sae-lens 6.24.1__py3-none-any.whl → 6.25.1__py3-none-any.whl

sae_lens/__init__.py

@@ -1,5 +1,5 @@
 # ruff: noqa: E402
-__version__ = "6.24.1"
+__version__ = "6.25.1"
 
 import logging
 

sae_lens/cache_activations_runner.py

@@ -14,9 +14,9 @@ from transformer_lens.HookedTransformer import HookedRootModule
 
 from sae_lens import logger
 from sae_lens.config import CacheActivationsRunnerConfig
-from sae_lens.constants import DTYPE_MAP
 from sae_lens.load_model import load_model
 from sae_lens.training.activations_store import ActivationsStore
+from sae_lens.util import str_to_dtype
 
 
 def _mk_activations_store(
@@ -97,7 +97,7 @@ class CacheActivationsRunner:
         bytes_per_token = (
             self.cfg.d_in * self.cfg.dtype.itemsize
             if isinstance(self.cfg.dtype, torch.dtype)
-            else DTYPE_MAP[self.cfg.dtype].itemsize
+            else str_to_dtype(self.cfg.dtype).itemsize
         )
         total_training_tokens = self.cfg.n_seq_in_dataset * self.context_size
         total_disk_space_gb = total_training_tokens * bytes_per_token / 10**9

sae_lens/config.py

@@ -17,9 +17,9 @@ from datasets import (
 )
 
 from sae_lens import __version__, logger
-from sae_lens.constants import DTYPE_MAP
 from sae_lens.registry import get_sae_training_class
 from sae_lens.saes.sae import TrainingSAEConfig
+from sae_lens.util import str_to_dtype
 
 if TYPE_CHECKING:
     pass
@@ -563,7 +563,7 @@ class CacheActivationsRunnerConfig:
 
     @property
     def bytes_per_token(self) -> int:
-        return self.d_in * DTYPE_MAP[self.dtype].itemsize
+        return self.d_in * str_to_dtype(self.dtype).itemsize
 
     @property
     def n_tokens_in_buffer(self) -> int:

sae_lens/constants.py

@@ -11,6 +11,14 @@ DTYPE_MAP = {
     "torch.bfloat16": torch.bfloat16,
 }
 
+# Reverse mapping from torch.dtype to canonical string format
+DTYPE_TO_STR = {
+    torch.float32: "float32",
+    torch.float64: "float64",
+    torch.float16: "float16",
+    torch.bfloat16: "bfloat16",
+}
+
 
 SPARSITY_FILENAME = "sparsity.safetensors"
 SAE_WEIGHTS_FILENAME = "sae_weights.safetensors"

sae_lens/loading/pretrained_sae_loaders.py

@@ -12,11 +12,9 @@ from huggingface_hub import hf_hub_download, hf_hub_url
 from huggingface_hub.utils import EntryNotFoundError, build_hf_headers
 from packaging.version import Version
 from safetensors import safe_open
-from safetensors.torch import load_file
 
 from sae_lens import logger
 from sae_lens.constants import (
-    DTYPE_MAP,
     SAE_CFG_FILENAME,
     SAE_WEIGHTS_FILENAME,
     SPARSIFY_WEIGHTS_FILENAME,
@@ -28,7 +26,7 @@ from sae_lens.loading.pretrained_saes_directory import (
     get_repo_id_and_folder_name,
 )
 from sae_lens.registry import get_sae_class
-from sae_lens.util import filter_valid_dataclass_fields
+from sae_lens.util import filter_valid_dataclass_fields, str_to_dtype
 
 LLM_METADATA_KEYS = {
     "model_name",
@@ -51,6 +49,21 @@ LLM_METADATA_KEYS = {
 }
 
 
+def load_safetensors_weights(
+    path: str | Path, device: str = "cpu", dtype: torch.dtype | str | None = None
+) -> dict[str, torch.Tensor]:
+    """Load safetensors weights and optionally convert to a different dtype"""
+    loaded_weights = {}
+    dtype = str_to_dtype(dtype) if isinstance(dtype, str) else dtype
+    with safe_open(path, framework="pt", device=device) as f:
+        for k in f.keys():  # noqa: SIM118
+            weight = f.get_tensor(k)
+            if dtype is not None:
+                weight = weight.to(dtype=dtype)
+            loaded_weights[k] = weight
+    return loaded_weights
+
+
 # loaders take in a release, sae_id, device, and whether to force download, and returns a tuple of config, state_dict, and log sparsity
 class PretrainedSaeHuggingfaceLoader(Protocol):
     def __call__(
@@ -341,7 +354,7 @@ def read_sae_components_from_disk(
     Given a loaded dictionary and a path to a weight file, load the weights and return the state_dict.
     """
     if dtype is None:
-        dtype = DTYPE_MAP[cfg_dict["dtype"]]
+        dtype = str_to_dtype(cfg_dict["dtype"])
 
     state_dict = {}
     with safe_open(weight_path, framework="pt", device=device) as f:  # type: ignore
@@ -682,15 +695,6 @@ def gemma_3_sae_huggingface_loader(
         cfg_overrides,
     )
 
