fix: warn rather than crash if sae is unnormalized

sae_bench/evals/meta_structure/eval_output.py

@@ -53,7 +53,7 @@ class MetaStructureEvalOutput(
     eval_id: str
     datetime_epoch_millis: int
     eval_result_metrics: MetaStructureMetricCategories
-    eval_result_details: list[BaseResultDetail] | None = None
+    eval_result_details: list[BaseResultDetail] | None = None  # pyright: ignore[reportIncompatibleVariableOverride]
     eval_type_id: str = Field(
         default=EVAL_TYPE_ID_META_STRUCTURE,
         title="Eval Type ID",

sae_bench/sae_bench_utils/general_utils.py

@@ -3,6 +3,7 @@
 import random
 import re
 import time
+import warnings
 from typing import Any, Callable
 
 import pandas as pd
@@ -124,10 +125,10 @@ def check_decoder_norms(W_dec: torch.Tensor) -> bool:
         return True
     else:
         max_diff = torch.max(torch.abs(norms - torch.ones_like(norms)))
-        print(f"Decoder weights are not normalized. Max diff: {max_diff.item()}")
-        raise ValueError(
-            "Decoder weights are not normalized. Refer to base_sae.py and relu_sae.py for more info."
+        warnings.warn(
+            f"Decoder weights are not normalized. Max diff: {max_diff.item()}. Refer to base_sae.py and relu_sae.py for more info."
         )
+        return False
 
 
 def load_and_format_sae(

tests/acceptance/test_meta_structure.py

@@ -54,9 +54,17 @@ def test_meta_structure_eval_matches_fixture(tmp_path):
         actual = json.load(f)
     actual_metrics = _load_metrics(actual)
 
-    assert pytest.approx(
-        expected_metrics["decoder_fraction_variance_explained"], rel=1e-2,
-    ) == actual_metrics["decoder_fraction_variance_explained"]
-    assert pytest.approx(
-        expected_metrics["final_reconstruction_mse"], rel=1e-2,
-    ) == actual_metrics["final_reconstruction_mse"]
+    assert (
+        pytest.approx(
+            expected_metrics["decoder_fraction_variance_explained"],
+            rel=1e-2,
+        )
+        == actual_metrics["decoder_fraction_variance_explained"]
+    )
+    assert (
+        pytest.approx(
+            expected_metrics["final_reconstruction_mse"],
+            rel=1e-2,
+        )
+        == actual_metrics["final_reconstruction_mse"]
+    )