parameters not having gradient when calling training_step() outside of trainer

[malfonsoarquimea]

Hi! I am writing pytest tests for a model that I coded and I am now writing the tests for the training_step() function to make sure that everything is working as expected.
My test is simple, I initialize the model and then call model.training_step(batch,0), where batch is a list with data with the same structure than the one that will be provided by the dataloader.
Everything seems to work and the training_step returns a loss as expected, but then when I try to get the gradients of the model with layer.weight.grad I get None. (given that layer is a linear layer of the model). The parameters have the attribute requires_grad set to True when I just print layer.weight.requires_grad

If I understand it correctly, I should be seen some gradients here, and not None values. What am I doing wrong?
Thanks in advance !

[akihironitta]

I believe it's totally expected because gradients will be available only after backward call.