-    # replace folder name of 65k with 64k
-    # TODO: remove this workaround once weights are fixed
-    if "270m-pt" in repo_id:
-        if "65k" in folder_name:
-            folder_name = folder_name.replace("65k", "64k")
-        # replace folder name of 262k with 250k
-        if "262k" in folder_name:
-            folder_name = folder_name.replace("262k", "250k")
-
     params_file = "params.safetensors"
     if "clt" in folder_name:
         params_file = folder_name.split("/")[-1] + ".safetensors"
@@ -704,7 +708,9 @@ def gemma_3_sae_huggingface_loader(
         force_download=force_download,
     )
 
-    raw_state_dict = load_file(sae_path, device=device)
+    raw_state_dict = load_safetensors_weights(
+        sae_path, device=device, dtype=cfg_dict.get("dtype")
+    )
 
     with torch.no_grad():
         w_dec = raw_state_dict["w_dec"]
@@ -791,11 +797,13 @@ def get_goodfire_huggingface_loader(
     )
     raw_state_dict = torch.load(sae_path, map_location=device)
 
+    target_dtype = str_to_dtype(cfg_dict.get("dtype", "float32"))
+
     state_dict = {
-        "W_enc": raw_state_dict["encoder_linear.weight"].T,
-        "W_dec": raw_state_dict["decoder_linear.weight"].T,
-        "b_enc": raw_state_dict["encoder_linear.bias"],
-        "b_dec": raw_state_dict["decoder_linear.bias"],
+        "W_enc": raw_state_dict["encoder_linear.weight"].T.to(dtype=target_dtype),
+        "W_dec": raw_state_dict["decoder_linear.weight"].T.to(dtype=target_dtype),
+        "b_enc": raw_state_dict["encoder_linear.bias"].to(dtype=target_dtype),
+        "b_dec": raw_state_dict["decoder_linear.bias"].to(dtype=target_dtype),
     }
 
     return cfg_dict, state_dict, None
@@ -898,26 +906,19 @@ def llama_scope_sae_huggingface_loader(
         force_download=force_download,
     )
 
-    # Load the weights using load_file instead of safe_open
-    state_dict_loaded = load_file(sae_path, device=device)
+    state_dict_loaded = load_safetensors_weights(
+        sae_path, device=device, dtype=cfg_dict.get("dtype")
+    )
 
     # Convert and organize the weights
     state_dict = {
-        "W_enc": state_dict_loaded["encoder.weight"]
-        .to(dtype=DTYPE_MAP[cfg_dict["dtype"]])
-        .T,
-        "W_dec": state_dict_loaded["decoder.weight"]
-        .to(dtype=DTYPE_MAP[cfg_dict["dtype"]])
-        .T,
-        "b_enc": state_dict_loaded["encoder.bias"].to(
-            dtype=DTYPE_MAP[cfg_dict["dtype"]]
-        ),
-        "b_dec": state_dict_loaded["decoder.bias"].to(
-            dtype=DTYPE_MAP[cfg_dict["dtype"]]
-        ),
+        "W_enc": state_dict_loaded["encoder.weight"].T,
+        "W_dec": state_dict_loaded["decoder.weight"].T,
+        "b_enc": state_dict_loaded["encoder.bias"],
+        "b_dec": state_dict_loaded["decoder.bias"],
         "threshold": torch.ones(
             cfg_dict["d_sae"],
-            dtype=DTYPE_MAP[cfg_dict["dtype"]],
+            dtype=str_to_dtype(cfg_dict["dtype"]),
             device=cfg_dict["device"],
         )
         * cfg_dict["jump_relu_threshold"],
@@ -1228,26 +1229,17 @@ def llama_scope_r1_distill_sae_huggingface_loader(
         force_download=force_download,
     )
 
-    # Load the weights using load_file instead of safe_open
-    state_dict_loaded = load_file(sae_path, device=device)
+    state_dict_loaded = load_safetensors_weights(
+        sae_path, device=device, dtype=cfg_dict.get("dtype")
+    )
 
     # Convert and organize the weights
     state_dict = {
-        "W_enc": state_dict_loaded["encoder.weight"]
-        .to(dtype=DTYPE_MAP[cfg_dict["dtype"]])
-        .T,
-        "W_dec": state_dict_loaded["decoder.weight"]
-        .to(dtype=DTYPE_MAP[cfg_dict["dtype"]])
-        .T,
-        "b_enc": state_dict_loaded["encoder.bias"].to(
-            dtype=DTYPE_MAP[cfg_dict["dtype"]]
-        ),
-        "b_dec": state_dict_loaded["decoder.bias"].to(
-            dtype=DTYPE_MAP[cfg_dict["dtype"]]
-        ),
-        "threshold": state_dict_loaded["log_jumprelu_threshold"]
-        .to(dtype=DTYPE_MAP[cfg_dict["dtype"]])
-        .exp(),
+        "W_enc": state_dict_loaded["encoder.weight"].T,
+        "W_dec": state_dict_loaded["decoder.weight"].T,
+        "b_enc": state_dict_loaded["encoder.bias"],
+        "b_dec": state_dict_loaded["decoder.bias"],
+        "threshold": state_dict_loaded["log_jumprelu_threshold"].exp(),
     }
 
