仓库:https://github.com/modelscope/DiffSynth-Studio
Commit:4dbf980d4d0eb34eda136300dd0d72014cff8965
版本:2.1.6安装器处理 11 个官方源码文件,并加入 17 个昇腾环境、测试、训练、下载、验收和说明文件。官方源码文件通过 sed -i 更新;新增文件从补丁载荷复制。
下载官方源码:
wget -O DiffSynth-Studio.tar.gz \
https://github.com/modelscope/DiffSynth-Studio/archive/4dbf980d4d0eb34eda136300dd0d72014cff8965.tar.gz
tar -xzf DiffSynth-Studio.tar.gz
mv DiffSynth-Studio-4dbf980d4d0eb34eda136300dd0d72014cff8965 DiffSynth-Studio解压本补丁包,并执行:
bash diffsynth-studio-ascend-910b-patch-20260903/install.sh \
/absolute/path/to/DiffSynth-Studio成功标志:
DiffSynth-Studio Ascend 910B patch: PASS独立校验:
bash diffsynth-studio-ascend-910b-patch-20260903/verify.sh \
/absolute/path/to/DiffSynth-Studio安装脚本可以重复执行;已经应用的文件会显示 present。
cd /absolute/path/to/DiffSynth-Studio
python tests/test_ascend_wan_rope.py
python scripts/ascend_npu_smoke.py本说明覆盖从镜像构建、容器启动、模型与数据下载,到 Wan2.2 LoRA 冒烟训练和结果验收的完整流程。适用训练目标如下:
| 模型 | 任务 | 训练单元 |
|---|---|---|
| Wan2.2-TI2V-5B | T2V | 1 个 LoRA |
| Wan2.2-TI2V-5B | I2V | 1 个 LoRA |
建议先完成 TI2V-5B T2V 冒烟训练,再执行 TI2V-5B I2V。
宿主机工作根目录:
export WORK_ROOT=/path/to/workspace
export REPO_ROOT=${WORK_ROOT}/DiffSynth-Studio
export MODEL_ROOT=${WORK_ROOT}/models
export OUTPUT_ROOT=${WORK_ROOT}/outputs
export LOG_ROOT=${WORK_ROOT}/logs
mkdir -p "${MODEL_ROOT}" "${OUTPUT_ROOT}" "${LOG_ROOT}" "${REPO_ROOT}/data"本文后续命令均使用以上目录。
cd "${REPO_ROOT}"
test -f UPSTREAM_COMMIT源码下载与补丁安装按照补丁包根目录的 README.md 完成。
镜像基于以下 Quay 基础镜像:
quay.io/ascend/mindspeed-mm:v26.1.0-cann9.1.0-torch_npu2.7.1.post8-910b-ubuntu22.04-py3.11仓库中的 docker/Dockerfile.ascend 固定了以下框架组合:
torch 2.7.1
torch-npu 2.7.1.post8
torchvision 0.22.1
torchaudio 2.7.1它同时安装训练依赖、音视频系统库、DiffSynth-Studio-Ascend,并在构建阶段执行版本检查和 Wan RoPE 回归测试。
构建镜像:
cd "${REPO_ROOT}"
docker build --pull \
-f docker/Dockerfile.ascend \
-t diffsynth-wan22-ascend:torch2.7.1-cann9.1 .构建成功的最后阶段会显示:
DiffSynth Ascend framework versions: OK
Ran 4 tests
OKdocker run --rm -it \
--name diffsynth-wan22 \
--network=host \
--ipc=host \
--device=/dev/davinci0 \
--device=/dev/davinci_manager \
--device=/dev/devmm_svm \
--device=/dev/hisi_hdc \
-v /usr/local/dcmi:/usr/local/dcmi \
-v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
-v /usr/local/Ascend/driver/lib64:/usr/local/Ascend/driver/lib64 \
-v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
-v /etc/ascend_install.info:/etc/ascend_install.info \
