If true this is very nice incremental improvement. It looks like it doesn't meaningfully improve the capabilities of the model, but is cheaper to compute than RMSNorm (which essentially all current state of art LLMs use) which means faster/cheaper training.
RMSNorm is pretty insigificant in terms of the overall compute in a transformer though -- usually the reduction work can be fused with earlier or later operations.
Rmsnorm acts like a barrier. No compute on the next network layer can start before all compute in the previous layer is done.
Splitting networks across multiple GPU's, this means you must wait for the slowest node and the longest latency.
As soon as you can remove most of these barriers, compute over non-latency-guaranteed networks becomes more practical, as does non-homogeneous compute (ie. Mixing different GPU models).
Okay, I just tried this on my pet transformer training benchmark and the results are very disappointing; it converges much more slowly than just using RMSNorm.
It either needs some significant hyperparameter tuning (besides tweaking alpha, which doesn't seem to do much for me), or some fancier initialization (tried both pytorch default and orthogonal, no difference), or maybe my scalar optimizer doesn't work on it (I have a custom optimizer for scalars which speeds up convergence vs Adam, but for DyT layers it seems to be just as good as Adam), or maybe it only catches up after billions of tokens (which I don't have the budget to test for so long).
Slight update, more fancy initialization of DyT weights (instead of having them be ones) seems to help a lot in my case (although it's still not as good as just using RMSNorm). Do something like this on the very first training step (`x` is the input to the layer):
y = x.to(torch.float32)
y = y * torch.rsqrt(y.pow(2).mean(-1, keepdim=True) + 1e-6)
z = torch.tanh(self.alpha * x)
scale = (y / (z + 1e-6)).mean(dim = -2).flatten()
self.weight.detach().copy_(scale)
This basically tries to initialize the weights so that the output of DyT is closer to what RMSNorm would have outputted, and it seems to help.
It's a fully custom architecture, heavily inspired by the modded-nanogpt speedrun (https://github.com/KellerJordan/modded-nanogpt) but written fully from scratch and further tweaked/modified. I use it for experiments and as a testbed when developing my training harness (which I use for training other models too, and which receives all of my non-LLM-specific improvements like e.g. better than Adam optimizers, a custom GPU memory allocator, custom gradient accumulation that accumulates directly into the optimizers' state without using extra VRAM for gradient, etc.).
When using low precision formats like float8 you usually have to upscale the activations to BF16 before normalising. So the normalisation layers are proportionally using more compute when going to lower precision. Replacing these layers would help reduce the compute cost significantly.
Need to read the details, but removing the norm can be big. It’s always a pain to make sure that your network is normalized properly when trying new architectures. Likely there will still be other implications of the tanh, since the norm is sometimes solving a conditioning problem, but IMO more alternatives are welcome
Proper initialization of layers keeps gradient magnitudes from vanishing/exploding in deep networks. If you make sure the output of each layer has mean 0, std 1, the gradients will be reasonable as well, for example.
I recommend e.g. the og resnet paper and its follow-up from Kaiming He et al.
There essentially the point is that largest eigenvalue (spectral radius) needs to be around 1, meaning repeated applications of a linear transformation doesn’t cause increase or decrease of the activations.
Sure initialization helps, but are there also results about long term training dynamics? Even the paper you suggested had to use some sort of normalization to keep things stable
I think ResNet pretty much solved vanishing gradients. As for exploding gradients, that is typically with good parameter initialization and normalization. The paper in question proposes an alternative to normalization.
Good question. That was an issue with tanh as activation function, and before residual connections and normalization layers. Tanh as a normalization but with other activations and residual present apparently is ok.
Batch norm and others are important for faster convergence due to forcing the model to focus creating second and higher order nonlinearities, as a simple shift in mean/std is normalized out, and thus the gradient does not point in a direction that would only change those properties of the output distribution.
Is it just me or have they provided graphs of LNinput againt LNoutput when the tanh(a*x) is also followed by a weight and bias.
Surely you would want to compare the output of the LayerNorm without the weight and bias to get an impression on their similarity.
I guess it doesn't matter if the final result works, but I feel like looking at the bit that they are changing in isolation might provide a better insight as to what is happening.
Exactly, And that's what happens in LayerNorm too. So if figured the best base for comparison would have been to leave that bit out when looking at their difference or similarity, because obviously the bits that have the same implementation will be the same.
By incorporating DyT, Transformers without normalization can match or exceed the performance of their normalized counterparts, mostly without hyperparameter tuning.
If true this is very nice incremental improvement. It looks like it doesn't meaningfully improve the capabilities of the model, but is cheaper to compute than RMSNorm (which essentially all current state of art LLMs use) which means faster/cheaper training.
RMSNorm is pretty insigificant in terms of the overall compute in a transformer though -- usually the reduction work can be fused with earlier or later operations.
Rmsnorm acts like a barrier. No compute on the next network layer can start before all compute in the previous layer is done.