     # No sparsity tensor for Llama Scope SAEs
@@ -1367,34 +1359,34 @@ def sparsify_disk_loader(
     cfg_dict = get_sparsify_config_from_disk(path, device, cfg_overrides)
 
     weight_path = Path(path) / SPARSIFY_WEIGHTS_FILENAME
-    state_dict_loaded = load_file(weight_path, device=device)
-
-    dtype = DTYPE_MAP[cfg_dict["dtype"]]
+    state_dict_loaded = load_safetensors_weights(
+        weight_path, device=device, dtype=cfg_dict.get("dtype")
+    )
 
     W_enc = (
         state_dict_loaded["W_enc"]
         if "W_enc" in state_dict_loaded
         else state_dict_loaded["encoder.weight"].T
-    ).to(dtype)
+    )
 
     if "W_dec" in state_dict_loaded:
-        W_dec = state_dict_loaded["W_dec"].T.to(dtype)
+        W_dec = state_dict_loaded["W_dec"].T
     else:
-        W_dec = state_dict_loaded["decoder.weight"].T.to(dtype)
+        W_dec = state_dict_loaded["decoder.weight"].T
 
     if "b_enc" in state_dict_loaded:
-        b_enc = state_dict_loaded["b_enc"].to(dtype)
+        b_enc = state_dict_loaded["b_enc"]
     elif "encoder.bias" in state_dict_loaded:
-        b_enc = state_dict_loaded["encoder.bias"].to(dtype)
+        b_enc = state_dict_loaded["encoder.bias"]
     else:
-        b_enc = torch.zeros(cfg_dict["d_sae"], dtype=dtype, device=device)
+        b_enc = torch.zeros(cfg_dict["d_sae"], dtype=W_dec.dtype, device=device)
 
     if "b_dec" in state_dict_loaded:
-        b_dec = state_dict_loaded["b_dec"].to(dtype)
+        b_dec = state_dict_loaded["b_dec"]
     elif "decoder.bias" in state_dict_loaded:
-        b_dec = state_dict_loaded["decoder.bias"].to(dtype)
+        b_dec = state_dict_loaded["decoder.bias"]
     else:
-        b_dec = torch.zeros(cfg_dict["d_in"], dtype=dtype, device=device)
+        b_dec = torch.zeros(cfg_dict["d_in"], dtype=W_dec.dtype, device=device)
 
     state_dict = {"W_enc": W_enc, "b_enc": b_enc, "W_dec": W_dec, "b_dec": b_dec}
     return cfg_dict, state_dict
@@ -1625,7 +1617,9 @@ def mwhanna_transcoder_huggingface_loader(
     )
 
     # Load weights from safetensors
-    state_dict = load_file(file_path, device=device)
+    state_dict = load_safetensors_weights(
+        file_path, device=device, dtype=cfg_dict.get("dtype")
+    )
     state_dict["W_enc"] = state_dict["W_enc"].T
 
     return cfg_dict, state_dict, None
@@ -1709,8 +1703,12 @@ def mntss_clt_layer_huggingface_loader(
         force_download=force_download,
     )
 
-    encoder_state_dict = load_file(encoder_path, device=device)
-    decoder_state_dict = load_file(decoder_path, device=device)
+    encoder_state_dict = load_safetensors_weights(
+        encoder_path, device=device, dtype=cfg_dict.get("dtype")
+    )
+    decoder_state_dict = load_safetensors_weights(
+        decoder_path, device=device, dtype=cfg_dict.get("dtype")
+    )
 
     with torch.no_grad():
         state_dict = {
@@ -1853,7 +1851,9 @@ def temporal_sae_huggingface_loader(
     )
 
     # Load checkpoint from safetensors
-    state_dict_raw = load_file(ckpt_path, device=device)
+    state_dict_raw = load_safetensors_weights(
+        ckpt_path, device=device, dtype=cfg_dict.get("dtype")
+    )
 
