TP Checkpoints Loaded always in a Non-strict mode And Can base_model.pt be trained?

[Zijiuyang]

Thank you for your open-resource and wonderful work. I found problems when I pretrain or posttrain the transfer1 model, when I used the depth controlnet. When posttraining, there are warnings like:
[12-12 07:35:14|CRITICAL|cosmos_transfer1/diffusion/training/models/model_ctrl.py:133:load_base_model] [RANK 5]_IncompatibleKeys(missing_keys=['net.blocks.block0.blocks.0.block.attn.to_q.0.weight', 'net.blocks.block0.blocks.0.block.attn.to_k.0.weight', 'net.blocks.block0.blocks.0.block.attn.to_v.0.weight', 'net.blocks.block0.blocks.0.block.attn.to_out.0.weight', 'net.blocks.block0.blocks.1.block.attn.to_q.0.weight
.........
After I finished the posttraining, I used the python script convert_ckpt_tp_to_fsdp.py, in an effort to merge the TP checkpoints so that the posttrained checkpoints can be used for inference. And I find the same issue like:
[12-12 07:35:14|CRITICAL|cosmos_transfer1/diffusion/training/models/model_ctrl.py:133:load_base_model] [RANK 5]_IncompatibleKeys(missing_keys=['net.blocks.block0.blocks.0.block.attn.to_q.0.weight', 'net.blocks.block0.blocks.0.block.attn.to_k.0.weight', 'net.blocks.block0.blocks.0.block.attn.to_v.0.weight', 'net.blocks.block0.blocks.0.block.attn.to_out.0.weight', 'net.blocks.block0.blocks.1.block.attn.to_q.0.weight.

Although the checkpoints can be used for inference without error, I still want to know whether the above issues will cause any impact on the model performance.

BTW, I notice that all the training guide is for the control weights. May you share that if the base_model can be trained?

Thanks for reading this issue.