Well, I am interested in designing a WGAN using pytorch-lightning, but feel trouble with limiting generator gradient clip and discriminator not clip.

[forechoandlook]

I do not want to use manual optimizer in LightningModule .

[rohitgr7]

for now, the internal gradient clipping procedures will clip gradients for both the generator and discriminator, but you can configure it yourself by disabling the one from lightning for now and configuring it manually

def on_before_optimizer_step(self, trainer, pl_module, optimizer, opt_idx):
    if opt_idx == 0:  # assuming that opt_idx=0 is for generator
        params = ...
        torch.nn.utils.clip_grad_norm_(params, clip_val)

we will add support for it in the future :)