sae-lens 6.12.1__tar.gz → 6.12.3__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- {sae_lens-6.12.1 → sae_lens-6.12.3}/PKG-INFO +4 -2
- {sae_lens-6.12.1 → sae_lens-6.12.3}/pyproject.toml +1 -1
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/__init__.py +1 -1
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/loading/pretrained_sae_loaders.py +6 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/saes/sae.py +1 -1
- {sae_lens-6.12.1 → sae_lens-6.12.3}/LICENSE +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/README.md +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/analysis/__init__.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/analysis/hooked_sae_transformer.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/analysis/neuronpedia_integration.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/cache_activations_runner.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/config.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/constants.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/evals.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/llm_sae_training_runner.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/load_model.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/loading/__init__.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/loading/pretrained_saes_directory.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/pretokenize_runner.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/pretrained_saes.yaml +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/registry.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/saes/__init__.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/saes/batchtopk_sae.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/saes/gated_sae.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/saes/jumprelu_sae.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/saes/standard_sae.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/saes/topk_sae.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/saes/transcoder.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/tokenization_and_batching.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/training/__init__.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/training/activation_scaler.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/training/activations_store.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/training/mixing_buffer.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/training/optim.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/training/sae_trainer.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/training/types.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/training/upload_saes_to_huggingface.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/tutorial/tsea.py +0 -0
- {sae_lens-6.12.1 → sae_lens-6.12.3}/sae_lens/util.py +0 -0
|
@@ -1,8 +1,9 @@
|
|
|
1
|
-
Metadata-Version: 2.
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
2
|
Name: sae-lens
|
|
3
|
-
Version: 6.12.
|
|
3
|
+
Version: 6.12.3
|
|
4
4
|
Summary: Training and Analyzing Sparse Autoencoders (SAEs)
|
|
5
5
|
License: MIT
|
|
6
|
+
License-File: LICENSE
|
|
6
7
|
Keywords: deep-learning,sparse-autoencoders,mechanistic-interpretability,PyTorch
|
|
7
8
|
Author: Joseph Bloom
|
|
8
9
|
Requires-Python: >=3.10,<4.0
|
|
@@ -12,6 +13,7 @@ Classifier: Programming Language :: Python :: 3.10
|
|
|
12
13
|
Classifier: Programming Language :: Python :: 3.11
|
|
13
14
|
Classifier: Programming Language :: Python :: 3.12
|
|
14
15
|
Classifier: Programming Language :: Python :: 3.13
|
|
16
|
+
Classifier: Programming Language :: Python :: 3.14
|
|
15
17
|
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
|
|
16
18
|
Provides-Extra: mamba
|
|
17
19
|
Requires-Dist: babe (>=0.0.7,<0.0.8)
|
|
@@ -233,6 +233,12 @@ def handle_pre_6_0_config(cfg_dict: dict[str, Any]) -> dict[str, Any]:
|
|
|
233
233
|
"reshape_activations",
|
|
234
234
|
"hook_z" if "hook_z" in new_cfg.get("hook_name", "") else "none",
|
|
235
235
|
)
|
|
236
|
+
if (
|
|
237
|
+
new_cfg.get("activation_fn") == "topk"
|
|
238
|
+
and new_cfg.get("activation_fn_kwargs", {}).get("k") is not None
|
|
239
|
+
):
|
|
240
|
+
new_cfg["architecture"] = "topk"
|
|
241
|
+
new_cfg["k"] = new_cfg["activation_fn_kwargs"]["k"]
|
|
236
242
|
|
|
237
243
|
if "normalize_activations" in new_cfg and isinstance(
|
|
238
244
|
new_cfg["normalize_activations"], bool
|
|
@@ -245,7 +245,7 @@ class SAE(HookedRootModule, Generic[T_SAE_CONFIG], ABC):
|
|
|
245
245
|
|
|
246
246
|
self.cfg = cfg
|
|
247
247
|
|
|
248
|
-
if cfg.metadata and cfg.metadata:
|
|
248
|
+
if cfg.metadata and cfg.metadata.model_from_pretrained_kwargs:
|
|
249
249
|
warnings.warn(
|
|
250
250
|
"\nThis SAE has non-empty model_from_pretrained_kwargs. "
|
|
251
251
|
"\nFor optimal performance, load the model like so:\n"
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|