     # Convert to SAELens naming convention
     # TemporalSAE uses: D (decoder), E (encoder), b (bias), attn_layers.*

sae_lens/pretrained_saes.yaml

@@ -10197,6 +10197,7 @@ gemma-scope-2-27b-it-res:
   - id: layer_16_width_16k_l0_medium
     path: resid_post/layer_16_width_16k_l0_medium
     l0: 53
+    neuronpedia: gemma-3-27b-it/16-gemmascope-2-res-16k
   - id: layer_16_width_16k_l0_small
     path: resid_post/layer_16_width_16k_l0_small
     l0: 17
@@ -10206,6 +10207,7 @@ gemma-scope-2-27b-it-res:
   - id: layer_16_width_1m_l0_medium
     path: resid_post/layer_16_width_1m_l0_medium
     l0: 53
+    neuronpedia: gemma-3-27b-it/16-gemmascope-2-res-1m
   - id: layer_16_width_1m_l0_small
     path: resid_post/layer_16_width_1m_l0_small
     l0: 17
@@ -10215,6 +10217,7 @@ gemma-scope-2-27b-it-res:
   - id: layer_16_width_262k_l0_medium
     path: resid_post/layer_16_width_262k_l0_medium
     l0: 53
+    neuronpedia: gemma-3-27b-it/16-gemmascope-2-res-262k
   - id: layer_16_width_262k_l0_medium_seed_1
     path: resid_post/layer_16_width_262k_l0_medium_seed_1
     l0: 53
@@ -10227,6 +10230,7 @@ gemma-scope-2-27b-it-res:
   - id: layer_16_width_65k_l0_medium
     path: resid_post/layer_16_width_65k_l0_medium
     l0: 53
+    neuronpedia: gemma-3-27b-it/16-gemmascope-2-res-65k
   - id: layer_16_width_65k_l0_small
     path: resid_post/layer_16_width_65k_l0_small
     l0: 17
@@ -10236,6 +10240,7 @@ gemma-scope-2-27b-it-res:
   - id: layer_31_width_16k_l0_medium
     path: resid_post/layer_31_width_16k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-27b-it/31-gemmascope-2-res-16k
   - id: layer_31_width_16k_l0_small
     path: resid_post/layer_31_width_16k_l0_small
     l0: 20
@@ -10245,6 +10250,7 @@ gemma-scope-2-27b-it-res:
   - id: layer_31_width_1m_l0_medium
     path: resid_post/layer_31_width_1m_l0_medium
     l0: 60
+    neuronpedia: gemma-3-27b-it/31-gemmascope-2-res-1m
   - id: layer_31_width_1m_l0_small
     path: resid_post/layer_31_width_1m_l0_small
     l0: 20
@@ -10254,6 +10260,7 @@ gemma-scope-2-27b-it-res:
   - id: layer_31_width_262k_l0_medium
     path: resid_post/layer_31_width_262k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-27b-it/31-gemmascope-2-res-262k
   - id: layer_31_width_262k_l0_medium_seed_1
     path: resid_post/layer_31_width_262k_l0_medium_seed_1
     l0: 60
@@ -10266,6 +10273,7 @@ gemma-scope-2-27b-it-res:
   - id: layer_31_width_65k_l0_medium
     path: resid_post/layer_31_width_65k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-27b-it/31-gemmascope-2-res-65k
   - id: layer_31_width_65k_l0_small
     path: resid_post/layer_31_width_65k_l0_small
     l0: 20
@@ -10275,6 +10283,7 @@ gemma-scope-2-27b-it-res:
   - id: layer_40_width_16k_l0_medium
     path: resid_post/layer_40_width_16k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-27b-it/40-gemmascope-2-res-16k
   - id: layer_40_width_16k_l0_small
     path: resid_post/layer_40_width_16k_l0_small
     l0: 20
@@ -10284,6 +10293,7 @@ gemma-scope-2-27b-it-res:
   - id: layer_40_width_1m_l0_medium
     path: resid_post/layer_40_width_1m_l0_medium
     l0: 60
+    neuronpedia: gemma-3-27b-it/40-gemmascope-2-res-1m
   - id: layer_40_width_1m_l0_small
     path: resid_post/layer_40_width_1m_l0_small
     l0: 20
@@ -10293,6 +10303,7 @@ gemma-scope-2-27b-it-res:
   - id: layer_40_width_262k_l0_medium
     path: resid_post/layer_40_width_262k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-27b-it/40-gemmascope-2-res-262k
   - id: layer_40_width_262k_l0_medium_seed_1
     path: resid_post/layer_40_width_262k_l0_medium_seed_1
     l0: 60
@@ -10305,6 +10316,7 @@ gemma-scope-2-27b-it-res:
   - id: layer_40_width_65k_l0_medium
     path: resid_post/layer_40_width_65k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-27b-it/40-gemmascope-2-res-65k
   - id: layer_40_width_65k_l0_small
     path: resid_post/layer_40_width_65k_l0_small
     l0: 20
@@ -10314,6 +10326,7 @@ gemma-scope-2-27b-it-res:
   - id: layer_53_width_16k_l0_medium
     path: resid_post/layer_53_width_16k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-27b-it/53-gemmascope-2-res-16k
   - id: layer_53_width_16k_l0_small
     path: resid_post/layer_53_width_16k_l0_small
     l0: 20
@@ -10323,6 +10336,7 @@ gemma-scope-2-27b-it-res:
   - id: layer_53_width_1m_l0_medium
     path: resid_post/layer_53_width_1m_l0_medium
     l0: 60
+    neuronpedia: gemma-3-27b-it/53-gemmascope-2-res-1m
   - id: layer_53_width_1m_l0_small
     path: resid_post/layer_53_width_1m_l0_small
     l0: 20
@@ -10332,6 +10346,7 @@ gemma-scope-2-27b-it-res:
   - id: layer_53_width_262k_l0_medium
     path: resid_post/layer_53_width_262k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-27b-it/53-gemmascope-2-res-262k
   - id: layer_53_width_262k_l0_medium_seed_1
     path: resid_post/layer_53_width_262k_l0_medium_seed_1
     l0: 60
@@ -10344,6 +10359,7 @@ gemma-scope-2-27b-it-res:
   - id: layer_53_width_65k_l0_medium
     path: resid_post/layer_53_width_65k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-27b-it/53-gemmascope-2-res-65k
   - id: layer_53_width_65k_l0_small
     path: resid_post/layer_53_width_65k_l0_small
     l0: 20

