interp-engine 1.13.1__tar.gz → 1.13.2__tar.gz

{interp_engine-1.13.1 → interp_engine-1.13.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: interp-engine
-Version: 1.13.1
+Version: 1.13.2
 Summary: A fast, standardized, and easy to use interpretability engine.
 Project-URL: Homepage, https://github.com/decoderesearch/interp-engine
 Project-URL: Repository, https://github.com/decoderesearch/interp-engine

{interp_engine-1.13.1 → interp_engine-1.13.2}/interp_engine/vllm_capture/mhc.py

@@ -648,9 +648,13 @@ class _MhcTaps:
         detectably different from no steer at all. So the rows the recorder actually changed take the
         re-run and every other row keeps the kernel's own output, which is also why a zero delta needs
         no re-run: no row changed.
+
+        Inside a graph capture the host cannot read ``changed``, so the re-run is always recorded and
+        the row mask alone decides at replay.
         """
         changed = (edited != stack).flatten(start_dim=1).any(dim=1)
-        if not bool(changed.any()):
+        capturing = stack.is_cuda and torch.cuda.is_current_stream_capturing()
+        if not capturing and not bool(changed.any()):
             return fused
         bound = signature.bind(*args, **kwargs)
         bound.apply_defaults()

{interp_engine-1.13.1 → interp_engine-1.13.2}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "interp-engine"
-version = "1.13.1"
+version = "1.13.2"
 description = "A fast, standardized, and easy to use interpretability engine."
 readme = "README.md"
 license = "Apache-2.0"

{interp_engine-1.13.1 → interp_engine-1.13.2}/tests/test_vllm_hyper_connections.py

@@ -1087,6 +1087,36 @@ def test_a_stream_stack_steer_reaches_the_collapse_the_very_same_kernel_computes
     assert not torch.allclose(trunk.layers[1].ffn_input_stack, unsteered_stack), "nor did anything downstream"
 
 
+@pytest.mark.gpu
+@pytest.mark.skipif(not torch.cuda.is_available(), reason="records a CUDA graph")
+def test_a_stream_stack_steer_records_into_a_cuda_graph():
+    """vLLM records FULL decode graphs with plain capture, where the host cannot read a tensor.
+
+    Static writes edit the stack while the graph is recorded, so the re-run decision may not read
+    ``changed`` on the host there. It did, and the DeepSeek-V4 static engine failed at start-up.
+    """
+    trunk = _HyperTrunk(n_layers=2).cuda()
+    positions, input_ids = POSITIONS.cuda(), INPUT_IDS.cuda()
+    _, handles = _steering_demux(trunk, {Address("resid_streams", 0): _bump(0.5)})
+    try:
+        with torch.no_grad():
+            eager = trunk(positions, input_ids).clone()
+            side = torch.cuda.Stream()
+            side.wait_stream(torch.cuda.current_stream())
+            with torch.cuda.stream(side):
+                trunk(positions, input_ids)
+            torch.cuda.current_stream().wait_stream(side)
+            graph = torch.cuda.CUDAGraph()
+            with torch.cuda.graph(graph):
+                out = trunk(positions, input_ids)
+            graph.replay()
+            torch.cuda.synchronize()
+    finally:
+        for handle in handles:
+            handle.remove()
+    torch.testing.assert_close(out, eager)
+
+
 def test_only_the_layer_that_is_actually_steered_pays_for_a_second_pre_phase():
     """The extra kernel is per call, so an unsteered forward must not pay for it -- and neither must
     the layers of a steered forward that nobody asked about.