[Proposal] Rename l1_coefficient to sparsity_coefficient

[chanind]

Proposal

Now that we support JumpReLU training, the l1_coefficient is confusing since jumprelu use a l0 loss, not l1, for training. We should rename this parameter to sparsity_coefficient since it is a coefficient used to generally promote sparsity. We should also rename l1_warmup_steps to sparsity_warmup_steps.

Motivation

It is confusing to see l1_coefficient used for JumpReLU training which doesn't use L1 loss.

Alternatives

Alternatively, we could add a separate l0_coefficient / l0_warmup_steps which is only used for jumprelu training and error if l1_coefficient is provided. This would also potentially allow training a jumprelu with both L0 and L1 loss if desired.

Checklist

• I have checked that there is no similar issue in the repo (required)

[muyo8692]

Hi @chanind!
I've created PR #376 that implements one of the alternatives you suggested - adding a separate l0_lambda parameter specifically for JumpReLU training.

The PR:

• Adds a dedicated l0_lambda parameter (default: 0.0)
• Requires explicit l0_lambda specification when using JumpReLU
• Maintains clear separation between l1 and l0 regularization terms

Would love to get your thoughts on this implementation!