sae-lens 6.29.0__tar.gz → 6.30.1__tar.gz

{sae_lens-6.29.0 → sae_lens-6.30.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sae-lens
-Version: 6.29.0
+Version: 6.30.1
 Summary: Training and Analyzing Sparse Autoencoders (SAEs)
 License: MIT
 License-File: LICENSE

{sae_lens-6.29.0 → sae_lens-6.30.1}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "sae-lens"
-version = "6.29.0"
+version = "6.30.1"
 description = "Training and Analyzing Sparse Autoencoders (SAEs)"
 authors = ["Joseph Bloom"]
 readme = "README.md"

{sae_lens-6.29.0 → sae_lens-6.30.1}/sae_lens/__init__.py

@@ -1,5 +1,5 @@
 # ruff: noqa: E402
-__version__ = "6.29.0"
+__version__ = "6.30.1"
 
 import logging
 

{sae_lens-6.29.0 → sae_lens-6.30.1}/sae_lens/loading/pretrained_sae_loaders.py

@@ -575,6 +575,8 @@ def _infer_gemma_3_raw_cfg_dict(repo_id: str, folder_name: str) -> dict[str, Any
         "model_name": model_name,
         "hf_hook_point_in": hf_hook_point_in,
     }
+    if "transcoder" in folder_name or "clt" in folder_name:
+        cfg["affine_connection"] = "affine" in folder_name
     if hf_hook_point_out is not None:
         cfg["hf_hook_point_out"] = hf_hook_point_out
 
@@ -614,11 +616,11 @@ def get_gemma_3_config_from_hf(
     if "resid_post" in folder_name:
         hook_name = f"blocks.{layer}.hook_resid_post"
     elif "attn_out" in folder_name:
-        hook_name = f"blocks.{layer}.hook_attn_out"
+        hook_name = f"blocks.{layer}.attn.hook_z"
     elif "mlp_out" in folder_name:
         hook_name = f"blocks.{layer}.hook_mlp_out"
     elif "transcoder" in folder_name or "clt" in folder_name:
-        hook_name = f"blocks.{layer}.ln2.hook_normalized"
+        hook_name = f"blocks.{layer}.hook_mlp_in"
         hook_name_out = f"blocks.{layer}.hook_mlp_out"
     else:
         raise ValueError("Hook name not found in folder_name.")
@@ -643,7 +645,11 @@ def get_gemma_3_config_from_hf(
 
     architecture = "jumprelu"
     if "transcoder" in folder_name or "clt" in folder_name:
-        architecture = "jumprelu_skip_transcoder"
+        architecture = (
+            "jumprelu_skip_transcoder"
+            if raw_cfg_dict.get("affine_connection", False)
+            else "jumprelu_transcoder"
+        )
         d_out = shapes_dict["w_dec"][-1]
 
