Harden Qwen packed hybrid runtime contract

Verify native packed-state dataflow, operand layouts, and Phase 1 attributes before emitting runtime metadata. Route paged context explicitly, expand malformed-graph coverage, and fix the CI dependency and stale VibeVoice streaming denylist regressions.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: titaiwang <titaiwang@microsoft.com>

Author: titaiwangms

src/mobius/components/_attention.py

@@ -593,9 +593,18 @@ def forward(
         position_embeddings: tuple,
         past_key_value: tuple | None = None,
         static_cache: StaticCacheState | None = None,
+        paged_state: PagedAttentionState | None = None,
     ):
+        if paged_state is not None:
+            if past_key_value is not None or static_cache is not None:
+                raise ValueError(
+                    "Paged attention state cannot be combined with dense cache state"
+                )
+            return self.forward_paged(op, hidden_states, position_embeddings, paged_state)
         if isinstance(past_key_value, PagedAttentionState):
-            return self.forward_paged(op, hidden_states, position_embeddings, past_key_value)
+            raise TypeError(
+                "PagedAttentionState must be passed through paged_state, not past_key_value"
+            )
         # Q projection (doubled) → split into Q and gate per head
         q_gate = self.q_proj(op, hidden_states)
         # Reshape to per-head view so split separates Q/gate within each head

src/mobius/components/_attention_test.py

@@ -15,6 +15,7 @@
     make_config,
 )
 from mobius.components._attention import Attention, FusedQKVAttention, Qwen35Attention
+from mobius.components._paged_attention import PagedAttentionState
 
 
 class TestAttention:
@@ -215,6 +216,22 @@ def test_forward_builds_graph(self):
         assert count_op_type(graph, "Attention") >= 1
         assert count_op_type(graph, "Sigmoid") >= 1
 
+    def test_legacy_paged_state_fails_fast(self):
+        config = make_config(partial_rotary_factor=0.5)
+        attn = Qwen35Attention(config)
+        builder, op, _ = create_test_builder()
+        hidden = create_test_input(builder, "hidden", [1, 8, 64])
+        paged_state = PagedAttentionState(*([None] * 6))
+
+        with pytest.raises(TypeError, match="paged_state"):
+            attn(
+                op,
+                hidden,
+                attention_bias=None,
+                position_embeddings=(),
+                past_key_value=paged_state,
+            )
+
 
 class TestGQAContextDispatch:
     """Tests for the GQAContext direct GroupQueryAttention emission path."""

src/mobius/components/_gated_deltanet.py

@@ -348,6 +348,11 @@ def forward_paged(
         z = self.in_proj_z(op, hidden_states)  # (N, Hv*Dv)
         raw_b = op.Cast(self.in_proj_b(op, hidden_states), to=ir.DataType.FLOAT)
         raw_a = op.Cast(self.in_proj_a(op, hidden_states), to=ir.DataType.FLOAT)
+        # The generic builder does not infer Cast output types until the
+        # production optimization pipeline runs. Keep the native ABI explicit
+        # even for directly-built graphs.
+        raw_b.type = ir.TensorType(ir.DataType.FLOAT)
+        raw_a.type = ir.TensorType(ir.DataType.FLOAT)
 
         # Invoke the child module so nn realizes its qualified conv1d.weight.
         conv_out, new_conv_state = self.conv1d(