Can I train this model?

[WateverOk]

Hello, may I ask, I have 4 3090s with 24G memory. Can I train this model?

[demoyu123]

Hello, may I ask, I have 4 3090s with 24G memory. Can I train this model?

It shouldn't work, the author is doing it on 4 A100s, and I can't drive it with 1 A6000 now.

[deyang2000]

I tried the demo on RTX 3090, and it looks trainable.

Another problem was that I could train on one GPU, but the validation for multi-card training would report errors.

[deyang2000]

I switched to a new 2*RTX 3090 server, which works at full power compared to the previous one. Now, I can train and verify using two GPUs.

[DianiSirimewan]

Can you provide the config file you used for the training? I used the demo file and replaced only the data roots. But it gives me this error.

[deyang2000]

Can you provide the config file you used for the training? I used the demo file and replaced only the data roots. But it gives me this error.

The model you loaded does not seem to match the parameters in the config file. It would help if you switched to a different model. I think that you should switch to the "large" model.

[RockyChason]

Can you provide the config file you used for the training? I used the demo file and replaced only the data roots. But it gives me this error.

The model you loaded does not seem to match the parameters in the config file. It would help if you switched to a different model. I think that you should switch to the "large" model.

When I use the large model, I encounter a RuntimeError: 'The size of tensor a (128) must match the size of tensor b (256) at non-singleton dimension 1.' How do you suggest I resolve this issue?