     cfg = {

{sae_lens-6.29.0 → sae_lens-6.30.1}/sae_lens/pretrained_saes.yaml

@@ -4148,6 +4148,7 @@ gemma-scope-2-4b-it-res:
   - id: layer_17_width_16k_l0_medium
     path: resid_post/layer_17_width_16k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-4b-it/17-gemmascope-2-res-16k
   - id: layer_17_width_16k_l0_small
     path: resid_post/layer_17_width_16k_l0_small
     l0: 20
@@ -4166,6 +4167,7 @@ gemma-scope-2-4b-it-res:
   - id: layer_17_width_262k_l0_medium
     path: resid_post/layer_17_width_262k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-4b-it/17-gemmascope-2-res-262k
   - id: layer_17_width_262k_l0_medium_seed_1
     path: resid_post/layer_17_width_262k_l0_medium_seed_1
     l0: 60
@@ -4178,6 +4180,7 @@ gemma-scope-2-4b-it-res:
   - id: layer_17_width_65k_l0_medium
     path: resid_post/layer_17_width_65k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-4b-it/17-gemmascope-2-res-65k
   - id: layer_17_width_65k_l0_small
     path: resid_post/layer_17_width_65k_l0_small
     l0: 20
@@ -4187,6 +4190,7 @@ gemma-scope-2-4b-it-res:
   - id: layer_22_width_16k_l0_medium
     path: resid_post/layer_22_width_16k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-4b-it/22-gemmascope-2-res-16k
   - id: layer_22_width_16k_l0_small
     path: resid_post/layer_22_width_16k_l0_small
     l0: 20
@@ -4205,6 +4209,7 @@ gemma-scope-2-4b-it-res:
   - id: layer_22_width_262k_l0_medium
     path: resid_post/layer_22_width_262k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-4b-it/22-gemmascope-2-res-262k
   - id: layer_22_width_262k_l0_medium_seed_1
     path: resid_post/layer_22_width_262k_l0_medium_seed_1
     l0: 60
@@ -4217,6 +4222,7 @@ gemma-scope-2-4b-it-res:
   - id: layer_22_width_65k_l0_medium
     path: resid_post/layer_22_width_65k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-4b-it/22-gemmascope-2-res-65k
   - id: layer_22_width_65k_l0_small
     path: resid_post/layer_22_width_65k_l0_small
     l0: 20
@@ -4226,6 +4232,7 @@ gemma-scope-2-4b-it-res:
   - id: layer_29_width_16k_l0_medium
     path: resid_post/layer_29_width_16k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-4b-it/29-gemmascope-2-res-16k
   - id: layer_29_width_16k_l0_small
     path: resid_post/layer_29_width_16k_l0_small
     l0: 20
@@ -4244,6 +4251,7 @@ gemma-scope-2-4b-it-res:
   - id: layer_29_width_262k_l0_medium
     path: resid_post/layer_29_width_262k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-4b-it/29-gemmascope-2-res-262k
   - id: layer_29_width_262k_l0_medium_seed_1
     path: resid_post/layer_29_width_262k_l0_medium_seed_1
     l0: 60
@@ -4256,6 +4264,7 @@ gemma-scope-2-4b-it-res:
   - id: layer_29_width_65k_l0_medium
     path: resid_post/layer_29_width_65k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-4b-it/29-gemmascope-2-res-65k
   - id: layer_29_width_65k_l0_small
     path: resid_post/layer_29_width_65k_l0_small
     l0: 20
@@ -4265,6 +4274,7 @@ gemma-scope-2-4b-it-res:
   - id: layer_9_width_16k_l0_medium
     path: resid_post/layer_9_width_16k_l0_medium
     l0: 53
+    neuronpedia: gemma-3-4b-it/9-gemmascope-2-res-16k
   - id: layer_9_width_16k_l0_small
     path: resid_post/layer_9_width_16k_l0_small
     l0: 17
@@ -4283,6 +4293,7 @@ gemma-scope-2-4b-it-res:
   - id: layer_9_width_262k_l0_medium
     path: resid_post/layer_9_width_262k_l0_medium
     l0: 53
+    neuronpedia: gemma-3-4b-it/9-gemmascope-2-res-262k
   - id: layer_9_width_262k_l0_medium_seed_1
     path: resid_post/layer_9_width_262k_l0_medium_seed_1
     l0: 53
@@ -4295,6 +4306,7 @@ gemma-scope-2-4b-it-res:
   - id: layer_9_width_65k_l0_medium
     path: resid_post/layer_9_width_65k_l0_medium
     l0: 53
+    neuronpedia: gemma-3-4b-it/9-gemmascope-2-res-65k
   - id: layer_9_width_65k_l0_small
     path: resid_post/layer_9_width_65k_l0_small
     l0: 17
@@ -14491,6 +14503,7 @@ gemma-scope-2-270m-it-res:
   - id: layer_12_width_16k_l0_medium
     path: resid_post/layer_12_width_16k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-270m-it/12-gemmascope-2-res-16k
   - id: layer_12_width_16k_l0_small
     path: resid_post/layer_12_width_16k_l0_small
     l0: 20
@@ -14509,6 +14522,7 @@ gemma-scope-2-270m-it-res:
   - id: layer_12_width_262k_l0_medium
     path: resid_post/layer_12_width_262k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-270m-it/12-gemmascope-2-res-262k
   - id: layer_12_width_262k_l0_medium_seed_1
     path: resid_post/layer_12_width_262k_l0_medium_seed_1
     l0: 60
@@ -14521,6 +14535,7 @@ gemma-scope-2-270m-it-res:
   - id: layer_12_width_65k_l0_medium
     path: resid_post/layer_12_width_65k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-270m-it/12-gemmascope-2-res-65k
   - id: layer_12_width_65k_l0_small
     path: resid_post/layer_12_width_65k_l0_small
     l0: 20
@@ -14530,6 +14545,7 @@ gemma-scope-2-270m-it-res:
   - id: layer_15_width_16k_l0_medium
     path: resid_post/layer_15_width_16k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-270m-it/15-gemmascope-2-res-16k
   - id: layer_15_width_16k_l0_small
     path: resid_post/layer_15_width_16k_l0_small
     l0: 20
@@ -14548,6 +14564,7 @@ gemma-scope-2-270m-it-res:
   - id: layer_15_width_262k_l0_medium
     path: resid_post/layer_15_width_262k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-270m-it/15-gemmascope-2-res-262k
   - id: layer_15_width_262k_l0_medium_seed_1
     path: resid_post/layer_15_width_262k_l0_medium_seed_1
     l0: 60
@@ -14560,6 +14577,7 @@ gemma-scope-2-270m-it-res:
   - id: layer_15_width_65k_l0_medium
     path: resid_post/layer_15_width_65k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-270m-it/15-gemmascope-2-res-65k
   - id: layer_15_width_65k_l0_small
