文件最后提交记录最后更新时间
[Docs] Correct the description in the document Co-authored-by: LKONE<wanglikai4@huawei.com> # message auto-generated for no-merge-commit merge: !2363 merge 26.0.0 into 26.0.0 [Docs] Correct the description in the document Created-by: wanglikai1019 Commit-by: wanglikai1019;LKONE Merged-by: ascend-robot Description: ## What this PR does / why we need it? 修改部分readme文档中的错误描述。 ## Does this PR introduce any user-facing change? 无 ## How was this patch tested? 无 See merge request: Ascend/MindSpeed-MM!23631 个月前
【Feature】 Improve GLM4.5V performance & update readme Co-authored-by: weixin_44031810<gaojie75@huawei.com> # message auto-generated for no-merge-commit merge: !1879 merge perf into master 【Feature】 Improve GLM4.5V performance & update readme Created-by: weixin_44031810 Commit-by: weixin_44031810 Merged-by: ascend-robot Description: ## Motivation Please describe the motivation of this PR and the goal you want to achieve through this PR. ## Modification Please briefly describe what modification is made in this PR. ## 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**: - [ ] The new code needs to comply with the Clean Code specification. - [ ] 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!18795 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
[Feature] Support Glm4.5v Co-authored-by: weixin_44031810<gaojie75@huawei.com> # message auto-generated for no-merge-commit merge: !1841 merge master into master [Feature] Support Glm4.5v Created-by: weixin_44031810 Commit-by: weixin_44031810 Merged-by: ascend-robot Description: ## Motivation Please describe the motivation of this PR and the goal you want to achieve through this PR. ## Modification Please briefly describe what modification is made in this PR. ## 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**: - [ ] The new code needs to comply with the Clean Code specification. - [ ] 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!18415 个月前
[Bugfix][Refactor] resolve aux loss compatibility with various loss types and refactor vlm loss Co-authored-by: liyingxuan<liyingxuan3@huawei.com> # message auto-generated for no-merge-commit merge: !1990 merge vlm_loss_modify into master [Bugfix][Refactor] resolve aux loss compatibility with various loss types and refactor vlm loss Created-by: liyx616 Commit-by: liyx616;liyingxuan Merged-by: ascend-robot Description: ## Motivation resolve aux loss compatibility with various loss types and refactor vlm loss. ## Modification ### 修复以下问题 1. per-token-loss和aux loss不兼容 2. aux loss没有被正确加到最终的loss中 3. aux loss不支持CP 4. token-loss和square-loss不支持chunkloss ### 重构以下内容 1. 对入口为pretrain_transformers.py的理解模型loss进行归一,统一使用model.json中的loss_cfg: {loss_type: "default"} 字段来管理loss_type, 解决之前loss type管理混乱的问题,全部禁用megatron的--calculate-per-token-loss和之前我们自己定义的--calculate-xxx-loss,在loss计算部分和megatron进行完全解耦 当前loss_cfg的配置示例如下: ```json "loss_cfg": { "compute_mode": "default", "chunk_size": 1024, "router_aux_loss_coef": 1.0, "loss_type": "default" } ``` 2. 对transformers_model.py中的compute_language_model_loss_cp、compute_language_model_loss和build_loss_ctx三个函数进行归一,避免加入一个新的loss type时,进行霰弹式修改,导致各种特性间的兼容性出现错乱 ### 基本设计思路 #### per-token-loss兼容问题 之前兼容性问题的根因是megatron的per-token-loss实现比较抽象,是先计算token sum loss,然后在梯度累积的最后一步对梯度div global_token_num 之所以要这么做的原因是per-token-loss需要所有梯度累积步的global_token_num, 而当前step只能拿到当前step的token_num。 transformers仓的做法比较暴力,是先遍历了一遍数据集,这样就能知道所有step的token num,但是这个是不现实的,如果数据集很大的话,初始化时在遍历数据集这里就会卡很久。因此这里设计了一个PrefetchGradAccDataLoader,对torch原生的DataLoader进行了封装,在执行iter的时候,会预取grad acc步的数据到buffer中,这样就可以提前知道当前grad acc 个step中总的token_num,不需要提前遍历数据集。而且预取的数据放在host侧,且大小可控。 #### loss计算归一问题 所有的vlm_loss的计算都可以用mindspeed_mm/models/common/chunkloss.py中的fixed_cross_entropy进行归一,只是reduction和alpha不同 ```python # Compute standard cross-entropy loss loss = torch.nn.functional.cross_entropy(source, target, ignore_index=ignore_index, reduction=reduction) if alpha is not None: alpha = alpha.to(loss.device) if reduction == "sum": # Global normalization: total loss divided by scalar alpha loss = loss / alpha elif reduction == "none": if alpha.ndim == 0: # Alpha is a scalar: sum all element-wise losses, then divide loss = loss.sum() / alpha else: # Alpha is 1-D with shape (N,): assume N examples # Reshape loss to (N, -1) to group elements per example loss = loss.view(alpha.shape[0], -1) # Sum over non-batch dimensions (e.g., sequence length in token classification) # Normalize each example's loss by its corresponding alpha loss = loss.sum(1) / alpha loss = loss.sum() else: raise ValueError(f"Unsupported reduction mode: {reduction}. Use 'sum' or 'none'.") ``` 通过数学推导,可以得出以下5中loss的alpha和reduction | loss type | reduction | alpha | |--|--|--| | default | sum | loss_mask.sum | | per-sample-loss | none | loss_mask.sum(1) * loss_mask.shape[0] | | per-token-loss | sum | all_reduce(kwargs['grad_acc_avg_tokens'], op=avg) | | token-loss | none | weight_sum = loss_mask.sum(1) * loss_mask.shape[0] | | square-loss | none | 见代码 | ## 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**: - [ ] The new code needs to comply with the Clean Code specification. - [ ] 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!19904 个月前
README.md

