search

chaojie / ComfyUI-MuseTalk

MIT License
189 stars 26 forks source link

您好，有个问题请教 #4

Open T8star1984 opened 5 months ago

T8star1984 commented 5 months ago

ComfyUI-MuseTalk这个插件会直接加载模型进GPU，占用显存，一启动comfyui就占用了7.5显存，请看看是否有这个问题 如果是的话，是否可以把模型剪切出来，就不会加载节点，然后就没有东西占用显存了。当然具体原理我也不是特别明白，还请排查一下

DeepAichemist commented 4 months ago

可以看看这个帖子https://blog.csdn.net/debug_ding/article/details/138845805?spm=1001.2014.3001.5502

xtkdige commented 1 month ago

ComfyUI-aki-v1.3\custom_nodes\ComfyUI-MuseTalk\nodes.py的160行代码 pred_latents = unet.model(latent_batch, timesteps, encoder_hidden_states=audio_feature_batch).sample 修改为如下吗？ pred_latents = unet.model(latent_batch, timesteps, audio_feature_batch = audio_feature_batch.to(torch.float16)).sample

DeepAichemist commented 1 month ago

ComfyUI-aki-v1.3\custom_nodes\ComfyUI-MuseTalk\nodes.py的160行代码 pred_latents = unet.model(latent_batch, timesteps, encoder_hidden_states=audio_feature_batch).sample 修改为如下吗？ pred_latents = unet.model(latent_batch, timesteps, audio_feature_batch = audio_feature_batch.to(torch.float16)).sample

是两行代码，不能合并到一起，你看一下文章中有提到