FIX: Removed duplicate convolution for DoRA & fixed error for ConvNd layers using "groups"

[gslama12]

This pull request fixes two problems:

1. Duplicate Convolution in DoRA implementation for ConvNd Layers:
Since the base layer convolution is already computed in layer.py, we don't need to compute it again in the dora.py. Computing it again doubles the FLOPs consumption during the forward pass resulting in significantly higher FLOPs overall. We can pass the result from the base layer computed in layer.py to the forward pass of the _DoraConvNdLayer in dora.py and save computational resources.

2. Bugfix for DoRA regarding Convolutional Layers using the Groups Argument:
CNNs that for example use depthwise separable convolutional layers result in an error when applying DoRA. Adjusting the dimension of the conv_layer in layer.py fixes this issue.

[BenjaminBossan]

Thanks for this PR. We're aware of this potential inefficiency but I think it's not as easy as re-using the base result. The reasoning is explained here. Back then, we were only dealing with linear layers but I'm certain that the same logic applies to convolutional layers.

The good news is that this optimization is indeed possible if dropout is set to 0 or if we're in eval mode, see #2122. LMK what you think.

[gslama12]

Thanks for the clarification! Would it be possible to apply the dropout for DoRA similar to how LoRA handles it, i.e. in lora_B(lora_A(x)) during the forward pass? It seems like LoRA also uses the pre-dropout information (result) here and does not re-compute the convolution: result = result + lora_B(lora_A(dropout(x))) * scaling (ln 1120 in layer.py)?

Also, what do you think of the fix for convolutional layers using the "groups" argument?

[BenjaminBossan]

Would it be possible to apply the dropout for DoRA similar to how LoRA handles it, i.e. in lora_B(lora_A(x)) during the forward pass? It seems like LoRA also uses the pre-dropout information (result) here and does not re-compute the convolution

Could you clarify what you mean here, maybe with a small code example? Note that we have to ensure that the implementation sticks with the specification of the original paper.

When we have no dropout, though, we should be able to make the same optimization as in #2122 though.

Also, what do you think of the fix for convolutional layers using the "groups" argument?

I wasn't aware of the groups argument in Conv2d. Quite possibly your solution is correct. Could you give a small example, so that we can build a unit test based on that? Also, could you explain why we only need it for DoRA?

[gslama12]

So the reasoning behind why the DoRA optimization is not possible when we use lora_dropout != 0 was, that the magnitude vector in the code snippet below will receive with pre-dropout information in form of the base layer result.

x = dropout(x)
result = result + self.lora_magnitude_vector[active_adapter](
                        x,
                        lora_A=lora_A,
                        lora_B=lora_B,
                        scaling=scaling,
                        base_layer=self.get_base_layer(),
                        base_layer_result=result
                    )

After looking at the DoRA paper, I can't figure out why this is an issue. If we see DoRA as a magnitude * direction, where LoRA is applied to the direction component, shouldn't we be able to apply LoRA to the direction component using lora_B(lora_A(dropout(x)))? Basically computing the DoRA result as: result_dora = (mag_norm_scale - 1) * base_layer_result + mag_norm_scale * lora_B(lora_A(dropout(x))) * scaling in dora.py.

Based on the paper, why would we need to compute the full-rank convolution again with a dropout(x) as the input (dora.py, ln 161)?

[BenjaminBossan]

Note that when we have LoRA+DoRA+dropout, we ensure that dropout is consistently applied to the LoRA part and the "base_result" part. If we use the result from the base layer directly (i.e. base_layer_result in your code), the LoRA-dropout is not applied to it, therefore, the result differs.

[gslama12]

I think I understand your point. But if we look at LoRA (e.g. ln 1120 in layer.py) we see, that we also don't apply the lora_dropout to the result here:

result = self.base_layer(x, *args, **kwargs)
.
.
.
if not self.use_dora[active_adapter]:
    result = result + lora_B(lora_A(dropout(x))) * scaling

So my question is wether the "base_result" part even needs the dropout for DoRA. And if we do need the dropout in the "base_result" part, why do we not need it for LoRA?

[BenjaminBossan]

Exactly. The result from the base model does not contain any LoRA-dropout applied to the x. However, for the DoRA-part, we need to ensure that the same dropout is applied to the x of base_layer_result as for the LoRA part. In the existing code, we apply dropout to x, then pass x to the DoRA layer, where x is passed to the convolution operation. Therefore, this requirement is met.

In your proposed code, we would instead use the result from the base layer which does not include x with dropout. Therefore, this is a different output and the DoRA calculation would no longer correspond to what is described in the paper.

Only if there is no dropout can we re-use the base result in the way you proposed, as is shown in #2122. (In addition, we also need to take care of potentially existing bias terms, but I left that out to simplify the discussion.)

You can also create a LoRA layer with DoRA and check that the outputs differ when dropout is applied between the old code and the suggested change (fix the seed to ensure that there is no randomness involved).

[gslama12]

Yes, I understand the reasoning and that my suggestion would produce a different output. I just could not find the reasoning for why the dropout needs to be applied like this in the DoRA paper. But I'll assume you're right.

The reason why I am questioning this is because we do not seem to use the dropout in the same way for LoRA. So my question is:

Why do we not apply the dropout to the result for the base layer for LoRA in the code example?

result = self.base_layer(x, *args, **kwargs) 
.
.
.
if not self.use_dora[active_adapter]:
    result = result + lora_B(lora_A(dropout(x))) * scaling   # why can we use the base_layer result without dropout here?