How to get prompt token-wise cross-attention for each time-step in stable diffusion

[shashankg7]

I am new to diffusion and diffusers library. I am working on a project where I need a token-wise attention for each timestep in a pre-trained stable diffusion model.

I am trying to modify src/diffusers/models/attention_processor.py to get that, but I am confused with the dimensions. Can anyone point me to a direction where I can extract token-wise attention map for prompts?