文件最后提交记录最后更新时间
[feature] CANN/HDK路径适配:添加CANN路径提醒注释及批量替换工具 Co-authored-by: yeqm<yeqiangmao@huawei.com> Co-authored-by: liuxi_<liuxi75@huawei.com> # message auto-generated for no-merge-commit merge: !2440 merge supp_path_oversea into master [feature] CANN/HDK路径适配:添加CANN路径提醒注释及批量替换工具 Created-by: liuxi_ Commit-by: liuxi_;yeqm Merged-by: ascend-robot Description: # What this PR does / why we need it? 考虑出海版本(CANN 安装路径为 /usr/local/npu),修改文档中涉及/usr/local/Ascend路径时的提示。 并提供路径替换脚本,通过脚本完成仓库适配后,组件可正常安装、模型微调可正常运行。 ## Does this PR introduce any user-facing change? 在部分文档中会优化对CANN环境脚本路径的提示,提示按实际路径配置脚本。 ## How was this patch tested? 执行脚本后,验证基本功能正常 See merge request: Ascend/MindSpeed-MM!244027 天前
style: pre-commit autofix cleancode (base check) Co-authored-by: liyingxuan<liyingxuan3@huawei.com> # message auto-generated for no-merge-commit merge: !2616 merge master into master style: pre-commit autofix cleancode (base check) Created-by: liyx616 Commit-by: liyingxuan Merged-by: ascend-robot Description: ## What this PR does / why we need it? Please describe the background and detailed changes of the PR. If it is a bugfix, please attach the related issue. ## 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!261613 小时前
style: pre-commit autofix cleancode (base check) Co-authored-by: liyingxuan<liyingxuan3@huawei.com> # message auto-generated for no-merge-commit merge: !2616 merge master into master style: pre-commit autofix cleancode (base check) Created-by: liyx616 Commit-by: liyingxuan Merged-by: ascend-robot Description: ## What this PR does / why we need it? Please describe the background and detailed changes of the PR. If it is a bugfix, please attach the related issue. ## 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!261613 小时前
style: pre-commit autofix cleancode (base check) Co-authored-by: liyingxuan<liyingxuan3@huawei.com> # message auto-generated for no-merge-commit merge: !2616 merge master into master style: pre-commit autofix cleancode (base check) Created-by: liyx616 Commit-by: liyingxuan Merged-by: ascend-robot Description: ## What this PR does / why we need it? Please describe the background and detailed changes of the PR. If it is a bugfix, please attach the related issue. ## 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!261613 小时前
style: pre-commit autofix cleancode (base check) Co-authored-by: liyingxuan<liyingxuan3@huawei.com> # message auto-generated for no-merge-commit merge: !2616 merge master into master style: pre-commit autofix cleancode (base check) Created-by: liyx616 Commit-by: liyingxuan Merged-by: ascend-robot Description: ## What this PR does / why we need it? Please describe the background and detailed changes of the PR. If it is a bugfix, please attach the related issue. ## 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!261613 小时前
style: pre-commit autofix cleancode (base check) Co-authored-by: liyingxuan<liyingxuan3@huawei.com> # message auto-generated for no-merge-commit merge: !2616 merge master into master style: pre-commit autofix cleancode (base check) Created-by: liyx616 Commit-by: liyingxuan Merged-by: ascend-robot Description: ## What this PR does / why we need it? Please describe the background and detailed changes of the PR. If it is a bugfix, please attach the related issue. ## 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!261613 小时前
style: pre-commit autofix cleancode (base check) Co-authored-by: liyingxuan<liyingxuan3@huawei.com> # message auto-generated for no-merge-commit merge: !2616 merge master into master style: pre-commit autofix cleancode (base check) Created-by: liyx616 Commit-by: liyingxuan Merged-by: ascend-robot Description: ## What this PR does / why we need it? Please describe the background and detailed changes of the PR. If it is a bugfix, please attach the related issue. ## 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!261613 小时前
README.md

Qwen3_VL 使用指南

目录


环境安装

MindSpeed-MM MindSpore后端的依赖配套如下表,安装步骤参考基础安装指导

依赖软件
昇腾NPU驱动固件 在研版本
昇腾 CANN 在研版本
MindSpore 2.7.2
Python >=3.9
transformers v4.57.0

1. 仓库拉取及环境搭建

针对MindSpeed MindSpore后端,昇腾社区提供了模型一键拉起部署MindSpeed-Core-MS,旨在帮助用户自动拉取相关代码仓并对torch代码进行一键适配,进而使用户无需再额外手动开发适配即可在华为MindSpore+CANN环境下一键拉起模型训练。在进行一键拉起前,用户需要拉取相关的代码仓以及进行环境搭建:

# 创建conda环境
conda create -n test python=3.10
conda activate test

# 使用环境变量
# 根据实际情况修改 ascend-toolkit 路径
source /usr/local/Ascend/cann/set_env.sh
# 根据实际情况修改 ascend-toolkit 路径
source /usr/local/Ascend/nnal/atb/set_env.sh --cxx_abi=0

# 安装MindSpeed-Core-MS一键拉起部署
git clone https://gitcode.com/Ascend/MindSpeed-Core-MS.git -b r0.5.0

# 使用MindSpeed-Core-MS内部脚本自动拉取相关代码仓并一键适配
cd MindSpeed-Core-MS
pip install -r requirements.txt 
source auto_convert.sh mm
# 使用master分支的MindSpeed-MM
cd MindSpeed-MM
git switch master
cd ..