-v "${MODEL_ROOT}:/workspace/models" \
-v "${OUTPUT_ROOT}:/workspace/outputs" \
-v "${LOG_ROOT}:/workspace/logs" \
-v "${REPO_ROOT}/data:/workspace/DiffSynth-Studio-Ascend/data" \
diffsynth-wan22-ascend:torch2.7.1-cann9.1进入容器后的代码目录为:
/workspace/DiffSynth-Studio-Ascend在容器内执行:
cd /workspace/DiffSynth-Studio-Ascend
npu-smi info
python scripts/ascend_npu_smoke.py
python tests/test_ascend_wan_rope.py验证结果:
PASS: BF16 SDPA
PASS: BF16 Conv3D
PASS: FP32 AdamW forward/backward
Ran 4 tests
OK在容器内设置统一模型目录:
export DIFFSYNTH_MODEL_BASE_PATH=/workspace/models
export DATA_BASE_PATH=/workspace/DiffSynth-Studio-Ascend/data/diffsynth_example_dataset下载 TI2V-5B 及其数据:
bash scripts/download_wan22_lora_assets.sh ti2v-5b命令完成后输出:
Wan2.2 <profile> assets: OK下载脚本会完整保存对应模型仓库,并准备:
metadata.csv。在容器内执行:
cd /workspace/DiffSynth-Studio-Ascend
export DIFFSYNTH_MODEL_BASE_PATH=/workspace/models
export DIFFSYNTH_SKIP_DOWNLOAD=True
export DIFFSYNTH_DEVICE=npu
export DIFFSYNTH_ATTENTION_IMPLEMENTATION=torch
export TOKENIZERS_PARALLELISM=false
export HCCL_CONNECT_TIMEOUT=1800训练使用 BF16 模型计算和 FP32 AdamW 优化器状态。模型、数据和输出路径在所有训练命令中保持一致。
NPROC_PER_NODE=1 \
HEIGHT=256 \
WIDTH=256 \
NUM_FRAMES=9 \
LORA_RANK=8 \
DATASET_REPEAT=1 \
OUTPUT_ROOT=/workspace/outputs \
bash examples/wanvideo/ascend/train_lora_smoke.sh ti2v-5b-t2v \
2>&1 | tee /workspace/logs/ti2v-5b-t2v.log验证训练输出:
python scripts/verify_lora_run.py \
/workspace/outputs/ti2v-5b-t2v通过时输出:
LoRA checkpoints: <数量>
Loss records: <数量>
Loss range: <最小值> .. <最大值>
LoRA run verification: PASSNPROC_PER_NODE=1 \
HEIGHT=256 \
WIDTH=256 \
NUM_FRAMES=9 \
LORA_RANK=8 \
DATASET_REPEAT=1 \
OUTPUT_ROOT=/workspace/outputs \
bash examples/wanvideo/ascend/train_lora_smoke.sh ti2v-5b-i2v \
2>&1 | tee /workspace/logs/ti2v-5b-i2v.log
python scripts/verify_lora_run.py \
/workspace/outputs/ti2v-5b-i2v冒烟训练全部通过后,沿用相同脚本并调整训练规模:
NPROC_PER_NODE=1 \
HEIGHT=480 \
WIDTH=832 \
NUM_FRAMES=49 \
LORA_RANK=32 \
DATASET_REPEAT=100 \
NUM_EPOCHS=5 \
SAVE_STEPS=100 \
LORA_TARGET_MODULES=q,k,v,o,ffn.0,ffn.2 \
OUTPUT_ROOT=/workspace/outputs \
bash examples/wanvideo/ascend/train_lora_smoke.sh ti2v-5b-t2v \
2>&1 | tee /workspace/logs/ti2v-5b-t2v-train.log将最后一个参数替换为相应训练单元即可执行其余任务:
ti2v-5b-i2v每个正式训练任务结束后执行 scripts/verify_lora_run.py,其 PASS 结果作为本次训练产物验收记录。
每个训练单元必须同时满足:
safetensors;loss.csv 存在有效 loss 记录;scripts/verify_lora_run.py 输出 LoRA run verification: PASS。