outputs.aggregated_viusal_token_attention not find

[CatchBeliF-TAT]

There seems to be an issue in the code（SGL/internvl/model/internvl_chat/modeling_internvl_chat.py） — it looks like outputs does not contain the attribute aggregated_viusal_token_attention. Could this be due to a problem with my data or is there something missing in the code? I'd really appreciate your help

[wangbo-zhao]

Hi, sorry for replying late. Have you solve this problem? Can you take a breakpoint here to debug?

[heliossun]

Same problem here. Which Transformers version should we use, I tried 4.37.2, 4.42.0 ant the latest one, but all has the same error, seems like the _sample() function were not used in .generate()

[wangbo-zhao]

Same problem here. Which Transformers version should we use, I tried 4.37.2, 4.42.0 ant the latest one, but all has the same error, seems like the _sample() function were not used in .generate()

I would like to confirm whether you are using FlashAttention, as the small model is not compatible with it. Additionally, please ensure that generation_config.output_attentions is set to True for the small model.

[wangbo-zhao]

please try transformers==4.44.2