Relax tolerance in Qwen Scope loader test to fix flaky CI

tests/loading/test_pretrained_sae_loaders.py

@@ -2364,16 +2364,25 @@ def mock_hf_hub_download(*args: Any, **kwargs: Any) -> str:  # noqa: ARG001
     # The transpose must preserve the underlying linear operation: in the
     # native Qwen Scope formula, pre_acts = residual @ raw_W_enc.T + b_enc.
     # In SAELens, pre_acts = residual @ W_enc + b_enc. They must agree.
+    #
+    # The loader stores W_enc/W_dec as `.T.contiguous()`, so each matmul below
+    # multiplies a contiguous tensor while the reference multiplies a transposed
+    # view. Over these large reduction dims (2048 and 32768) the two layouts can
+    # accumulate float32 in a different order on some BLAS backends, so the
+    # results agree only up to rounding. Use tolerances that absorb that noise
+    # while still catching a real loader regression (which would be orders of
+    # magnitude off); the default rtol=1e-5/atol=1e-8 is too tight here and
+    # causes intermittent CI failures.
     residual = torch.randn(2, 7, d_in)
     expected_pre_acts = residual @ raw_W_enc.T + raw_b_enc
     actual_pre_acts = residual @ state_dict["W_enc"] + state_dict["b_enc"]
-    assert_close(actual_pre_acts, expected_pre_acts)
+    assert_close(actual_pre_acts, expected_pre_acts, rtol=1e-3, atol=1e-2)
 
     # Same for the decoder: native is feats @ raw_W_dec.T + b_dec, SAELens is feats @ W_dec + b_dec.
     feats = torch.randn(2, 7, d_sae)
     expected_recon = feats @ raw_W_dec.T + raw_b_dec
     actual_recon = feats @ state_dict["W_dec"] + state_dict["b_dec"]
-    assert_close(actual_recon, expected_recon)
+    assert_close(actual_recon, expected_recon, rtol=1e-3, atol=1e-2)
 
 
 def test_qwen_scope_sae_loads_via_from_pretrained(