     path: resid_post/layer_15_width_65k_l0_small
     l0: 20
@@ -14569,6 +14587,7 @@ gemma-scope-2-270m-it-res:
   - id: layer_5_width_16k_l0_medium
     path: resid_post/layer_5_width_16k_l0_medium
     l0: 55
+    neuronpedia: gemma-3-270m-it/5-gemmascope-2-res-16k
   - id: layer_5_width_16k_l0_small
     path: resid_post/layer_5_width_16k_l0_small
     l0: 18
@@ -14587,6 +14606,7 @@ gemma-scope-2-270m-it-res:
   - id: layer_5_width_262k_l0_medium
     path: resid_post/layer_5_width_262k_l0_medium
     l0: 55
+    neuronpedia: gemma-3-270m-it/5-gemmascope-2-res-262k
   - id: layer_5_width_262k_l0_medium_seed_1
     path: resid_post/layer_5_width_262k_l0_medium_seed_1
     l0: 55
@@ -14599,6 +14619,7 @@ gemma-scope-2-270m-it-res:
   - id: layer_5_width_65k_l0_medium
     path: resid_post/layer_5_width_65k_l0_medium
     l0: 55
+    neuronpedia: gemma-3-270m-it/5-gemmascope-2-res-65k
   - id: layer_5_width_65k_l0_small
     path: resid_post/layer_5_width_65k_l0_small
     l0: 18
@@ -14608,6 +14629,7 @@ gemma-scope-2-270m-it-res:
   - id: layer_9_width_16k_l0_medium
     path: resid_post/layer_9_width_16k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-270m-it/9-gemmascope-2-res-16k
   - id: layer_9_width_16k_l0_small
     path: resid_post/layer_9_width_16k_l0_small
     l0: 20
@@ -14626,6 +14648,7 @@ gemma-scope-2-270m-it-res:
   - id: layer_9_width_262k_l0_medium
     path: resid_post/layer_9_width_262k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-270m-it/9-gemmascope-2-res-262k
   - id: layer_9_width_262k_l0_medium_seed_1
     path: resid_post/layer_9_width_262k_l0_medium_seed_1
     l0: 60
@@ -14638,6 +14661,7 @@ gemma-scope-2-270m-it-res:
   - id: layer_9_width_65k_l0_medium
     path: resid_post/layer_9_width_65k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-270m-it/9-gemmascope-2-res-65k
   - id: layer_9_width_65k_l0_small
     path: resid_post/layer_9_width_65k_l0_small
     l0: 20
@@ -18727,6 +18751,7 @@ gemma-scope-2-1b-it-res:
   - id: layer_13_width_16k_l0_medium
     path: resid_post/layer_13_width_16k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-1b-it/13-gemmascope-2-res-16k
   - id: layer_13_width_16k_l0_small
     path: resid_post/layer_13_width_16k_l0_small
     l0: 20
@@ -18745,6 +18770,7 @@ gemma-scope-2-1b-it-res:
   - id: layer_13_width_262k_l0_medium
     path: resid_post/layer_13_width_262k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-1b-it/13-gemmascope-2-res-262k
   - id: layer_13_width_262k_l0_medium_seed_1
     path: resid_post/layer_13_width_262k_l0_medium_seed_1
     l0: 60
@@ -18757,6 +18783,7 @@ gemma-scope-2-1b-it-res:
   - id: layer_13_width_65k_l0_medium
     path: resid_post/layer_13_width_65k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-1b-it/13-gemmascope-2-res-65k
   - id: layer_13_width_65k_l0_small
     path: resid_post/layer_13_width_65k_l0_small
     l0: 20
@@ -18766,6 +18793,7 @@ gemma-scope-2-1b-it-res:
   - id: layer_17_width_16k_l0_medium
     path: resid_post/layer_17_width_16k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-1b-it/17-gemmascope-2-res-16k
   - id: layer_17_width_16k_l0_small
     path: resid_post/layer_17_width_16k_l0_small
     l0: 20
@@ -18784,6 +18812,7 @@ gemma-scope-2-1b-it-res:
   - id: layer_17_width_262k_l0_medium
     path: resid_post/layer_17_width_262k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-1b-it/17-gemmascope-2-res-262k
   - id: layer_17_width_262k_l0_medium_seed_1
     path: resid_post/layer_17_width_262k_l0_medium_seed_1
     l0: 60
@@ -18796,6 +18825,7 @@ gemma-scope-2-1b-it-res:
   - id: layer_17_width_65k_l0_medium
     path: resid_post/layer_17_width_65k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-1b-it/17-gemmascope-2-res-65k
   - id: layer_17_width_65k_l0_small
     path: resid_post/layer_17_width_65k_l0_small
     l0: 20
@@ -18805,6 +18835,7 @@ gemma-scope-2-1b-it-res:
   - id: layer_22_width_16k_l0_medium
     path: resid_post/layer_22_width_16k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-1b-it/22-gemmascope-2-res-16k
   - id: layer_22_width_16k_l0_small
     path: resid_post/layer_22_width_16k_l0_small
     l0: 20
@@ -18823,6 +18854,7 @@ gemma-scope-2-1b-it-res:
   - id: layer_22_width_262k_l0_medium
     path: resid_post/layer_22_width_262k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-1b-it/22-gemmascope-2-res-262k
   - id: layer_22_width_262k_l0_medium_seed_1
     path: resid_post/layer_22_width_262k_l0_medium_seed_1
     l0: 60
@@ -18835,6 +18867,7 @@ gemma-scope-2-1b-it-res:
   - id: layer_22_width_65k_l0_medium
     path: resid_post/layer_22_width_65k_l0_medium
     l0: 60
+    neuronpedia: gemma-3-1b-it/22-gemmascope-2-res-65k
   - id: layer_22_width_65k_l0_small
     path: resid_post/layer_22_width_65k_l0_small
     l0: 20
@@ -18844,6 +18877,7 @@ gemma-scope-2-1b-it-res:
   - id: layer_7_width_16k_l0_medium
     path: resid_post/layer_7_width_16k_l0_medium
     l0: 54
+    neuronpedia: gemma-3-1b-it/7-gemmascope-2-res-16k
   - id: layer_7_width_16k_l0_small
     path: resid_post/layer_7_width_16k_l0_small
     l0: 18
@@ -18862,6 +18896,7 @@ gemma-scope-2-1b-it-res:
   - id: layer_7_width_262k_l0_medium
     path: resid_post/layer_7_width_262k_l0_medium
     l0: 54
+    neuronpedia: gemma-3-1b-it/7-gemmascope-2-res-262k
   - id: layer_7_width_262k_l0_medium_seed_1
     path: resid_post/layer_7_width_262k_l0_medium_seed_1
     l0: 54
@@ -18874,6 +18909,7 @@ gemma-scope-2-1b-it-res:
   - id: layer_7_width_65k_l0_medium
     path: resid_post/layer_7_width_65k_l0_medium
     l0: 54
+    neuronpedia: gemma-3-1b-it/7-gemmascope-2-res-65k
   - id: layer_7_width_65k_l0_small
     path: resid_post/layer_7_width_65k_l0_small
     l0: 18

