Add bnb.nn.StableEmbedding for quantized training #1770
Conversation
| if isinstance(fabric.strategy.precision, BitsandbytesPrecision): | ||
| optimizer = instantiate_bnb_optimizer(optimizer, model.parameters()) | ||
|
|
||
| from bitsandbytes.nn import StableEmbedding |
There was a problem hiding this comment.
If it doesn't cause any issues, later it should be moved to PTL/Fabric's BitsandbytesPrecision plugin.
In addition, it should be better from a memory usage perspective.
Now the model is materialized on the device, embeddings aren't quantized (only Linear layers are) and only after embeddings weights are replaced with quantized once.
If the code is in the PTL code, then embeddings will be quantized during materialization.
There was a problem hiding this comment.
@Andrei-Aksionov @lantiga I think the idea is not to quantize StableEmbedding during training due to precision issues, but I am not an expert in BitsandBytes tbh.
There was a problem hiding this comment.
But I think I maybe understand what you are saying: the modified model will be more inefficient during inference because quantization will not be applied to the StableEmbedding layer after training?
In this case, perhaps we should convert the StableEmbedding layer back into the regular Embedding layer before saving the model at the end of training. Let me investigate what impact that has on performance during inference.
Regarding integration into PTL/Fabric, that's a good point, but I am a bit concerned if we can/should make model architecture modifications like that?
There was a problem hiding this comment.
No, it was my mistake 😄.
I thought that it's a way to quantize embeddings 🙈.
But it still might be beneficial to upstream the code.
Right now, the model is initialized on a meta device and materialized on a target device via Fabric.
After the model is materialized the code creates StableEmbeddings and after replaces weights of the regular embeddings. So, at the same moment in time, we have two sets of weights: regular embeddings and stable ones.
If we move the code to Fabric, the replacement will happen during materialization.
Or you can try to replace the layer while the model is on meta device.
There was a problem hiding this comment.
So, at the same moment in time, we have two sets of weights: regular embeddings and stable ones.
Good point. We could/should explicitly delete and garbage-collect the redundant original weight embedding after initializing the StableEmbedding. Sure, it would be even more efficient to never have 2 sets of weight at the same time during this weight-swapping point, but I don't think this will create a memory bottleneck there because the forward and backward passes in the training afterwards will likely be much higher than having temporarily a second embedding layer in memory.
The one concern is though if we move the code to Fabric, it will apply the Fabric precision settings to StableEmbedding, but if I understand correctly, this is a special layer that should retain float32 bit precision. So we would have to make a special exception for that layer in Fabric.
There was a problem hiding this comment.
Correct. It's in a category of "nice to have".
If we move it to Fabric, it will just mean that whenever you apply BnB prevision plugin, this type of embedding, with the correct precision, will be used. Moving the code to Fabric doesn't require embeddings to be quantized :)
There was a problem hiding this comment.
Sounds good! And it might also be a good first Fabric contribution for me to familiarize myself with the code base.
The bnb.nn.StableEmbedding is used during finetuning to replace the standard Embedding layer. As discussed #1769
Fixes #1769