sae: A Python repository from aryopg

Introduction

This library trains k-sparse autoencoders (SAEs) on the residual stream activations of HuggingFace language models, roughly following the recipe detailed in Scaling and evaluating sparse autoencoders (Gao et al. 2024).

This is a lean, simple library with few configuration options. Unlike most other SAE libraries (e.g. SAELens), it does not cache activations on disk, but rather computes them on-the-fly. This allows us to scale to very large models and datasets with zero storage overhead, but has the downside that trying different hyperparameters for the same model and dataset will be slower than if we cached activations (since activations will be re-computed). We may add caching as an option in the future.

Following Gao et al., we use a TopK activation function which directly enforces a desired level of sparsity in the activations. This is in contrast to other libraries which use an L1 penalty in the loss function. We believe TopK is a Pareto improvement over the L1 approach, and hence do not plan on supporting it.

Loading pretrained SAEs

To load a pretrained SAE from the HuggingFace Hub, you can use the Sae.load_from_hub method as follows:

from sae import Sae

sae = Sae.load_from_hub("EleutherAI/sae-llama-3-8b-32x", hookpoint="layers.10")

This will load the SAE for residual stream layer 10 of Llama 3 8B, which was trained with an expansion factor of 32. You can also load the SAEs for all layers at once using Sae.load_many:

saes = Sae.load_many("EleutherAI/sae-llama-3-8b-32x")
saes["layers.10"]

The dictionary returned by load_many is guaranteed to be naturally sorted by the name of the hook point. For the common case where the hook points are named embed_tokens, layers.0, ..., layers.n, this means that the SAEs will be sorted by layer number. We can then gather the SAE activations for a model forward pass as follows:

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-8B")
inputs = tokenizer("Hello, world!", return_tensors="pt")

with torch.inference_mode():
    model = AutoModelForCausalLM.from_pretrained("meta-llama/Meta-Llama-3-8B")
    outputs = model(**inputs, output_hidden_states=True)

    latent_acts = []
    for sae, hidden_state in zip(saes.values(), outputs.hidden_states):
        latent_acts.append(sae.encode(hidden_state))

# Do stuff with the latent activations

Training SAEs

To train SAEs from the command line, you can use the following command:

python -m sae EleutherAI/pythia-160m togethercomputer/RedPajama-Data-1T-Sample

The CLI supports all of the config options provided by the TrainConfig class. You can see them by running python -m sae --help.

Programmatic usage is simple. Here is an example:

import torch
from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer

from sae import SaeConfig, SaeTrainer, TrainConfig
from sae.data import chunk_and_tokenize

MODEL = "EleutherAI/pythia-160m"
dataset = load_dataset(
    "togethercomputer/RedPajama-Data-1T-Sample",
    split="train",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(MODEL)
tokenized = chunk_and_tokenize(dataset, tokenizer)


gpt = AutoModelForCausalLM.from_pretrained(
    MODEL,
    device_map={"": "cuda"},
    torch_dtype=torch.bfloat16,
)

cfg = TrainConfig(
    SaeConfig(gpt.config.hidden_size), batch_size=16
)
trainer = SaeTrainer(cfg, tokenized, gpt)

trainer.fit()

Custom hookpoints

By default, the SAEs are trained on the residual stream activations of the model. However, you can also train SAEs on the activations of any other submodule(s) by specifying custom hookpoint patterns. These patterns are like standard PyTorch module names (e.g. h.0.ln_1) but also allow Unix pattern matching syntax, including wildcards and character sets. For example, to train SAEs on the output of every attention module and the inner activations of every MLP in GPT-2, you can use the following code:

python -m sae gpt2 togethercomputer/RedPajama-Data-1T-Sample --hookpoints "h.*.attn" "h.*.mlp.act"

To restrict to the first three layers:

python -m sae gpt2 togethercomputer/RedPajama-Data-1T-Sample --hookpoints "h.[012].attn" "h.[012].mlp.act"

We currently don't support fine-grained manual control over the learning rate, number of latents, or other hyperparameters on a hookpoint-by-hookpoint basis. By default, the expansion_ratio option is used to select the appropriate number of latents for each hookpoint based on the width of that hookpoint's output. The default learning rate for each hookpoint is then set using an inverse square root scaling law based on the number of latents. If you manually set the number of latents or the learning rate, it will be applied to all hookpoints.

Distributed training

We support distributed training via PyTorch's torchrun command. By default we use the Distributed Data Parallel method, which means that the weights of each SAE are replicated on every GPU.

torchrun --nproc_per_node gpu -m sae meta-llama/Meta-Llama-3-8B --batch_size 1 --layers 16 24 --k 192 --grad_acc_steps 8 --ctx_len 2048

This is simple, but very memory inefficient. If you want to train SAEs for many layers of a model, we recommend using the --distribute_modules flag, which allocates the SAEs for different layers to different GPUs. Currently, we require that the number of GPUs evenly divides the number of layers you're training SAEs for.

torchrun --nproc_per_node gpu -m sae meta-llama/Meta-Llama-3-8B --distribute_modules --batch_size 1 --layer_stride 2 --grad_acc_steps 8 --ctx_len 2048 --k 192 --load_in_8bit --micro_acc_steps 2

The above command trains an SAE for every even layer of Llama 3 8B, using all available GPUs. It accumulates gradients over 8 minibatches, and splits each minibatch into 2 microbatches before feeding them into the SAE encoder, thus saving a lot of memory. It also loads the model in 8-bit precision using bitsandbytes. This command requires no more than 48GB of memory per GPU on an 8 GPU node.

TODO

There are several features that we'd like to add in the near future:

Finetuning pretrained SAEs
Support for caching activations
Evaluate SAEs with KL divergence when grafted into the model