sae_lens/saes/sae.py

@@ -27,11 +27,10 @@ from typing_extensions import deprecated, overload, override
 
 from sae_lens import __version__
 from sae_lens.constants import (
-    DTYPE_MAP,
     SAE_CFG_FILENAME,
     SAE_WEIGHTS_FILENAME,
 )
-from sae_lens.util import filter_valid_dataclass_fields
+from sae_lens.util import dtype_to_str, filter_valid_dataclass_fields, str_to_dtype
 
 if TYPE_CHECKING:
     from sae_lens.config import LanguageModelSAERunnerConfig
@@ -253,7 +252,7 @@ class SAE(HookedRootModule, Generic[T_SAE_CONFIG], ABC):
                 stacklevel=1,
             )
 
-        self.dtype = DTYPE_MAP[cfg.dtype]
+        self.dtype = str_to_dtype(cfg.dtype)
         self.device = torch.device(cfg.device)
         self.use_error_term = use_error_term
 
@@ -437,8 +436,8 @@ class SAE(HookedRootModule, Generic[T_SAE_CONFIG], ABC):
 
         # Update dtype in config if provided
         if dtype_arg is not None:
-            # Update the cfg.dtype
-            self.cfg.dtype = str(dtype_arg)
+            # Update the cfg.dtype (use canonical short form like "float32")
+            self.cfg.dtype = dtype_to_str(dtype_arg)
 
             # Update the dtype property
             self.dtype = dtype_arg
@@ -534,6 +533,15 @@ class SAE(HookedRootModule, Generic[T_SAE_CONFIG], ABC):
         dtype: str | None = None,
         converter: PretrainedSaeDiskLoader = sae_lens_disk_loader,
     ) -> T_SAE:
+        """
+        Load a SAE from disk.
+
+        Args:
+            path: The path to the SAE weights and config.
+            device: The device to load the SAE on, defaults to "cpu".
+            dtype: The dtype to load the SAE on, defaults to None. If None, the dtype will be inferred from the SAE config.
+            converter: The converter to use to load the SAE, defaults to sae_lens_disk_loader.
+        """
         overrides = {"dtype": dtype} if dtype is not None else None
         cfg_dict, state_dict = converter(path, device, cfg_overrides=overrides)
         cfg_dict = handle_config_defaulting(cfg_dict)
@@ -542,10 +550,17 @@ class SAE(HookedRootModule, Generic[T_SAE_CONFIG], ABC):
         )
         sae_cfg = sae_config_cls.from_dict(cfg_dict)
         sae_cls = cls.get_sae_class_for_architecture(sae_cfg.architecture())
+        # hack to avoid using double memory when loading the SAE.
+        # first put the SAE on the meta device, then load the weights.
+        device = sae_cfg.device
+        sae_cfg.device = "meta"
         sae = sae_cls(sae_cfg)
+        sae.cfg.device = device
         sae.process_state_dict_for_loading(state_dict)
-        sae.load_state_dict(state_dict)
-        return sae
+        sae.load_state_dict(state_dict, assign=True)
+        # the loaders should already handle the dtype / device conversion
+        # but this is a fallback to guarantee the SAE is on the correct device and dtype
+        return sae.to(dtype=str_to_dtype(sae_cfg.dtype), device=device)
 
     @classmethod
     def from_pretrained(
@@ -553,6 +568,7 @@ class SAE(HookedRootModule, Generic[T_SAE_CONFIG], ABC):
         release: str,
         sae_id: str,
         device: str = "cpu",
+        dtype: str = "float32",
         force_download: bool = False,
         converter: PretrainedSaeHuggingfaceLoader | None = None,
     ) -> T_SAE:
@@ -562,10 +578,18 @@ class SAE(HookedRootModule, Generic[T_SAE_CONFIG], ABC):
         Args:
             release: The release name. This will be mapped to a huggingface repo id based on the pretrained_saes.yaml file.
             id: The id of the SAE to load. This will be mapped to a path in the huggingface repo.
-            device: The device to load the SAE on.
+            device: The device to load the SAE on, defaults to "cpu".
+            dtype: The dtype to load the SAE on, defaults to "float32".
+            force_download: Whether to force download the SAE weights and config, defaults to False.
+            converter: The converter to use to load the SAE, defaults to None. If None, the converter will be inferred from the release.
         """
         return cls.from_pretrained_with_cfg_and_sparsity(
-            release, sae_id, device, force_download, converter=converter
+            release,
+            sae_id,
+            device,
+            force_download=force_download,
+            dtype=dtype,
+            converter=converter,
         )[0]
 
