DiffSynth-Studio

DiffSynth-Studio copied to clipboard

DiffSynth-Studio/examples/wanvideo/model_inference/Wan2.2-Fun-A14B-Control.py 多卡推理出现噪声

3

用这个脚本单卡推理时没有问题，多卡推理时后半段会成为噪声，请问这个是什么原因呢，是不是多卡推理时应当增加什么设置，或者是当前脚本还不支持多卡推理？ 推理时设置 `pipe = WanVideoPipeline.from_pretrained( torch_dtype=torch.bfloat16, device="cuda", use_usp=True, #主要加这个更改 model_configs=[ …… ], )` 运行时使用 torchrun --nproc_per_node=2 Wan2.2-Fun-A14B-Control.py 万望解答，不胜感激！

PrivateMaRyan

Distill

1

请问要准备的蒸馏数据，num_inference_steps,cfg_scale表示要蒸馏到对应的步数和cfg吗，还是说需要使用diffusiondb数据集中的原始数据50 | 7

coder4nlp

Regarding multi-machine multi-GPU training for FLUX kontext

1

Hello, I have carefully reviewed the previous issues and found that you don't seem to plan to support single-card multi-batch training for various models. But, may I ask, can I...

wangjiyuan9

可以支持Wan2.2的gguf模型吗？

1

可以支持Wan2.2的gguf模型吗？当前默认的bf16推理压力在一般机器上还是太大了

oath-sign

Inpainting or swapping using Wan VACE

3

Hi, I tried to perform video inpainting/swapping with Wan VACE, but unfortunately the results were poor. For example, if I provide a video with a gray mask as a 'vace_video'...

julkaztwittera

Operation timeout when using Wan2.2-Fun-A14B-Control with multiple GPUs

2

Hi, I am grateful for this amazing repository for video conditioned Wan2.2s. I am trying to use `Wan2.2-Fun-A14B-Control` and give a depth video for generation. I tried on a single...

FYRichie

NaN Loss after Forward Pass on Step 4 when Full Fine-Tuning Wan 2.2 5B

9

Hi, I am currently full fine-tuning the Wan 2.2 5B model, but the **loss becomes NaN right after the forward process at the 4th training step** every time. I have...

wyyfffff

关于FLUX kontext的多机多卡训练

2

您好，我仔细查看了之前的issue，发现您似乎不准备支持各个模型的单卡多batch训练。但请问，我是否可以简单的把： ``` accelerate launch \ examples/flux/model_training/train.py \ ``` 替换为 ``` accelerate launch \ --multi_gpu \ --num_processes 3 \ --gpu_ids 5,6,7 \ --mixed_precision bf16 \ examples/flux/model_training/train.py \ ``` 来实现多卡训练，以及后面的多机训练都是直接改这里即可，且保证训练逻辑完全正确？

wangjiyuan9

使用deepspeed zero_stage 3训练，遇到张量shape不一致的问题该怎么解决？

16

同样的工程使用deepspeed zero_stage 2训练，不会出现shape不一致的问题，但会出现显存OOM。 推测是zero_stage 3分片后导致张量shape改变，求助解决方案

RockyLQ1

wan2.2-S2V training always got OOM on H100-80GB GPU

2

tried `examples/wanvideo/model_training/full/Wan2.2-S2V-14B.sh` but always got OOM whatever how to reduce height/width/num_frames

Eurus-Holmes