# 安装新版transformers(支持qwen3vl模型)
git clone https://github.com/huggingface/transformers.git
cd transformers
git checkout c0dbe09
pip install -e .

mkdir ckpt
mkdir data
mkdir logs

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/Qwen3-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。参考权重转换工具了解该工具的具体使用。注意当前在MindSpore后端下,转换出的权重无法用于Torch后端的训练

注:基于mindspore后端执行权重转换时,mm-convert执行的脚本为convert_cli.py

  
# 8b
mm-convert  Qwen3VLMegatronConverter hf_to_mm \
  --cfg.mm_dir "ckpt/mm_path/Qwen3-VL-8B" \
  --cfg.hf_config.hf_dir "ckpt/hf_path/Qwen3-VL-8B-Instruct" \
  --cfg.parallel_config.llm_pp_layers [[6,10,10,10]] \
  --cfg.parallel_config.vit_pp_layers [[27,0,0,0]] \
  --cfg.parallel_config.tp_size 1

# 30b
mm-convert  Qwen3VLMegatronConverter hf_to_mm \
--cfg.mm_dir "ckpt/mm_path/Qwen3-VL-30B" \
--cfg.hf_config.hf_dir "ckpt/hf_path/Qwen3-VL-30B-Instruct" \
--cfg.parallel_config.llm_pp_layers [[6,14,14,14]] \
--cfg.parallel_config.vit_pp_layers [[27,0,0,0]] \
--cfg.parallel_config.tp_size 1 \
--cfg.parallel_config.ep_size 1

# 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
# ep_size: 专家并行数量,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/mindspore/qwen3vl/finetune_qwen3vl_*B.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/Qwen3-VL-xxB-Instruct进行区分。

LOAD_PATH="ckpt/mm_path/Qwen3-VL-xxB"

3. 权重转换(mm2hf)

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。

# 8b
mm-convert  Qwen3VLMegatronConverter mm_to_hf \
  --cfg.save_hf_dir "ckpt/mm_to_hf/Qwen3-VL-8B" \
  --cfg.mm_dir "ckpt/mm_path/Qwen3-VL-8B" \
  --cfg.hf_config.hf_dir "ckpt/hf_path/Qwen3-VL-8B-Instruct" \
  --cfg.parallel_config.llm_pp_layers [6,10,10,10] \
  --cfg.parallel_config.vit_pp_layers [27,0,0,0] \
  --cfg.parallel_config.tp_size 1

# 30b
mm-convert  Qwen3VLMegatronConverter mm_to_hf \
--cfg.save_hf_dir "ckpt/mm_to_hf/Qwen3-VL-30B" \
--cfg.mm_dir "ckpt/mm_path/Qwen3-VL-30B" \
--cfg.hf_config.hf_dir "ckpt/hf_path/Qwen3-VL-30B-Instruct" \
--cfg.parallel_config.llm_pp_layers [6,14,14,14] \
--cfg.parallel_config.vit_pp_layers [27,0,0,0] \
--cfg.parallel_config.tp_size 1 \
--cfg.parallel_config.ep_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
# ep_size: 专家并行数量,注意要和微调启动脚本中的配置一致

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,转换后参考数据目录结构如下:

$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_xxB.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_xxB.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_xxB.json中的数据集路径,包括model_name_or_pathdataset_dirdataset等字段。

示例: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Qwen3VL-xxB为例,data_xxB.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,即原始hf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/Qwen3-VL-xxB-Instruct",
            ...
        },
        "basic_parameters": {
            ...
            "dataset_dir": "./data",
            "dataset": "./data/mllm_format_llava_instruct_data.json",
            "cache_dir": "./data/cache_dir",
            ...
        },
        ...
    },
    ...
}

【模型保存加载及日志信息配置】

根据实际情况配置examples/mindspore/qwen3vl/finetune_qwen3vl_xxB.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)

...
# 断点续训权重加载路径
LOAD_PATH="./ckpt/save_dir/Qwen3-VL-xxB-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
    --save $SAVE_PATH \ # 保存路径
"

根据实际情况配置examples/mindspore/qwen3vl/model_xxB.json中的init_from_hf_path参数,该参数表示初始权重的加载路径。 根据实际情况配置examples/mindspore/qwen3vl/model_xxB.json中的image_encoder.vision_encoder.freezeimage_encoder.vision_projector.freezetext_decoder.freeze参数,该参数分别代表是否冻结vision model模块、projector模块、及language model模块。 注:当前examples/mindspore/qwen3vl/model_xxB.json中点各网络层数均为未过校验的无效配置,如需减层请修改原始hf路径下相关配置文件。

【单机运行配置】

配置examples/mindspore/qwen3vl/finetune_qwen3vl_xxB.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))

3. 启动微调

以Qwen3VL-xxB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/mindspore/qwen3vl/finetune_qwen3vl_xxB.sh

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)