If VisionLlama causal masked?

[ma-xu]

Congrats on your great work!

A quick question: is causal masking introduced in VisionLlama like Llama? making the model autoregressive. Just want to double confirm that.

Thanks and congrats again!

[cxxgtxy]

Congrats on your great work!

A quick question: is causal masking introduced in VisionLlama like Llama? making the model autoregressive. Just want to double confirm that.

Thanks and congrats again!

Not yet. It's not causal masked.

[Charles-Xie]

I would like to ask more about this interesting question.

I understand that VisionLLama is not causal-masked. I wonder if the authors have tried changing it to a causal-masked version.
And if so, is the performance comparable to the bi-directional attention version, or will the performance drop significantly?

It would be nice and helpful to me if the authors can shed light on this interesting question.
Thanks a lot in advance!

[cxxgtxy]

The causal-masked version achieves about 2-3% lower Top-1 accuracy, which is regarded as siginificant.

[Charles-Xie]

The causal-masked version achieves about 2-3% lower Top-1 accuracy, which is regarded as siginificant.

Thanks for sharing this information! It is very helpful to me.