fix(evo2-sae): drop stale test_clamp_math + leaked server deps from the engine PR

- remove test_clamp_math: it called the deleted Evo2SAE._clamp_hook (we unified onto
  sae.steering.clamp_hook); the delta-clamp math is covered in sae/tests/test_steering.py
- pyproject: drop fastapi/uvicorn/pandas — the engine imports none of them; fastapi+uvicorn
  move to the serve PR (#1637), pandas was unused

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Signed-off-by: Polina Binder <pbinder@nvidia.com>

Author: polinabinder1

bionemo-recipes/interpretability/sparse_autoencoders/recipes/evo2/pyproject.toml

@@ -13,13 +13,11 @@ dependencies = [
     "torch>=2.0",
     "numpy>=1.20",
     "pyarrow>=23.0.0",
-    "fastapi>=0.110",
-    "uvicorn>=0.29",
-    "pandas>=1.5",
 ]
 
-# The `evo2_sae` package (src/) holds the live inference engine + server + CLI;
-# scripts/ (extract, train) are standalone entry points alongside it.
+# The `evo2_sae` package (src/) holds the live inference engine + steering hook;
+# scripts/ (extract, train) are standalone entry points alongside it. The FastAPI
+# server + CLI (and their fastapi/uvicorn deps) are added by the serve PR (#1637).
 [tool.setuptools.packages.find]
 where = ["src"]
 

bionemo-recipes/interpretability/sparse_autoencoders/recipes/evo2/tests/test_steering.py

@@ -29,35 +29,9 @@
 import torch
 
 
-# --------------------------------------------------------------------- CPU: clamp math
-def test_clamp_math():
-    """The decode-only hook applies h <- h + Σ_f (target_f - relu((h-pre_bias)@enc_f + b_f))·dec_f."""
-    from evo2_sae.core import Evo2SAE
-
-    H, F = 4, 2
-    torch.manual_seed(0)
-    pre_bias = torch.randn(H)
-    enc = torch.randn(F, H)
-    bias = torch.randn(F)
-    dec = torch.randn(H, F)
-    specs = [(enc[f], float(bias[f]), dec[:, f], float(f + 1)) for f in range(F)]
-
-    eng = Evo2SAE.__new__(Evo2SAE)  # bare instance — exercise the hook only
-    hook = eng._clamp_hook(specs, pre_bias)
-
-    h = torch.randn(1, 1, H)  # one decode token: [S=1, B=1, H]
-    out = hook(None, None, h)
-
-    xc = h - pre_bias
-    expected = h.clone()
-    for enc_f, b_f, dec_f, target in specs:
-        a = torch.relu(xc @ enc_f + b_f)
-        expected = expected + (target - a).unsqueeze(-1) * dec_f
-    torch.testing.assert_close(out, expected)
-
-    # prefill (S>1) must be left untouched (continuation-only steering)
-    prefill = torch.randn(5, 1, H)
-    assert torch.equal(hook(None, None, prefill), prefill)
+# The delta-clamp math + decode-only/prefill behavior is covered against the production
+# hook in sae/tests/test_steering.py (sae.steering.clamp_hook). This file covers the
+# engine-level input guards (CPU) and the live steering smoke tests (GPU).
 
 
 # --------------------------------------------------- CPU: steering input guards (no model)