Keyerror: hook_point_in

[BoyuanChen99]

Following ./examples/loading_llamascope_saes.ipynb, and received the following error on the line below... Please instruct on how to fix. Thanks!

sae = SparseAutoEncoder.from_pretrained("fnlp/Llama3_1-8B-Base-L15R-8x")

Local path `fnlp/Llama3_1-8B-Base-L15R-8x` not found. Downloading from huggingface model hub.
Downloading Llama Scope SAEs.
Fetching 3 files: 100%|██████████| 3/3 [00:00<00:00, 37900.34it/s]
---------------------------------------------------------------------------
KeyError                                  Traceback (most recent call last)
Cell In[4], [line 1](vscode-notebook-cell:?execution_count=4&line=1)
----> [1](vscode-notebook-cell:?execution_count=4&line=1) sae = SparseAutoEncoder.from_pretrained("fnlp/Llama3_1-8B-Base-L15R-8x")

File ~/Desktop/Language-Model-SAEs/src/lm_saes/sae.py:585, in SparseAutoEncoder.from_pretrained(cls, pretrained_name_or_path, strict_loading, **kwargs)
    [583](https://file+.vscode-resource.vscode-cdn.net/home/bc3194/Desktop/Language-Model-SAEs/examples/~/Desktop/Language-Model-SAEs/src/lm_saes/sae.py:583) @classmethod
    [584](https://file+.vscode-resource.vscode-cdn.net/home/bc3194/Desktop/Language-Model-SAEs/examples/~/Desktop/Language-Model-SAEs/src/lm_saes/sae.py:584) def from_pretrained(cls, pretrained_name_or_path: str, strict_loading: bool = True, **kwargs):
--> [585](https://file+.vscode-resource.vscode-cdn.net/home/bc3194/Desktop/Language-Model-SAEs/examples/~/Desktop/Language-Model-SAEs/src/lm_saes/sae.py:585)     cfg = SAEConfig.from_pretrained(pretrained_name_or_path, strict_loading=strict_loading, **kwargs)
    [586](https://file+.vscode-resource.vscode-cdn.net/home/bc3194/Desktop/Language-Model-SAEs/examples/~/Desktop/Language-Model-SAEs/src/lm_saes/sae.py:586)     return cls.from_config(cfg)

File ~/Desktop/Language-Model-SAEs/src/lm_saes/config.py:84, in BaseSAEConfig.from_pretrained(cls, pretrained_name_or_path, strict_loading, **kwargs)
     [82](https://file+.vscode-resource.vscode-cdn.net/home/bc3194/Desktop/Language-Model-SAEs/examples/~/Desktop/Language-Model-SAEs/src/lm_saes/config.py:82) sae_config["sae_pretrained_name_or_path"] = pretrained_name_or_path
     [83](https://file+.vscode-resource.vscode-cdn.net/home/bc3194/Desktop/Language-Model-SAEs/examples/~/Desktop/Language-Model-SAEs/src/lm_saes/config.py:83) sae_config["strict_loading"] = strict_loading
---> [84](https://file+.vscode-resource.vscode-cdn.net/home/bc3194/Desktop/Language-Model-SAEs/examples/~/Desktop/Language-Model-SAEs/src/lm_saes/config.py:84) return cls.model_validate({**sae_config, **kwargs})

    [... skipping hidden 1 frame]

File ~/Desktop/Language-Model-SAEs/src/lm_saes/config.py:53, in BaseSAEConfig.<lambda>(validated_model)
     [51](https://file+.vscode-resource.vscode-cdn.net/home/bc3194/Desktop/Language-Model-SAEs/examples/~/Desktop/Language-Model-SAEs/src/lm_saes/config.py:51) sae_type: Literal["sae", "crosscoder", "mixcoder"]
     [52](https://file+.vscode-resource.vscode-cdn.net/home/bc3194/Desktop/Language-Model-SAEs/examples/~/Desktop/Language-Model-SAEs/src/lm_saes/config.py:52) hook_point_in: str
---> [53](https://file+.vscode-resource.vscode-cdn.net/home/bc3194/Desktop/Language-Model-SAEs/examples/~/Desktop/Language-Model-SAEs/src/lm_saes/config.py:53) hook_point_out: str = Field(default_factory=lambda validated_model: validated_model["hook_point_in"])
     [54](https://file+.vscode-resource.vscode-cdn.net/home/bc3194/Desktop/Language-Model-SAEs/examples/~/Desktop/Language-Model-SAEs/src/lm_saes/config.py:54) d_model: int
     [55](https://file+.vscode-resource.vscode-cdn.net/home/bc3194/Desktop/Language-Model-SAEs/examples/~/Desktop/Language-Model-SAEs/src/lm_saes/config.py:55) expansion_factor: int

KeyError: 'hook_point_in'

[Hzfinfdu]

Hi! Our codebase is constantly updating and we now realized that we are not capable to maintain a stable version for external use XD. Apologies for this! However you can use llamascope with SAELens. For example,

from sae_lens import SAE

sae, cfg_dict, sparsity = SAE.from_pretrained(
    release="llama_scope_lxr_32x",  # see other options in sae_lens/pretrained_saes.yaml
    sae_id="l0r_32x",  # won't always be a hook point
    device=device,
)

should work.

Please reach out if you have any further questions :)