File size: 3,029 Bytes
0784a51
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
### Quantized EleutherAI/gpt-j-6b with 8-bit weights

This is a version of EleutherAI's GPT-J with 6 billion parameters that is modified so you can generate **and fine-tune** the model in colab or equivalent desktop gpu (e.g. single 1080Ti).

The [original GPT-J](https://huggingface.co/EleutherAI/gpt-j-6B/tree/main) takes 22+ GB memory for float32 parameters. Even if you cast everything to 16-bit, it will still not fit onto most single-GPU setups short of A6000 and A100. You can inference it [on TPU](https://colab.research.google.com/github/kingoflolz/mesh-transformer-jax/blob/master/colab_demo.ipynb) or CPUs, but fine-tuning is way more expensive.

Here, we apply several techniques to make GPT-J usable and fine-tunable on a single GPU with ~11 GB memory:
- large weight tensors are quantized using dynamic 8-bit quantization and de-quantized just-in-time for multiplication
- using gradient checkpoints to store one only activation per layer: using dramatically less memory at the cost of 30% slower training
- scalable fine-tuning with [LoRA](https://arxiv.org/abs/2106.09685) and [8-bit Adam](https://arxiv.org/abs/2110.02861)

In other words, all of the large weight-matrices are frozen in 8-bit, and you only train small adapters and optionally 1d tensors (layernorm scales, biases).

![img](https://i.imgur.com/n4XXo1x.png)


__What about model quality?__ TODO compare on wikitext TODO brag about visual quality


__What about performance?__ Both checkpointing and de-quantization has some overhead, but it's surprisingly manageable. Depending on GPU and batch size, the quantized model is 1-10% slower than the original model on top of using gradient checkpoints (which is 30% overhead). In short, this is because block-wise quantization from bitsandbytes is really fast on GPU.


### How should I fine-tune the model?

We recommend starting with the original hyperparameters from [the LoRA paper](https://arxiv.org/pdf/2106.09685.pdf).
On top of that, there is one more trick to consider: the overhead from de-quantizing weights does not depend on batch size.
As a result, the larger batch size you can fit, the more efficient you will train.


### Where can I train for free?

You can train fine in colab, but if you get a k80, it's probably best to switch to other free gpu providers: [kaggle](https://towardsdatascience.com/amazon-sagemaker-studio-lab-a-great-alternative-to-google-colab-7194de6ef69a), [aws sagemaker](https://towardsdatascience.com/amazon-sagemaker-studio-lab-a-great-alternative-to-google-colab-7194de6ef69a) or [paperspace](https://docs.paperspace.com/gradient/more/instance-types/free-instances)


### Can I use this technique with other models?

The model was converted using [this notebook](https://nbviewer.org/urls/huggingface.co/hivemind/gpt-j-6B-8bit/raw/main/convert-gpt-j.ipynb). It can be adapted to work with other model types. However, please bear in mind that some models replace Linear and Embedding with custom alternatives that require their own BNBWhateverWithAdapters.