配置为整数,取值范围[68,17340],默认值为1800,单位s
ACLNN_CACHE_LIMIT 配置单算子执行API在Host侧缓存的算子信息条目个数 需要配置为整数,取值范围[1, 10,000,000],默认值为10000
TOKENIZERS_PARALLELISM 用于控制Hugging Face的transformers库中的分词器(tokenizer)在多线程环境下的行为 False: 禁用并行分词
True: 开启并行分词
MULTI_STREAM_MEMORY_REUSE 配置多流内存复用是否开启 0: 关闭多流内存复用
1: 开启多流内存复用
NPU_ASD_ENABLE 控制是否开启Ascend Extension for PyTorch的特征值检测功能 设置0或未设置: 关闭特征值检测
1: 表示开启特征值检测,只打印异常日志,不告警
2:开启特征值检测,并告警
3:开启特征值检测,并告警,同时会在device侧info级别日志中记录过程数据
ASCEND_LAUNCH_BLOCKING 控制算子执行时是否启动同步模式 0: 采用异步方式执行
1: 强制算子采用同步模式运行
NPUS_PER_NODE 配置一个计算节点上使用的NPU数量 整数值(如 1, 8 等)

注意事项

  1. finetune_xx.sh里,与模型结构相关的参数并不生效,以examples/qwen2vl/model_xb.json里同名参数配置为准,非模型结构的训练相关参数在 finetune_xx.sh修改。