     @classmethod
@@ -574,6 +598,7 @@ class SAE(HookedRootModule, Generic[T_SAE_CONFIG], ABC):
         release: str,
         sae_id: str,
         device: str = "cpu",
+        dtype: str = "float32",
         force_download: bool = False,
         converter: PretrainedSaeHuggingfaceLoader | None = None,
     ) -> tuple[T_SAE, dict[str, Any], torch.Tensor | None]:
@@ -584,7 +609,10 @@ class SAE(HookedRootModule, Generic[T_SAE_CONFIG], ABC):
         Args:
             release: The release name. This will be mapped to a huggingface repo id based on the pretrained_saes.yaml file.
             id: The id of the SAE to load. This will be mapped to a path in the huggingface repo.
-            device: The device to load the SAE on.
+            device: The device to load the SAE on, defaults to "cpu".
+            dtype: The dtype to load the SAE on, defaults to "float32".
+            force_download: Whether to force download the SAE weights and config, defaults to False.
+            converter: The converter to use to load the SAE, defaults to None. If None, the converter will be inferred from the release.
         """
 
         # get sae directory
@@ -634,6 +662,7 @@ class SAE(HookedRootModule, Generic[T_SAE_CONFIG], ABC):
         repo_id, folder_name = get_repo_id_and_folder_name(release, sae_id)
         config_overrides = get_config_overrides(release, sae_id)
         config_overrides["device"] = device
+        config_overrides["dtype"] = dtype
 
         # Load config and weights
         cfg_dict, state_dict, log_sparsities = conversion_loader(
@@ -651,9 +680,14 @@ class SAE(HookedRootModule, Generic[T_SAE_CONFIG], ABC):
         )
         sae_cfg = sae_config_cls.from_dict(cfg_dict)
         sae_cls = cls.get_sae_class_for_architecture(sae_cfg.architecture())
+        # hack to avoid using double memory when loading the SAE.
+        # first put the SAE on the meta device, then load the weights.
+        device = sae_cfg.device
+        sae_cfg.device = "meta"
         sae = sae_cls(sae_cfg)
+        sae.cfg.device = device
         sae.process_state_dict_for_loading(state_dict)
-        sae.load_state_dict(state_dict)
+        sae.load_state_dict(state_dict, assign=True)
 
         # Apply normalization if needed
         if cfg_dict.get("normalize_activations") == "expected_average_only_in":
@@ -666,7 +700,13 @@ class SAE(HookedRootModule, Generic[T_SAE_CONFIG], ABC):
                     f"norm_scaling_factor not found for {release} and {sae_id}, but normalize_activations is 'expected_average_only_in'. Skipping normalization folding."
                 )
 
-        return sae, cfg_dict, log_sparsities
+        # the loaders should already handle the dtype / device conversion
+        # but this is a fallback to guarantee the SAE is on the correct device and dtype
+        return (
+            sae.to(dtype=str_to_dtype(dtype), device=device),
+            cfg_dict,
+            log_sparsities,
+        )
 
     @classmethod
     def from_dict(cls: type[T_SAE], config_dict: dict[str, Any]) -> T_SAE:

sae_lens/training/activations_store.py

@@ -24,7 +24,7 @@ from sae_lens.config import (
     HfDataset,
     LanguageModelSAERunnerConfig,
 )
-from sae_lens.constants import ACTIVATIONS_STORE_STATE_FILENAME, DTYPE_MAP
+from sae_lens.constants import ACTIVATIONS_STORE_STATE_FILENAME
 from sae_lens.pretokenize_runner import get_special_token_from_cfg
 from sae_lens.saes.sae import SAE, T_SAE_CONFIG, T_TRAINING_SAE_CONFIG
 from sae_lens.tokenization_and_batching import concat_and_batch_sequences
@@ -32,6 +32,7 @@ from sae_lens.training.mixing_buffer import mixing_buffer
 from sae_lens.util import (
     extract_stop_at_layer_from_tlens_hook_name,
     get_special_token_ids,
+    str_to_dtype,
 )
 
 
@@ -258,7 +259,7 @@ class ActivationsStore:
         self.prepend_bos = prepend_bos
         self.normalize_activations = normalize_activations
         self.device = torch.device(device)
-        self.dtype = DTYPE_MAP[dtype]
+        self.dtype = str_to_dtype(dtype)
         self.cached_activations_path = cached_activations_path
         self.autocast_lm = autocast_lm
         self.seqpos_slice = seqpos_slice

sae_lens/util.py

@@ -5,8 +5,11 @@ from dataclasses import asdict, fields, is_dataclass
 from pathlib import Path
 from typing import Sequence, TypeVar
 
