inference: interp-engine 1.8.0 -> 1.11.0, steer with SteerMethod

apps/inference/neuronpedia_inference/endpoints/steer/completion.py

@@ -13,6 +13,7 @@
     ProjectionCapSpec,
     SteeringOp,
     SteeringSpec,
+    SteerMethod,
     SteerSpec,
     VLLMModel,
 )
@@ -255,7 +256,9 @@ def _feature_to_steerspec(
     layer = address.layer
     vector = torch.tensor(feature.steering_vector, dtype=torch.float32)
     coeff = settings.strength_multiplier * feature.strength
-    method = "orthogonal" if settings.steer_method == NPSteerMethod.ORTHOGONAL_DECOMP else "additive"
+    method = (
+        SteerMethod.ORTHOGONAL if settings.steer_method == NPSteerMethod.ORTHOGONAL_DECOMP else SteerMethod.ADDITIVE
+    )
 
     if name == "resid_post":
         point, spec_layer = "resid_post", layer

apps/inference/pyproject.toml

@@ -55,7 +55,7 @@ dependencies = [
   # 1.8.0 adds `quantization=` and `kv_cache_dtype=` to load_model, which start.py's
   # --quantization / --kv_cache_dtype pass through: FP8 on load is what fits Llama-3.3-70B
   # on one 96 GB card. Nothing else on the surface this service calls changed.
-  "interp-engine[vllm,quant]==1.8.0",
+  "interp-engine[vllm,quant]==1.11.0",
 ]
 
 [dependency-groups]

apps/inference/tests/unit/test_vllm_capture_contract.py

@@ -142,8 +142,19 @@ def test_unknown_hidden_size_skips_the_check(self):
 
 
 class TestAttentionRecomputeSharding:
-    def test_refuses_under_tensor_parallelism(self):
-        # q/k/v are head-sharded, and `dims` describes the whole model, so the reshape
-        # downstream would fail on the element count. Say why instead.
-        with pytest.raises(RuntimeError, match="tensor_parallel_size=4"):
-            vllm_backend.recompute_attn_from_payloads({}, [0], {}, 4)
+    def test_tensor_parallelism_is_accepted(self):
+        # The worker all-gathers q/k/v across ranks at collect, so rank 0's payload holds every
+        # head at any TP size and the recompute no longer refuses on the argument alone. The pod-level
+        # gate for sharded attention is inference's own (`vllm_attention_unsupported_reason`).
+        assert vllm_backend.recompute_attn_from_payloads({}, [], {}, 4) == {}
+
+    def test_a_shard_that_gets_through_is_named(self):
+        # `dims` describes the whole model, so a head-sharded capture is not a whole number of heads
+        # wide. The reshape would fail on the element count; the engine says what it saw instead.
+        with pytest.raises(ValueError, match="tensor-parallel shard"):
+            vllm_backend._heads_in(torch.zeros(19, 96), head_dim=64, which="q", layer=0)
+
+    def test_a_shard_that_divides_is_caught_against_the_config(self):
+        # Half the heads at the same head dim divides cleanly, so only the config's count exposes it.
+        with pytest.raises(ValueError, match="the config says 32"):
+            vllm_backend._heads_in(torch.zeros(19, 16 * 64), head_dim=64, which="q", layer=0, expected=32)