Questions regarding custom Attention Net

Hi there, so I’m quite new to RLLib and want to try to build my own custom transformer, but so far I have been recieveing quite a few bugs (one of them is that the model being stuck in “PENDING” stage.

I tried to implement the default attention_net provided by the RLLIB and run it as a custom model (TrXLNet) but received the same error. It also found a value error at line 108 regarding the fan_in_layer=None(E_out),

I have changed the parameters to be used for the custom_model_config (e.g. self.num_transformer_units = num_transformer_units to customer_model_config[“num_transformer_unit”].

I’m not sure if I’m doing this correctly. If anyone could guide me on how to use this TrXLNet as a custom model, I would appreciated it. I can’t use the default model since I plan to customise it later.

EDIT: I got the model running using the GTrXL model (where fan_in_layer was not None), but not the original TrXL model. So I guess the issue lies within that loop close box of the multihead attention module.