Set QMoE raw weight prepack mode

Emit the new weights_prepacked=0 QMoE attribute for raw int expert weights now that ORT can prepack them internally. Add focused assertions for both component-built and rewrite-fused QMoE nodes.

Signed-off-by: GitHub <noreply@github.com>

Co-authored-by: justinchuby <11205048+justinchuby@users.noreply.github.com>

Author: Copilot

src/mobius/components/_moe.py

@@ -366,6 +366,8 @@ def _qmoe_forward(self, op: OpBuilder, hidden_states: ir.Value):
             expert_weight_bits=quantization.bits,
             block_size=quantization.group_size,
             swiglu_fusion=2,
+            quant_type="int",
+            weights_prepacked=0,
             _domain="com.microsoft",
         )
         if output_scale != 1.0:  # noqa: RUF069

src/mobius/components/_moe_test.py

@@ -160,6 +160,8 @@ def test_int4_olive_moe_emits_expert_major_qmoe(self):
         assert qmoe.attributes["block_size"].value == 32
         assert qmoe.attributes["swiglu_fusion"].value == 2
         assert qmoe.attributes["expert_weight_bits"].value == 4
+        assert qmoe.attributes["quant_type"].value == "int"
+        assert qmoe.attributes["weights_prepacked"].value == 0
         assert layer.experts is None
         assert layer.fc1_experts_weights.shape == ir.Shape([64, 64, 32])
         assert layer.fc2_experts_weights.shape == ir.Shape([64, 64, 16])
@@ -209,6 +211,8 @@ def test_int4_moe_emits_expert_major_qmoe(self):
         assert qmoe.attributes["k"].value == 6
         assert qmoe.attributes["block_size"].value == 32
         assert qmoe.attributes["swiglu_fusion"].value == 2
+        assert qmoe.attributes["quant_type"].value == "int"
+        assert qmoe.attributes["weights_prepacked"].value == 0
         assert layer.fc1_experts_weights.shape == ir.Shape([64, 64, 32])
         assert layer.fc1_scales.shape == ir.Shape([64, 64, 2])
         assert layer.fc1_experts_zero_points.shape == ir.Shape([64, 64, 1])

src/mobius/rewrite_rules/_qmoe_fusion.py

@@ -404,6 +404,8 @@ def _fuse_layer(graph: ir.Graph, layer: _DenseMoELayer, index: int) -> None:
             "expert_weight_bits": bits,
             "block_size": block_size,
             "swiglu_fusion": 2,
+            "quant_type": "int",
+            "weights_prepacked": 0,
         },
         domain=_MS_DOMAIN,
         num_outputs=1,

src/mobius/rewrite_rules/_qmoe_fusion_test.py

@@ -478,6 +478,8 @@ def test_attributes_match_qmoe_abi() -> None:
     assert attrs["swiglu_fusion"].value == 2
     assert attrs["normalize_routing_weights"].value == 1
     assert attrs["activation_type"].value == "swiglu"
+    assert attrs["quant_type"].value == "int"
+    assert attrs["weights_prepacked"].value == 0
 
 
 @pytest.mark.parametrize(