+import torch
 from transformers import PreTrainedTokenizerBase
 
+from sae_lens.constants import DTYPE_MAP, DTYPE_TO_STR
+
 K = TypeVar("K")
 V = TypeVar("V")
 
@@ -90,3 +93,21 @@ def get_special_token_ids(tokenizer: PreTrainedTokenizerBase) -> list[int]:
                     special_tokens.add(token_id)
 
     return list(special_tokens)
+
+
+def str_to_dtype(dtype: str) -> torch.dtype:
+    """Convert a string to a torch.dtype."""
+    if dtype not in DTYPE_MAP:
+        raise ValueError(
+            f"Invalid dtype: {dtype}. Must be one of {list(DTYPE_MAP.keys())}"
+        )
+    return DTYPE_MAP[dtype]
+
+
+def dtype_to_str(dtype: torch.dtype) -> str:
+    """Convert a torch.dtype to a string."""
+    if dtype not in DTYPE_TO_STR:
+        raise ValueError(
+            f"Invalid dtype: {dtype}. Must be one of {list(DTYPE_TO_STR.keys())}"
+        )
+    return DTYPE_TO_STR[dtype]

{sae_lens-6.24.1.dist-info → sae_lens-6.25.1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sae-lens
-Version: 6.24.1
+Version: 6.25.1
 Summary: Training and Analyzing Sparse Autoencoders (SAEs)
 License: MIT
 License-File: LICENSE