{sae_lens-6.29.0 → sae_lens-6.30.1}/sae_lens/synthetic/activation_generator.py

@@ -2,12 +2,12 @@
 Functions for generating synthetic feature activations.
 """
 
+import math
 from collections.abc import Callable, Sequence
 
 import torch
-from scipy.stats import norm
 from torch import nn
-from torch.distributions import LowRankMultivariateNormal, MultivariateNormal
+from torch.distributions import MultivariateNormal
 
 from sae_lens.synthetic.correlation import LowRankCorrelationMatrix
 from sae_lens.util import str_to_dtype
@@ -34,6 +34,7 @@ class ActivationGenerator(nn.Module):
     correlation_matrix: torch.Tensor | None
     low_rank_correlation: tuple[torch.Tensor, torch.Tensor] | None
     correlation_thresholds: torch.Tensor | None
+    use_sparse_tensors: bool
 
     def __init__(
         self,
@@ -45,7 +46,34 @@ class ActivationGenerator(nn.Module):
         correlation_matrix: CorrelationMatrixInput | None = None,
         device: torch.device | str = "cpu",
         dtype: torch.dtype | str = "float32",
+        use_sparse_tensors: bool = False,
     ):
+        """
+        Create a new ActivationGenerator.
+
+        Args:
+            num_features: Number of features to generate activations for.
+            firing_probabilities: Probability of each feature firing. Can be a single
+                float (applied to all features) or a tensor of shape (num_features,).
+            std_firing_magnitudes: Standard deviation of firing magnitudes. Can be a
+                single float or a tensor of shape (num_features,). Defaults to 0.0
+                (deterministic magnitudes).
+            mean_firing_magnitudes: Mean firing magnitude when a feature fires. Can be
+                a single float or a tensor of shape (num_features,). Defaults to 1.0.
+            modify_activations: Optional function(s) to modify activations after
+                generation. Can be a single callable, a sequence of callables (applied
+                in order), or None. Useful for applying hierarchy constraints.
+            correlation_matrix: Optional correlation structure between features. Can be:
+
+                - A full correlation matrix tensor of shape (num_features, num_features)
+                - A LowRankCorrelationMatrix for memory-efficient large-scale correlations
+                - A tuple of (factor, diag) tensors representing low-rank structure
+
+            device: Device to place tensors on. Defaults to "cpu".
+            dtype: Data type for tensors. Defaults to "float32".
+            use_sparse_tensors: If True, return sparse COO tensors from sample().
+                Only recommended when using massive numbers of features. Defaults to False.
+        """
         super().__init__()
         self.num_features = num_features
         self.firing_probabilities = _to_tensor(
@@ -61,6 +89,7 @@ class ActivationGenerator(nn.Module):
         self.correlation_thresholds = None
         self.correlation_matrix = None
         self.low_rank_correlation = None
+        self.use_sparse_tensors = use_sparse_tensors
 
         if correlation_matrix is not None:
             if isinstance(correlation_matrix, torch.Tensor):
@@ -76,12 +105,15 @@ class ActivationGenerator(nn.Module):
                 _validate_low_rank_correlation(
                     correlation_factor, correlation_diag, num_features
                 )
-                self.low_rank_correlation = (correlation_factor, correlation_diag)
+                # Pre-compute sqrt for efficiency (used every sample call)
+                self.low_rank_correlation = (
+                    correlation_factor,
+                    correlation_diag.sqrt(),
+                )
 
-            self.correlation_thresholds = torch.tensor(
-                [norm.ppf(1 - p.item()) for p in self.firing_probabilities],
-                device=device,
-                dtype=self.firing_probabilities.dtype,
+            # Vectorized inverse normal CDF: norm.ppf(1-p) = sqrt(2) * erfinv(1 - 2*p)
+            self.correlation_thresholds = math.sqrt(2) * torch.erfinv(
+                1 - 2 * self.firing_probabilities
             )
 
     @torch.no_grad()
@@ -105,7 +137,7 @@ class ActivationGenerator(nn.Module):
 
         if self.correlation_matrix is not None:
             assert self.correlation_thresholds is not None
-            firing_features = _generate_correlated_features(
+            firing_indices = _generate_correlated_features(
                 batch_size,
                 self.correlation_matrix,
                 self.correlation_thresholds,
@@ -113,7 +145,7 @@ class ActivationGenerator(nn.Module):
             )
         elif self.low_rank_correlation is not None:
             assert self.correlation_thresholds is not None
-            firing_features = _generate_low_rank_correlated_features(
+            firing_indices = _generate_low_rank_correlated_features(
                 batch_size,
                 self.low_rank_correlation[0],
                 self.low_rank_correlation[1],
@@ -121,23 +153,58 @@ class ActivationGenerator(nn.Module):
                 device,
             )
         else:
-            firing_features = torch.bernoulli(
+            firing_indices = torch.bernoulli(
                 self.firing_probabilities.unsqueeze(0).expand(batch_size, -1)
+            ).nonzero(as_tuple=True)
+
+        # Compute activations only at firing positions (sparse optimization)
+        feature_indices = firing_indices[1]
+        num_firing = feature_indices.shape[0]
+        mean_at_firing = self.mean_firing_magnitudes[feature_indices]
+        std_at_firing = self.std_firing_magnitudes[feature_indices]
+        random_deltas = (
+            torch.randn(
+                num_firing, device=device, dtype=self.mean_firing_magnitudes.dtype
             )
-
-        firing_magnitude_delta = torch.normal(
-            torch.zeros_like(self.firing_probabilities)
-            .unsqueeze(0)
-            .expand(batch_size, -1),
-            self.std_firing_magnitudes.unsqueeze(0).expand(batch_size, -1),
+            * std_at_firing
         )
-        firing_magnitude_delta[firing_features == 0] = 0
-        feature_activations = (
-            firing_features * self.mean_firing_magnitudes + firing_magnitude_delta
-        ).relu()
+        activations_at_firing = (mean_at_firing + random_deltas).relu()
+
+        if self.use_sparse_tensors:
+            # Return sparse COO tensor
+            indices = torch.stack(firing_indices)  # [2, nnz]
+            feature_activations = torch.sparse_coo_tensor(
+                indices,
+                activations_at_firing,
+                size=(batch_size, self.num_features),
+                device=device,
+                dtype=self.mean_firing_magnitudes.dtype,
+            )
+        else:
+            # Dense tensor path
+            feature_activations = torch.zeros(
+                batch_size,
+                self.num_features,
+                device=device,
+                dtype=self.mean_firing_magnitudes.dtype,
+            )
+            feature_activations[firing_indices] = activations_at_firing
 
