linear2d_layer_no_bias: add optional argument to disable biases

[OneAdder]

Add option to disable biases

Simple additional option. Removing biases helps speed up inference. This technique is used in Llama and, to an extent, in Qwen. Being able to disable biases here is a simple quality of life addition

[jvdp1]

Being able to disable biases here is a simple quality of life addition

Should we extend this to all types of layers?

[OneAdder]

@jvdp1 yep, I think we should

[OneAdder]

Perhaps, I should rename it to FeedForward or MLP. Thoughts?

[milancurcic]

I don't think MLP is appropriate because it refers to a network. We already have dense. Since you mention naming, and I don't recall if we discussed this in the original linear2d PR: linear layer is just a special case of a dense layer without an activation (which we have via the linear activation function). Should this be called dense2d for consistency?

Regarding the question of whether this should be an option for all layers. My opinion depends on whether this is used more broadly in various layers. If it's used only in dense2d (or linear2d) layers, this option should be available only here.

[milancurcic]

This also means that we could have a generic maxpool after all. We could make it so that maxpool1d is invoked by maxpool(pool_size=2), while maxpool2d is invoked by maxpool(pool_size=[2, 2]), because a scalar and array are type-kind-rank distinguishable. However, using this pattern we couldn't extend this to maxpool3d because maxpool(pool_size=[2, 2, 2]) would not be distinguishable from the 2d variant.

Alternatively, we could do:

maxpool(pool_width=2)  ! maxpool1d
maxpool(pool_width=2, pool_height=2)  ! maxpool2d
maxpool(pool_width=2, pool_height=2, pool_depth=2)  ! maxpool3d

Same approach could apply to conv layers.

The more I think about this the more I like it. What do you think?

[OneAdder]

@milancurcic Disregard the naming comment. It belongs in a different PR: #208
I like the maxpool ideas

[OneAdder]

@milancurcic Should we merge it? I really would like to do it before the dense2d refactoring

[milancurcic]

Yes, that's fine, please go ahead and squash-merge, thank you!