ascend-robot[Docs] Modify current repository URLs to relative paths

文件	最后提交记录	最后更新时间
npu_adapt	[Feature]Added performance optimization for the Self-Forcing Co-authored-by: qq_31502577<liuqiyuan6@huawei.com> # message auto-generated for no-merge-commit merge: !1832 merge master into master [Feature]Added performance optimization for the Self-Forcing Created-by: qq_31502577 Commit-by: qq_31502577 Merged-by: ascend-robot Description: ## Motivation 新增self forcing模型的性能优化点 ## Modification 1、使能rms_norm融合算子 2、将rope_param默认精度从float64修改为float32 ## Self-test (Optional) 1、经验证，修改后性能提升25% 2、性能优化后，精度符合精度标准 ![image.png](https://raw.gitcode.com/user-images/assets/7404510/05f27d8a-0a7e-4f38-b9bf-5c99f256d818/image.png 'image.png') ## 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!1832	5 个月前
README.md	[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!2360	1 个月前
replace_npu_patch.sh	[Feature] Add Self-Forcing Co-authored-by: qq_31502577<liuqiyuan6@huawei.com> # message auto-generated for no-merge-commit merge: !1716 merge master into master [Feature] Add Self-Forcing Created-by: qq_31502577 Commit-by: qq_31502577 Merged-by: ascend-robot Description: ## Motivation 新增 Add Self-Forcing模型 ## Modification 1、增加Self-Forcing相关的原仓适配patch 2、增加Self-Forcing README ## Self-test (Optional) 1、根据README测试环境安装、训练、推理流程正常 2、12小时长跑对比精度，critic loss平均绝对误差0.0099，generator loss平均绝对误差0.014 - critic loss对比： ![image.png](https://raw.gitcode.com/user-images/assets/7404510/a6968b94-381b-45bb-a476-6503d94f3185/image.png 'image.png') - generator loss： ![image.png](https://raw.gitcode.com/user-images/assets/7404510/8af25a4c-e778-4345-b68d-4782296bad7c/image.png 'image.png') ## 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: - [x] CLA has been signed and all committers have signed the CLA in this PR. - [x] The ci-pipeline is passed, Code Check is passed. See merge request: Ascend/MindSpeed-MM!1716	6 个月前

Self-Forcing

self-forcing
引用
- 公网地址说明

模型介绍

Self Forcing的核心技术路线是在训练自回归视频扩散模型时，摒弃传统的“教师强制”方法，转而让模型基于自身之前生成的、带有噪声的帧来预测下一帧，从而模拟真实的推理生成过程，实现完整的自回归自我展开训练。其核心创新点在于从根本上解决了“暴露偏差”问题，通过迫使模型学习如何从其自身生成的不完美上下文中进行修正，并引入整体分布匹配损失来优化整个生成序列的全局质量，从而显著提升了生成的一致性和稳定性。主要用途是实现高质量、高帧率且时间一致性强的实时视频生成

版本说明

参考实现

url=https://github.com/guandeh17/Self-Forcing
commit_id=33593df3e81fa3ec10239271dd2c100facac6de1

变更记录

2025.11.12：首次发布Self-Forcing

环境搭建

【模型开发时推荐使用配套的环境版本】

昇腾基础软件安装请参考安装指南

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

在工作目录执行下列命令

git clone --branch 26.0.0 https://gitcode.com/Ascend/MindSpeed-MM.git
cd MindSpeed-MM
git clone https://github.com/guandeh17/Self-Forcing.git
bash examples/self_forcing/replace_npu_patch.sh
cd Self-Forcing

安装依赖
```
pip install -r requirements.txt
```

下载权重

huggingface-cli download Wan-AI/Wan2.1-T2V-1.3B --local-dir-use-symlinks False --local-dir wan_models/Wan2.1-T2V-1.3B
huggingface-cli download gdhe17/Self-Forcing checkpoints/self_forcing_dmd.pt --local-dir .
huggingface-cli download gdhe17/Self-Forcing checkpoints/ode_init.pt --local-dir .
huggingface-cli download gdhe17/Self-Forcing vidprom_filtered_extended.txt --local-dir prompts

训练

执行以下命令即可开启训练

torchrun --nnodes=1 --nproc_per_node=8 \
  train.py \
  --config_path configs/self_forcing_dmd.yaml \
  --logdir logs/self_forcing_dmd \
 --disable-wandb

推理

执行以下命令即可开启推理

python inference.py \
    --config_path configs/self_forcing_dmd.yaml \
    --output_folder videos/self_forcing_dmd \
    --checkpoint_path checkpoints/self_forcing_dmd.pt \
    --data_path prompts/MovieGenVideoBench_extended.txt \
    --use_ema

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
`OMP_NUM_THREADS`	设置执行期间使用的线程数	需要配置为整数

引用

公网地址说明

代码涉及公网地址参考公网地址