         if self.modify_activations is not None:
-            feature_activations = self.modify_activations(feature_activations).relu()
+            feature_activations = self.modify_activations(feature_activations)
+            if feature_activations.is_sparse:
+                # Apply relu to sparse values
+                feature_activations = feature_activations.coalesce()
+                feature_activations = torch.sparse_coo_tensor(
+                    feature_activations.indices(),
+                    feature_activations.values().relu(),
+                    feature_activations.shape,
+                    device=feature_activations.device,
+                    dtype=feature_activations.dtype,
+                )
+            else:
+                feature_activations = feature_activations.relu()
+
         return feature_activations
 
     def forward(self, batch_size: int) -> torch.Tensor:
@@ -149,7 +216,7 @@ def _generate_correlated_features(
     correlation_matrix: torch.Tensor,
     thresholds: torch.Tensor,
     device: torch.device,
-) -> torch.Tensor:
+) -> tuple[torch.Tensor, torch.Tensor]:
     """
     Generate correlated binary features using multivariate Gaussian sampling.
 
@@ -163,7 +230,7 @@ def _generate_correlated_features(
         device: Device to generate samples on
 
     Returns:
-        Binary feature matrix of shape (batch_size, num_features)
+        Tuple of (row_indices, col_indices) for firing features
     """
     num_features = correlation_matrix.shape[0]
 
@@ -173,16 +240,17 @@ def _generate_correlated_features(
     )
 
     gaussian_samples = mvn.sample((batch_size,))
-    return (gaussian_samples > thresholds.unsqueeze(0)).float()
+    indices = (gaussian_samples > thresholds.unsqueeze(0)).nonzero(as_tuple=True)
+    return indices[0], indices[1]
 
 
 def _generate_low_rank_correlated_features(
     batch_size: int,
     correlation_factor: torch.Tensor,
-    correlation_diag: torch.Tensor,
+    cov_diag_sqrt: torch.Tensor,
     thresholds: torch.Tensor,
     device: torch.device,
-) -> torch.Tensor:
+) -> tuple[torch.Tensor, torch.Tensor]:
     """
     Generate correlated binary features using low-rank multivariate Gaussian sampling.
 
@@ -192,23 +260,29 @@ def _generate_low_rank_correlated_features(
     Args:
         batch_size: Number of samples to generate
         correlation_factor: Factor matrix of shape (num_features, rank)
-        correlation_diag: Diagonal term of shape (num_features,)
+        cov_diag_sqrt: Pre-computed sqrt of diagonal term, shape (num_features,)
         thresholds: Pre-computed thresholds for each feature (from inverse normal CDF)
         device: Device to generate samples on
 
     Returns:
-        Binary feature matrix of shape (batch_size, num_features)
+        Tuple of (row_indices, col_indices) for firing features
     """
-    mvn = LowRankMultivariateNormal(
-        loc=torch.zeros(
-            correlation_factor.shape[0], device=device, dtype=thresholds.dtype
-        ),
-        cov_factor=correlation_factor.to(device=device, dtype=thresholds.dtype),
-        cov_diag=correlation_diag.to(device=device, dtype=thresholds.dtype),
+    # Manual low-rank MVN sampling to enable autocast for the expensive matmul
+    # samples = eps @ cov_factor.T + eta * sqrt(cov_diag)
+    # where eps ~ N(0, I_rank) and eta ~ N(0, I_n)
+
+    num_features, rank = correlation_factor.shape
+
+    # Generate random samples in float32 for numerical stability
+    eps = torch.randn(batch_size, rank, device=device, dtype=correlation_factor.dtype)
+    eta = torch.randn(
+        batch_size, num_features, device=device, dtype=cov_diag_sqrt.dtype
     )
 
-    gaussian_samples = mvn.sample((batch_size,))
-    return (gaussian_samples > thresholds.unsqueeze(0)).float()
+    gaussian_samples = eps @ correlation_factor.T + eta * cov_diag_sqrt
+
+    indices = (gaussian_samples > thresholds.unsqueeze(0)).nonzero(as_tuple=True)
+    return indices[0], indices[1]
 
 
 def _to_tensor(

{sae_lens-6.29.0 → sae_lens-6.30.1}/sae_lens/synthetic/feature_dictionary.py

@@ -168,9 +168,18 @@ class FeatureDictionary(nn.Module):
 
         Args:
             feature_activations: Tensor of shape [batch, num_features] containing
-                sparse feature activation values
+                sparse feature activation values. Can be dense or sparse COO.
 
         Returns:
             Tensor of shape [batch, hidden_dim] containing dense hidden activations
         """
+        if feature_activations.is_sparse:
+            # autocast is disabled here because sparse matmul is not supported with bfloat16
+            with torch.autocast(
+                device_type=feature_activations.device.type, enabled=False
+            ):
+                return (
+                    torch.sparse.mm(feature_activations, self.feature_vectors)
+                    + self.bias
+                )
         return feature_activations @ self.feature_vectors + self.bias

{sae_lens-6.29.0 → sae_lens-6.30.1}/sae_lens/synthetic/hierarchy.py

@@ -147,6 +147,14 @@ class _SparseHierarchyData:
     # Total number of ME groups
     num_groups: int
 
