how can we compute the loss of chained models with keras?

@Begoodpy, I suggest you combine the 2 models into a single one and train it as you would usually do.

supermodel = keras.Sequential(
    [
      model1(),
      model2(),
    ]

If you need more control over the models, try this:

all_vars = model1.trainable_variables + model2.trainable_variables

grads = tape.gradient(loss_value2, all_vars)
optimizer.apply_gradients(zip(grads, all_vars))