sae-lens 6.16.0__tar.gz → 6.16.2__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.
Files changed (40) hide show
  1. {sae_lens-6.16.0 → sae_lens-6.16.2}/PKG-INFO +1 -1
  2. {sae_lens-6.16.0 → sae_lens-6.16.2}/pyproject.toml +1 -1
  3. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/__init__.py +1 -1
  4. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/saes/batchtopk_sae.py +3 -1
  5. {sae_lens-6.16.0 → sae_lens-6.16.2}/LICENSE +0 -0
  6. {sae_lens-6.16.0 → sae_lens-6.16.2}/README.md +0 -0
  7. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/analysis/__init__.py +0 -0
  8. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/analysis/hooked_sae_transformer.py +0 -0
  9. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/analysis/neuronpedia_integration.py +0 -0
  10. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/cache_activations_runner.py +0 -0
  11. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/config.py +0 -0
  12. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/constants.py +0 -0
  13. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/evals.py +0 -0
  14. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/llm_sae_training_runner.py +0 -0
  15. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/load_model.py +0 -0
  16. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/loading/__init__.py +0 -0
  17. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/loading/pretrained_sae_loaders.py +0 -0
  18. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/loading/pretrained_saes_directory.py +0 -0
  19. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/pretokenize_runner.py +0 -0
  20. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/pretrained_saes.yaml +0 -0
  21. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/registry.py +0 -0
  22. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/saes/__init__.py +0 -0
  23. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/saes/gated_sae.py +0 -0
  24. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/saes/jumprelu_sae.py +0 -0
  25. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/saes/matryoshka_batchtopk_sae.py +0 -0
  26. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/saes/sae.py +0 -0
  27. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/saes/standard_sae.py +0 -0
  28. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/saes/topk_sae.py +0 -0
  29. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/saes/transcoder.py +0 -0
  30. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/tokenization_and_batching.py +0 -0
  31. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/training/__init__.py +0 -0
  32. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/training/activation_scaler.py +0 -0
  33. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/training/activations_store.py +0 -0
  34. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/training/mixing_buffer.py +0 -0
  35. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/training/optim.py +0 -0
  36. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/training/sae_trainer.py +0 -0
  37. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/training/types.py +0 -0
  38. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/training/upload_saes_to_huggingface.py +0 -0
  39. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/tutorial/tsea.py +0 -0
  40. {sae_lens-6.16.0 → sae_lens-6.16.2}/sae_lens/util.py +0 -0
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: sae-lens
3
- Version: 6.16.0
3
+ Version: 6.16.2
4
4
  Summary: Training and Analyzing Sparse Autoencoders (SAEs)
5
5
  License: MIT
6
6
  License-File: LICENSE
@@ -1,6 +1,6 @@
1
1
  [tool.poetry]
2
2
  name = "sae-lens"
3
- version = "6.16.0"
3
+ version = "6.16.2"
4
4
  description = "Training and Analyzing Sparse Autoencoders (SAEs)"
5
5
  authors = ["Joseph Bloom"]
6
6
  readme = "README.md"
@@ -1,5 +1,5 @@
1
1
  # ruff: noqa: E402
2
- __version__ = "6.16.0"
2
+ __version__ = "6.16.2"
3
3
 
4
4
  import logging
5
5
 
@@ -23,7 +23,9 @@ class BatchTopK(nn.Module):
23
23
  def forward(self, x: torch.Tensor) -> torch.Tensor:
24
24
  acts = x.relu()
25
25
  flat_acts = acts.flatten()
26
- acts_topk_flat = torch.topk(flat_acts, int(self.k * acts.shape[0]), dim=-1)
26
+ # Calculate total number of samples across all non-feature dimensions
27
+ num_samples = acts.shape[:-1].numel()
28
+ acts_topk_flat = torch.topk(flat_acts, int(self.k * num_samples), dim=-1)
27
29
  return (
28
30
  torch.zeros_like(flat_acts)
29
31
  .scatter(-1, acts_topk_flat.indices, acts_topk_flat.values)
File without changes
File without changes
File without changes
File without changes
File without changes