+    # Sparse COO support: Feature-to-parent mapping
+    # feat_to_parent[f] = parent feature index, or -1 if root/no parent
+    feat_to_parent: torch.Tensor | None = None  # [num_features]
+
+    # Sparse COO support: Feature-to-ME-group mapping
+    # feat_to_me_group[f] = group index, or -1 if not in any ME group
+    feat_to_me_group: torch.Tensor | None = None  # [num_features]
+
 
 def _build_sparse_hierarchy(
     roots: Sequence[HierarchyNode],
@@ -232,7 +240,11 @@ def _build_sparse_hierarchy(
             me_indices = torch.empty(0, dtype=torch.long)
 
         level_data.append(
-            _LevelData(features=feats, parents=parents, me_group_indices=me_indices)
+            _LevelData(
+                features=feats,
+                parents=parents,
+                me_group_indices=me_indices,
+            )
         )
 
     # Build group siblings and parents tensors
@@ -254,12 +266,30 @@ def _build_sparse_hierarchy(
         me_group_parents = torch.empty(0, dtype=torch.long)
         num_groups = 0
 
+    # Build sparse COO support: feat_to_parent and feat_to_me_group mappings
+    # First determine num_features (max feature index + 1)
+    all_features = [f for f, _, _ in feature_info]
+    num_features = max(all_features) + 1 if all_features else 0
+
+    # Build feature-to-parent mapping
+    feat_to_parent = torch.full((num_features,), -1, dtype=torch.long)
+    for feat, parent, _ in feature_info:
+        feat_to_parent[feat] = parent
+
+    # Build feature-to-ME-group mapping
+    feat_to_me_group = torch.full((num_features,), -1, dtype=torch.long)
+    for g_idx, (_, _, siblings) in enumerate(me_groups):
+        for sib in siblings:
+            feat_to_me_group[sib] = g_idx
+
     return _SparseHierarchyData(
         level_data=level_data,
         me_group_siblings=me_group_siblings,
         me_group_sizes=me_group_sizes,
         me_group_parents=me_group_parents,
         num_groups=num_groups,
+        feat_to_parent=feat_to_parent,
+        feat_to_me_group=feat_to_me_group,
     )
 
 
@@ -396,8 +426,9 @@ def _apply_me_for_groups(
     # Random selection for winner
     # Use -1e9 instead of -inf to avoid creating a tensor (torch.tensor(-float("inf")))
     # on every call. Since random scores are in [0,1], -1e9 is effectively -inf for argmax.
+    _INACTIVE_SCORE = -1e9
     random_scores = torch.rand(num_conflicts, max_siblings, device=device)
-    random_scores[~conflict_active] = -1e9
+    random_scores[~conflict_active] = _INACTIVE_SCORE
 
     winner_idx = random_scores.argmax(dim=1)
 
@@ -420,6 +451,275 @@ def _apply_me_for_groups(
     activations[deact_batch, deact_feat] = 0
 
