实验

[tangjw20]

很棒的工作！我想知道这个训练硬件资源的要求，大概需要多少显存呢？

[Xiaoc7]

很棒的工作！我想知道这个训练硬件资源的要求，大概需要多少显存呢？

Thank you! We trained the diffusion model using an A100 80GB GPU. Without enabling --gradient_checkpointing, the peak VRAM usage is around 48GB at a 576×1024 input resolution. So I think you would need a GPU with about 48GB of VRAM to train the model, or you can reduce the input resolution to fit it into a smaller GPU(e.g. Crop mode input in our code).

[tangjw20]

很棒的工作！我想知道这个训练硬件资源的要求，大概需要多少显存呢？

Thank you! We trained the diffusion model using an A100 80GB GPU. Without enabling --gradient_checkpointing, the peak VRAM usage is around 48GB at a 576×1024 input resolution. So I think you would need a GPU with about 48GB of VRAM to train the model, or you can reduce the input resolution to fit it into a smaller GPU(e.g. Crop mode input in our code).

Thanks！I am a beginner. May I ask if this supports multi-GPU training? I have six RTX 3090 GPUs.

[Xiaoc7]

很棒的工作！我想知道这个训练硬件资源的要求，大概需要多少显存呢？

Thank you! We trained the diffusion model using an A100 80GB GPU. Without enabling --gradient_checkpointing, the peak VRAM usage is around 48GB at a 576×1024 input resolution. So I think you would need a GPU with about 48GB of VRAM to train the model, or you can reduce the input resolution to fit it into a smaller GPU(e.g. Crop mode input in our code).

Thanks！I am a beginner. May I ask if this supports multi-GPU training? I have six RTX 3090 GPUs.

Hi! Sorry for the delayed reply. Our work supports multi-GPU training, but it uses DDP rather than model parallelism/model splitting. Therefore, it may not be possible to train our model on 3090 GPUs without downsampling the input resolution.

[coolfeel]

很棒的工作！我想知道这个训练硬件资源的要求，大概需要多少显存呢？

Thank you! We trained the diffusion model using an A100 80GB GPU. Without enabling --gradient_checkpointing, the peak VRAM usage is around 48GB at a 576×1024 input resolution. So I think you would need a GPU with about 48GB of VRAM to train the model, or you can reduce the input resolution to fit it into a smaller GPU(e.g. Crop mode input in our code).

你好，可以请问一下用A100训练扩散模型需要多久？谢谢！

[Xiaoc7]

很棒的工作！我想知道这个训练硬件资源的要求，大概需要多少显存呢？

Thank you! We trained the diffusion model using an A100 80GB GPU. Without enabling --gradient_checkpointing, the peak VRAM usage is around 48GB at a 576×1024 input resolution. So I think you would need a GPU with about 48GB of VRAM to train the model, or you can reduce the input resolution to fit it into a smaller GPU(e.g. Crop mode input in our code).

你好，可以请问一下用A100训练扩散模型需要多久？谢谢！

It depends on the scale of your data. In our case, training the model takes about 1.5 days. However, the model converges quite quickly, so it may not require a very long training time.