Splitting networks across multiple GPU's, this means you must wait for the slowest node and the longest latency.
As soon as you can remove most of these barriers, compute over non-latency-guaranteed networks becomes more practical, as does non-homogeneous compute (ie. Mixing different GPU models).
What are other barriers in transformers? Or is the normalization layer the primary one?
dot-product attention is the biggest barrier. This is why there are so many attempts to linearize it.
that fail... linearization is a bad idea. But plenty of other optimizations are done
The paper's Table 7 shows DyT reducing overall LLaMA 7B inference time by 7.8% and training time by 8.2%. That is not insignificant.
But LLM performance scales according to the log of compute, so yeah it’s pretty insignificant. I think we’ve reached a bit of a plateau.
Okay, I just tried this on my pet transformer training benchmark and the results are very disappointing; it converges much more slowly than just using RMSNorm.
It either needs some significant hyperparameter tuning (besides tweaking alpha, which doesn't seem to do much for me), or some fancier initialization (tried both pytorch default and orthogonal, no difference), or maybe my scalar optimizer doesn't work on it (I have a custom optimizer for scalars which speeds up convergence vs Adam, but for DyT layers it seems to be just as good as Adam), or maybe it only catches up after billions of tokens (which I don't have the budget to test for so long).
Slight update, more fancy initialization of DyT weights (instead of having them be ones) seems to help a lot in my case (although it's still not as good as just using RMSNorm). Do something like this on the very first training step (`x` is the input to the layer):
This basically tries to initialize the weights so that the output of DyT is closer to what RMSNorm would have outputted, and it seems to help.Which model are you training and on what dataset?
It's a fully custom architecture, heavily inspired by the modded-nanogpt speedrun (https://github.com/KellerJordan/modded-nanogpt) but written fully from scratch and further tweaked/modified. I use it for experiments and as a testbed when developing my training harness (which I use for training other models too, and which receives all of my non-LLM-specific improvements like e.g. better than Adam optimizers, a custom GPU memory allocator, custom gradient accumulation that accumulates directly into the optimizers' state without using extra VRAM for gradient, etc.).
For the dataset I just use FineWeb-Edu.
Wow, thank you for the link to the code - I haven't seen it before - it contains a ton of useful tricks. Lots to learn from there.
When using low precision formats like float8 you usually have to upscale the activations to BF16 before normalising. So the normalisation layers are proportionally using more compute when going to lower precision. Replacing these layers would help reduce the compute cost significantly.
Need to read the details, but removing the norm can be big. It’s always a pain to make sure that your network is normalized properly when trying new architectures. Likely there will still be other implications of the tanh, since the norm is sometimes solving a conditioning problem, but IMO more alternatives are welcome
And so vanishing gradients are not a thing anymore?
Proper initialization of layers keeps gradient magnitudes from vanishing/exploding in deep networks. If you make sure the output of each layer has mean 0, std 1, the gradients will be reasonable as well, for example.
I recommend e.g. the og resnet paper and its follow-up from Kaiming He et al.
For a modern take on RNNs, read https://arxiv.org/abs/2303.06349 by DeepMind.
There essentially the point is that largest eigenvalue (spectral radius) needs to be around 1, meaning repeated applications of a linear transformation doesn’t cause increase or decrease of the activations.
Sure initialization helps, but are there also results about long term training dynamics? Even the paper you suggested had to use some sort of normalization to keep things stable
I think ResNet pretty much solved vanishing gradients. As for exploding gradients, that is typically with good parameter initialization and normalization. The paper in question proposes an alternative to normalization.
Good question. That was an issue with tanh as activation function, and before residual connections and normalization layers. Tanh as a normalization but with other activations and residual present apparently is ok.
Proper initialization is more important.
Batch norm and others are important for faster convergence due to forcing the model to focus creating second and higher order nonlinearities, as a simple shift in mean/std is normalized out, and thus the gradient does not point in a direction that would only change those properties of the output distribution.
Transformers learn residuals, as you can see in the figure. y = x + f(x).
Is it just me or have they provided graphs of LNinput againt LNoutput when the tanh(a*x) is also followed by a weight and bias.
Surely you would want to compare the output of the LayerNorm without the weight and bias to get an impression on their similarity.
I guess it doesn't matter if the final result works, but I feel like looking at the bit that they are changing in isolation might provide a better insight as to what is happening.
From their implementation it looks like they’re calculating tanh and then applying a weight and bias
Exactly, And that's what happens in LayerNorm too. So if figured the best base for comparison would have been to leave that bit out when looking at their difference or similarity, because obviously the bits that have the same implementation will be the same.
What are the practical implications of this?
from the abstract
Sure, but why would one prefer tanh instead of normalization layers if they have the same accuracy?
I suppose normalization kernels have reductions in them, but how hard are reductions in 2025?
[flagged]
I'll give you a call when I finished building my tesla tower. This was also unnoticed by the engineering/science communities.
I don’t follow.
Don't advertise blatantly like this on HN please
It was relevant.