 
+# ---------------------------------------------------------------------------
+# Sparse COO hierarchy implementation
+# ---------------------------------------------------------------------------
+
+
+def _apply_hierarchy_sparse_coo(
+    sparse_tensor: torch.Tensor,
+    sparse_data: _SparseHierarchyData,
+) -> torch.Tensor:
+    """
+    Apply hierarchy constraints to a sparse COO tensor.
+
+    This is the sparse analog of _apply_hierarchy_sparse. It processes
+    level-by-level, applying parent deactivation then mutual exclusion.
+    """
+    if sparse_tensor._nnz() == 0:
+        return sparse_tensor
+
+    sparse_tensor = sparse_tensor.coalesce()
+
+    for level_data in sparse_data.level_data:
+        # Step 1: Apply parent deactivation for features at this level
+        if level_data.features.numel() > 0:
+            sparse_tensor = _apply_parent_deactivation_coo(
+                sparse_tensor, level_data, sparse_data
+            )
+
+        # Step 2: Apply ME for groups whose parent is at this level
+        if level_data.me_group_indices.numel() > 0:
+            sparse_tensor = _apply_me_coo(
+                sparse_tensor, level_data.me_group_indices, sparse_data
+            )
+
+    return sparse_tensor
+
+
+def _apply_parent_deactivation_coo(
+    sparse_tensor: torch.Tensor,
+    level_data: _LevelData,
+    sparse_data: _SparseHierarchyData,
+) -> torch.Tensor:
+    """
+    Remove children from sparse COO tensor when their parent is inactive.
+
+    Uses searchsorted for efficient membership testing of parent activity.
+    """
+    if sparse_tensor._nnz() == 0 or level_data.features.numel() == 0:
+        return sparse_tensor
+
+    sparse_tensor = sparse_tensor.coalesce()
+    indices = sparse_tensor.indices()  # [2, nnz]
+    values = sparse_tensor.values()  # [nnz]
+    batch_indices = indices[0]
+    feat_indices = indices[1]
+
+    _, num_features = sparse_tensor.shape
+    device = sparse_tensor.device
+    nnz = indices.shape[1]
+
+    # Build set of active (batch, feature) pairs for efficient lookup
+    # Encode as: batch_idx * num_features + feat_idx
+    active_pairs = batch_indices * num_features + feat_indices
+    active_pairs_sorted, _ = active_pairs.sort()
+
+    # Use the precomputed feat_to_parent mapping
+    assert sparse_data.feat_to_parent is not None
+    hierarchy_num_features = sparse_data.feat_to_parent.numel()
+
+    # Handle features outside the hierarchy (they have no parent, pass through)
+    in_hierarchy = feat_indices < hierarchy_num_features
+    parent_of_feat = torch.full((nnz,), -1, dtype=torch.long, device=device)
+    parent_of_feat[in_hierarchy] = sparse_data.feat_to_parent[
+        feat_indices[in_hierarchy]
+    ]
+
+    # Find entries that have a parent (parent >= 0 means this feature has a parent)
+    has_parent = parent_of_feat >= 0
+
+    if not has_parent.any():
+        return sparse_tensor
+
+    # For entries with parents, check if parent is active
+    child_entry_indices = torch.where(has_parent)[0]
+    child_batch = batch_indices[has_parent]
+    child_parents = parent_of_feat[has_parent]
+
+    # Look up parent activity using searchsorted
+    parent_pairs = child_batch * num_features + child_parents
+    search_pos = torch.searchsorted(active_pairs_sorted, parent_pairs)
+    search_pos = search_pos.clamp(max=active_pairs_sorted.numel() - 1)
+    parent_active = active_pairs_sorted[search_pos] == parent_pairs
+
+    # Handle empty case
+    if active_pairs_sorted.numel() == 0:
+        parent_active = torch.zeros_like(parent_pairs, dtype=torch.bool)
+
+    # Build keep mask: keep entry if it's a root OR its parent is active
+    keep_mask = torch.ones(nnz, dtype=torch.bool, device=device)
+    keep_mask[child_entry_indices[~parent_active]] = False
+
+    if keep_mask.all():
+        return sparse_tensor
+
+    return torch.sparse_coo_tensor(
+        indices[:, keep_mask],
+        values[keep_mask],
+        sparse_tensor.shape,
+        device=device,
+        dtype=sparse_tensor.dtype,
+    )
+
+
+def _apply_me_coo(
+    sparse_tensor: torch.Tensor,
+    group_indices: torch.Tensor,
+    sparse_data: _SparseHierarchyData,
+) -> torch.Tensor:
+    """
+    Apply mutual exclusion to sparse COO tensor.
+
+    For each ME group with multiple active siblings in the same batch,
+    randomly selects one winner and removes the rest.
+    """
+    if sparse_tensor._nnz() == 0 or group_indices.numel() == 0:
+        return sparse_tensor
+
+    sparse_tensor = sparse_tensor.coalesce()
+    indices = sparse_tensor.indices()  # [2, nnz]
+    values = sparse_tensor.values()  # [nnz]
+    batch_indices = indices[0]
+    feat_indices = indices[1]
+
+    _, num_features = sparse_tensor.shape
+    device = sparse_tensor.device
+    nnz = indices.shape[1]
+
+    # Use precomputed feat_to_me_group mapping
+    assert sparse_data.feat_to_me_group is not None
+    hierarchy_num_features = sparse_data.feat_to_me_group.numel()
+
+    # Handle features outside the hierarchy (they are not in any ME group)
+    in_hierarchy = feat_indices < hierarchy_num_features
+    me_group_of_feat = torch.full((nnz,), -1, dtype=torch.long, device=device)
+    me_group_of_feat[in_hierarchy] = sparse_data.feat_to_me_group[
+        feat_indices[in_hierarchy]
+    ]
+
+    # Find entries that belong to ME groups we're processing (vectorized)
+    in_relevant_group = torch.isin(me_group_of_feat, group_indices)
+
+    if not in_relevant_group.any():
+        return sparse_tensor
+
+    # Get the ME entries
+    me_entry_indices = torch.where(in_relevant_group)[0]
+    me_batch = batch_indices[in_relevant_group]
+    me_group = me_group_of_feat[in_relevant_group]
+
+    # Check parent activity for ME groups (only apply ME if parent is active)
+    me_group_parents = sparse_data.me_group_parents[me_group]
+    has_parent = me_group_parents >= 0
+
+    if has_parent.any():
+        # Build active pairs for parent lookup
+        active_pairs = batch_indices * num_features + feat_indices
+        active_pairs_sorted, _ = active_pairs.sort()
+
+        parent_pairs = (
+            me_batch[has_parent] * num_features + me_group_parents[has_parent]
+        )
+        search_pos = torch.searchsorted(active_pairs_sorted, parent_pairs)
+        search_pos = search_pos.clamp(max=active_pairs_sorted.numel() - 1)
+        parent_active_for_has_parent = active_pairs_sorted[search_pos] == parent_pairs
+
+        # Build full parent_active mask
+        parent_active = torch.ones(
+            me_entry_indices.numel(), dtype=torch.bool, device=device
+        )
+        parent_active[has_parent] = parent_active_for_has_parent
+
+        # Filter to only ME entries where parent is active
+        valid_me = parent_active
+        me_entry_indices = me_entry_indices[valid_me]
+        me_batch = me_batch[valid_me]
+        me_group = me_group[valid_me]
+
+    if me_entry_indices.numel() == 0:
+        return sparse_tensor
+
+    # Encode (batch, group) pairs
+    num_groups = sparse_data.num_groups
+    batch_group_pairs = me_batch * num_groups + me_group
+
+    # Find unique (batch, group) pairs and count occurrences
+    unique_bg, inverse, counts = torch.unique(
+        batch_group_pairs, return_inverse=True, return_counts=True
+    )
+
+    # Only process pairs with count > 1 (conflicts)
+    has_conflict = counts > 1
+
+    if not has_conflict.any():
+        return sparse_tensor
+
+    # For efficiency, we process all conflicts together
+    # Assign random scores to each ME entry
+    random_scores = torch.rand(me_entry_indices.numel(), device=device)
+
+    # For each (batch, group) pair, we want the entry with highest score to be winner
+    # Use scatter_reduce to find max score per (batch, group)
+    bg_to_dense = torch.zeros(unique_bg.numel(), dtype=torch.long, device=device)
+    bg_to_dense[has_conflict.nonzero(as_tuple=True)[0]] = torch.arange(
+        has_conflict.sum(), device=device
+    )
+
+    # Map each ME entry to its dense conflict index
+    entry_has_conflict = has_conflict[inverse]
+
+    if not entry_has_conflict.any():
+        return sparse_tensor
+
+    conflict_entries_mask = entry_has_conflict
+    conflict_entry_indices = me_entry_indices[conflict_entries_mask]
+    conflict_random_scores = random_scores[conflict_entries_mask]
+    conflict_inverse = inverse[conflict_entries_mask]
+    conflict_dense_idx = bg_to_dense[conflict_inverse]
+
+    # Vectorized winner selection using sorting
+    # Sort entries by (group_idx, -random_score) so highest score comes first per group
+    # Use group * 2 - score to sort by group ascending, then score descending
+    sort_keys = conflict_dense_idx.float() * 2.0 - conflict_random_scores
+    sorted_order = sort_keys.argsort()
+    sorted_dense_idx = conflict_dense_idx[sorted_order]
+
+    # Find first entry of each group in sorted order (these are winners)
+    group_starts = torch.cat(
+        [
+            torch.tensor([True], device=device),
+            sorted_dense_idx[1:] != sorted_dense_idx[:-1],
+        ]
+    )
+
+    # Winners are entries at group starts in sorted order
+    winner_positions_in_sorted = torch.where(group_starts)[0]
+    winner_original_positions = sorted_order[winner_positions_in_sorted]
+
+    # Create winner mask (vectorized)
+    is_winner = torch.zeros(
+        conflict_entry_indices.numel(), dtype=torch.bool, device=device
+    )
+    is_winner[winner_original_positions] = True
+
+    # Build keep mask (vectorized)
+    keep_mask = torch.ones(nnz, dtype=torch.bool, device=device)
+    loser_entry_indices = conflict_entry_indices[~is_winner]
+    keep_mask[loser_entry_indices] = False
+
+    if keep_mask.all():
+        return sparse_tensor
+
+    return torch.sparse_coo_tensor(
+        indices[:, keep_mask],
+        values[keep_mask],
+        sparse_tensor.shape,
+        device=device,
+        dtype=sparse_tensor.dtype,
+    )
+
+
 @torch.no_grad()
 def hierarchy_modifier(
     roots: Sequence[HierarchyNode] | HierarchyNode,
@@ -475,12 +775,24 @@ def hierarchy_modifier(
                 me_group_sizes=sparse_data.me_group_sizes.to(device),
                 me_group_parents=sparse_data.me_group_parents.to(device),
                 num_groups=sparse_data.num_groups,
+                feat_to_parent=(
+                    sparse_data.feat_to_parent.to(device)
+                    if sparse_data.feat_to_parent is not None
+                    else None
+                ),
+                feat_to_me_group=(
+                    sparse_data.feat_to_me_group.to(device)
+                    if sparse_data.feat_to_me_group is not None
+                    else None
+                ),
             )
         return device_cache[device]
 