{sae_lens-6.24.1.dist-info → sae_lens-6.25.1.dist-info}/RECORD

@@ -1,25 +1,25 @@
-sae_lens/__init__.py,sha256=spLEw4TR2BzzKc3R-ik8MbHlYOAR__wVmkSmJqOB4Tc,4268
+sae_lens/__init__.py,sha256=vWuA8EbynIJadj666RoFNCTIvoH9-HFpUxuHwoYt8Ks,4268
 sae_lens/analysis/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 sae_lens/analysis/hooked_sae_transformer.py,sha256=dQRgGVwce8XwylL2AzJE7l9elhtMRFCs2hdUj-Qyy4g,14038
 sae_lens/analysis/neuronpedia_integration.py,sha256=Gx1W7hUBEuMoasNcnOnZ1wmqbXDd1pSZ1nqKEya1HQc,4962
-sae_lens/cache_activations_runner.py,sha256=RNN_nDQkH0lqEIxTAIDx3g1cgAzRxQWBSBEXA6nbWh0,12565
-sae_lens/config.py,sha256=fxvpQxFfPOVUkryiHD19q9O1AJDSkIguWeYlbJuTxmY,30329
-sae_lens/constants.py,sha256=qX12uAE_xkha6hjss_0MGTbakI7gEkJzHABkZaHWQFU,683
+sae_lens/cache_activations_runner.py,sha256=Lvlz-k5-3XxVRtUdC4b1CiKyx5s0ckLa8GDGv9_kcxs,12566
+sae_lens/config.py,sha256=JmcrXT4orJV2OulbEZAciz8RQmYv7DrtUtRbOLsNQ2Y,30330
+sae_lens/constants.py,sha256=CM-h9AjZNAl2aP7hVpKk7YsFHpu-_Lfhhmq2d5qPEVc,887
 sae_lens/evals.py,sha256=P0NUsJeGzYxFBiVKhbPzd72IFKY4gH40HHlEZ3jEAmg,39598
 sae_lens/llm_sae_training_runner.py,sha256=M7BK55gSFYu2qFQKABHX3c8i46P1LfODCeyHFzGGuqU,15196
 sae_lens/load_model.py,sha256=C8AMykctj6H7tz_xRwB06-EXj6TfW64PtSJZR5Jxn1Y,8649
 sae_lens/loading/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-sae_lens/loading/pretrained_sae_loaders.py,sha256=W2eIvUU1wAHrYxGiZs4s2D6DnGBQqqKjq0wvXzWbD5c,63561
+sae_lens/loading/pretrained_sae_loaders.py,sha256=hq-dhxsEdUmlAnZEiZBqX7lNyQQwZ6KXmXZWpzAc5FY,63638
 sae_lens/loading/pretrained_saes_directory.py,sha256=hejNfLUepYCSGPalRfQwxxCEUqMMUPsn1tufwvwct5k,3820
 sae_lens/pretokenize_runner.py,sha256=amJwIz3CKi2s2wNQn-10E7eAV7VFhNqtFDNTeTkwEI8,7133
-sae_lens/pretrained_saes.yaml,sha256=Hd1GgaPL4TAXoS2gizG9e_9jc_9LpfI4w_hwGkEz9xQ,1509314
+sae_lens/pretrained_saes.yaml,sha256=Hy9mk4Liy50B0CIBD4ER1ETcho2drFFiIy-bPVCN_lc,1510210
 sae_lens/registry.py,sha256=nhy7BPSudSATqW4lo9H_k3Na7sfGHmAf9v-3wpnLL_o,1490
 sae_lens/saes/__init__.py,sha256=fYVujOzNnUgpzLL0MBLBt_DNX2CPcTaheukzCd2bEPo,1906
 sae_lens/saes/batchtopk_sae.py,sha256=x4EbgZl0GUickRPcCmtKNGS2Ra3Uy1Z1OtF2FnrSabQ,5422
 sae_lens/saes/gated_sae.py,sha256=mHnmw-RD7hqIbP9_EBj3p2SK0OqQIkZivdOKRygeRgw,8825
 sae_lens/saes/jumprelu_sae.py,sha256=udjGHp3WTABQSL2Qq57j-bINWX61GCmo68EmdjMOXoo,13310
 sae_lens/saes/matryoshka_batchtopk_sae.py,sha256=Qr6htt1HHOuO9FXI9hyaPSnGFIiJG-v7y1t1CEmkFzM,5995
-sae_lens/saes/sae.py,sha256=Vb1aGSDPRv_0J2aL8-EICRSkIxsO6Q4lJaJE9NNmfdA,37749
+sae_lens/saes/sae.py,sha256=fzXv8lwHskSxsf8hm_wlKPkpq50iafmBjBNQzwZ6a00,40050
 sae_lens/saes/standard_sae.py,sha256=nEVETwAmRD2tyX7ESIic1fij48gAq1Dh7s_GQ2fqCZ4,5747
 sae_lens/saes/temporal_sae.py,sha256=DsecivcHWId-MTuJpQbz8OhqtmGhZACxJauYZGHo0Ok,13272
 sae_lens/saes/topk_sae.py,sha256=vrMRPrCQR1o8G_kXqY_EAoGZARupkQNFB2dNZVLsusE,21073
@@ -27,15 +27,15 @@ sae_lens/saes/transcoder.py,sha256=CTpJs8ASOK06npih7gZHygZuxqTR7HICWlOYfTiKjI4,1
 sae_lens/tokenization_and_batching.py,sha256=D_o7cXvRqhT89H3wNzoRymNALNE6eHojBWLdXOUwUGE,5438
 sae_lens/training/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 sae_lens/training/activation_scaler.py,sha256=FzNfgBplLWmyiSlZ6TUvE-nur3lOiGTrlvC97ys8S24,1973
-sae_lens/training/activations_store.py,sha256=yDWw7TZGPFM_O8_Oi78j8lLIHJJesxq9TKVP_TrMX-M,33768
+sae_lens/training/activations_store.py,sha256=rQadexm2BiwK7_MZIPlRkcKSqabi3iuOTC-R8aJchS8,33778
 sae_lens/training/mixing_buffer.py,sha256=vDpYG5ZE70szDvBsRKcNHEES3h_WTKJ16qDYk5jPOVA,2015
 sae_lens/training/optim.py,sha256=bJpqqcK4enkcPvQAJkeH4Ci1LUOlfjIMTv6-IlaAbRA,5588
 sae_lens/training/sae_trainer.py,sha256=zhkabyIKxI_tZTV3_kwz6zMrHZ95Ecr97krmwc-9ffs,17600
 sae_lens/training/types.py,sha256=1FpLx_Doda9vZpmfm-x1e8wGBYpyhe9Kpb_JuM5nIFM,90
 sae_lens/training/upload_saes_to_huggingface.py,sha256=r_WzI1zLtGZ5TzAxuG3xa_8T09j3zXJrWd_vzPsPGkQ,4469
 sae_lens/tutorial/tsea.py,sha256=fd1am_XXsf2KMbByDapJo-2qlxduKaa62Z2qcQZ3QKU,18145
-sae_lens/util.py,sha256=tCovQ-eZa1L7thPpNDL6PGOJrIMML2yLI5e0EHCOpS8,3309
-sae_lens-6.24.1.dist-info/METADATA,sha256=5TlxCqEZoJV4S0F9IP6Ak_aitVkMkFfUhlFOl5NIJBc,5361
-sae_lens-6.24.1.dist-info/WHEEL,sha256=zp0Cn7JsFoX2ATtOhtaFYIiE2rmFAD4OcMhtUki8W3U,88
-sae_lens-6.24.1.dist-info/licenses/LICENSE,sha256=DW6e-hDosiu4CfW0-imI57sV1I5f9UEslpviNQcOAKs,1069
-sae_lens-6.24.1.dist-info/RECORD,,
+sae_lens/util.py,sha256=spkcmQUsjVYFn5H2032nQYr1CKGVnv3tAdfIpY59-Mg,3919
+sae_lens-6.25.1.dist-info/METADATA,sha256=gClFVWzEWNNjrXsGqvCY6ry6ehXIFwp8PB0jIOhmQvc,5361
+sae_lens-6.25.1.dist-info/WHEEL,sha256=zp0Cn7JsFoX2ATtOhtaFYIiE2rmFAD4OcMhtUki8W3U,88
+sae_lens-6.25.1.dist-info/licenses/LICENSE,sha256=DW6e-hDosiu4CfW0-imI57sV1I5f9UEslpviNQcOAKs,1069
+sae_lens-6.25.1.dist-info/RECORD,,