Glm4.5v 使用指南

目录

版本说明

参考实现

url=https://github.com/huggingface/transformers.git
commit_id=8cb5963

变更记录

2025.11.29: 首次支持Glm4.5v模型


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。

Python版本推荐3.10,torch和torch_npu版本推荐2.7.1版本

‼️MoE部分的加速特性依赖较新版本的torch_npu和CANN,推荐使用以下版本

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
git checkout d76dbddd4517d48a2fc1cd494de8b9a6cfdbfbab

# 安装mindspeed及依赖
pip install -e .
cd ..
# 安装mindspeed mm及依赖
pip install -e .

# 安装新版transformers(支持glm4.5v模型)
git clone https://github.com/huggingface/transformers.git
cd transformers
git checkout 8cb5963
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/GLM-4.5V目录下。(*表示对应的尺寸)

为了使网络能够流畅运行,MindSpeed MM修改了moe中专家的结构名称,需对原始预训练权重进行转换:

mm-convert ExpertMergeDcpConverter hf_to_dcp --hf_dir "ckpt/hf_path/GLM-4.5V" --save_dir "ckpt/mm_path/GLM-4.5V"

由于glm4.5v参数量较大, 必须使用meta init初始化加载权重; 需要在examples/glm4.5v/finetune_glm4_5v106B.sh的GPT_ARGS中加入--init-model-with-meta-device参数. 默认脚本中已添加。 此外,使用meta init初始化加载权重时,需要将examples/glm4.5v/finetune_glm4_5v106B.sh中LOAD_PATH设置为上述权重转换完后保存的路径"ckpt/mm_path/GLM-4.5V"。


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_106B.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_106B.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_106B.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

修改示例如下,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/GLM-4.5V",
            ...
        },
        "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_106B.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",
        ...
    }
}

【模块冻结配置】

当前支持vision encoder、vision projector、text decoder及lm head模块的冻结,其中,vision encoder、vision projector默认训练时为冻结状态,

通过配置model.json文件中各个模块的freeze字段,来修改各个模块的冻结与否。

【模型保存加载及日志信息配置】

根据实际情况配置examples/glm4.5v/finetune_glm4_5v_106B.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/GLM-4.5V"
# 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/glm4.5v/model_106B.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/glm4.5v/model_106B.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/glm4.5v/model_106B.json中的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。 为了在NPU上更快得运行模型训练,moe模块可使能NPU亲和的融合算子,在examples/glm4.5v/model_106B.json中配置"use_npu_fused_moe"为true即可,此处默认配置为true。

【单机运行配置】 注意:单机只能跑减层,可作为模型调试用;完整模型运行请使用多机配置。 配置examples/glm4.5v/model_106B.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=16  # 单机减层可跑
MASTER_ADDR=localhost
MASTER_PORT=29501
NNODES=1
NODE_RANK=0
WORLD_SIZE=$(($NPUS_PER_NODE * $NNODES))

【多机运行配置】

配置examples/glm4.5v/model_106B.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
# 根据分布式集群实际情况配置分布式参数
NPUS_PER_NODE=16  # 每个节点的卡数,以实际情况填写
MASTER_ADDR="your master node IP"  # 都需要修改为主节点的IP地址(不能为localhost)
MASTER_PORT=6000
NNODES=8  # 集群里的节点数,以实际情况填写
NODE_RANK="current node id"  # 当前节点的RANK,多个节点不能重复,主节点为0, 其他节点可以是1,2..
WORLD_SIZE=$(($NPUS_PER_NODE * $NNODES))

3. 启动微调

启动微调训练任务需要确认loss计算方式。
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/glm4.5v/model_106B.sh

4. hf权重转换

训练完成之后,将保存在save_dir目录下的权重转换成huggingface格式

mm-convert ExpertMergeDcpConverter dcp_to_hf --hf_dir "ckpt/hf_path/GLM-4.5V" --dcp_dir "save_dir/iter_000xx" --save_dir "ckpt/dcp_to_hf/GLM-4.5V"

其中,--hf_dir表示原始huggingface权重的路径,--dcp_dir表示微调后的权重保存路径,iter_000xx表示保存的第xx步的权重,--save_dir表示转换后的权重保存路径。

完成权重转换之后,即可使用相关库进行推理。

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