Custom Training for multiple LoRAs for the same model. about peft HOT 4 CLOSED

Note that set_adapter enables requires_grad on the adapter whose name was passed and turns it off on all other adapters. However, since the PEFT model is a PyTorch nn.Module, as always, you can just iterate through the modules and enable or disable gradient according to your wishes:

model.set_adapter("adapterA")
# now only adapter A's weights have gradients enabled
for name, param in model.named_parameters():
    if ...:  # do some check, e.g. `if "lora_" in name`
        param.requires_grad = True

from peft.

Thank you for your answer.
However, this isn't quite what I had in mind.

How can I set the different adapters inside the forward path?
For example, does the following code work?
Because setting requires_grad = True inside the forward path seems little bit weird.

# Inside the model
def forward(self, x, y):
  model_a.set_adapter("adapter_x")
  output_x = model_a(x)
  
  model_a.set_adapter("adapter_y")
  output_y = model_a(y)
  
  final_output = model_b(output_x, output_y)
  return final_output

output = model(x, y)
loss = loss_fn(output)
loss.backward()

from peft.

It's really hard to answer your question with this little context. The code you pasted should work. It means that for output_x, only gradients for adapter_x etc. If you try this, do you get an error or does the training work?

from peft.

I've tried it, and the code works :)
I misunderstood the set_adapter function.
I thought the line model_a.set_adapter("adapter_y") freezes the other adapters (such as model_a.set_adapter("adapter_x")).
Thanks for your replies :)

from peft.