     def modifier(activations: torch.Tensor) -> torch.Tensor:
         device = activations.device
         cached = _get_sparse_for_device(device)
+        if activations.is_sparse:
+            return _apply_hierarchy_sparse_coo(activations, cached)
         return _apply_hierarchy_sparse(activations, cached)
 
     return modifier

{sae_lens-6.29.0 → sae_lens-6.30.1}/sae_lens/synthetic/training.py

@@ -23,6 +23,8 @@ def train_toy_sae(
     device: str | torch.device = "cpu",
     n_snapshots: int = 0,
     snapshot_fn: Callable[[SAETrainer[Any, Any]], None] | None = None,
+    autocast_sae: bool = False,
+    autocast_data: bool = False,
 ) -> None:
     """
     Train an SAE on synthetic activations from a feature dictionary.
@@ -46,6 +48,8 @@ def train_toy_sae(
         snapshot_fn: Callback function called at each snapshot point. Receives
             the SAETrainer instance, allowing access to the SAE, training step,
             and other training state. Required if n_snapshots > 0.
+        autocast_sae: Whether to autocast the SAE to bfloat16. Only recommend for large SAEs on CUDA
+        autocast_data: Whether to autocast the activations generator and feature dictionary to bfloat16. Only recommend for large data on CUDA.
     """
 
     device_str = str(device) if isinstance(device, torch.device) else device
@@ -55,6 +59,7 @@ def train_toy_sae(
         feature_dict=feature_dict,
         activations_generator=activations_generator,
         batch_size=batch_size,
+        autocast=autocast_data,
     )
 
     # Create trainer config
@@ -64,7 +69,7 @@ def train_toy_sae(
         save_final_checkpoint=False,
         total_training_samples=training_samples,
         device=device_str,
-        autocast=False,
+        autocast=autocast_sae,
         lr=lr,
         lr_end=lr,
         lr_scheduler_name="constant",
@@ -119,6 +124,7 @@ class SyntheticActivationIterator(Iterator[torch.Tensor]):
         feature_dict: FeatureDictionary,
         activations_generator: ActivationGenerator,
         batch_size: int,
+        autocast: bool = False,
     ):
         """
         Create a new SyntheticActivationIterator.
@@ -127,16 +133,23 @@ class SyntheticActivationIterator(Iterator[torch.Tensor]):
             feature_dict: The feature dictionary to use for generating hidden activations
             activations_generator: Generator that produces feature activations
             batch_size: Number of samples per batch
+            autocast: Whether to autocast the activations generator and feature dictionary to bfloat16.
         """
         self.feature_dict = feature_dict
         self.activations_generator = activations_generator
         self.batch_size = batch_size
+        self.autocast = autocast
 
     @torch.no_grad()
     def next_batch(self) -> torch.Tensor:
         """Generate the next batch of hidden activations."""
-        features = self.activations_generator(self.batch_size)
-        return self.feature_dict(features)
+        with torch.autocast(
+            device_type=self.feature_dict.feature_vectors.device.type,
+            dtype=torch.bfloat16,
+            enabled=self.autocast,
+        ):
+            features = self.activations_generator(self.batch_size)
+            return self.feature_dict(features)
 
     def __iter__(self) -> "SyntheticActivationIterator":
         return self