WP1: native VLM package emission (typed metadata) (#420)

Implements native VLM package emission aligned with the typed onnx-genai
metadata schema.

Independent review (Resch):
- Model-name gate: no VLM production logic branches on model identity;
the sole grep hit is an unrelated pre-existing Qwen3-TTS docstring.
- `python3 -m pytest
src/mobius/integrations/onnx_genai/inference_metadata_test.py -q`: 40
passed. (`python` is unavailable on this host, so the equivalent
`python3` interpreter was used.)
- `ruff check src/mobius/integrations/onnx_genai/inference_metadata.py`:
passed.
- Spot-checked schema fields/enums/capabilities against onnx-genai Rust
and JSON schemas.

Authored by Dave; independently reviewed by Resch.

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

Author: justinchuby

src/mobius/integrations/onnx_genai/inference_metadata.py

@@ -686,6 +686,7 @@ def _state_and_kv_pairs(
 class _PositionProgram:
     rank: int
     axes: tuple[str, ...]
+    generation: str
     continuation: str
     matches: Callable[[Any], bool]
     sections_attribute: str | None = None
@@ -695,13 +696,15 @@ class _PositionProgram:
     _PositionProgram(
         rank=1,
         axes=("sequence",),
+        generation="linear",
         continuation="linear_increment",
         matches=lambda config: True,
     ),
     _PositionProgram(
         rank=3,
         axes=("temporal", "height", "width"),
-        continuation="from_grid",
+        generation="processor_coordinates",
+        continuation="carry_max",
         matches=lambda config: (
             bool(getattr(config, "mrope_interleaved", False))
             and bool(getattr(config, "mrope_section", None))
@@ -729,7 +732,9 @@ def _positions_from_registry(position: _Port, config: Any) -> dict[str, Any]:
         positions: dict[str, Any] = {
             "input": position.name,
             "rank": program.rank,
+            "tensor_rank": position.rank,
             "dtype": position.dtype,
+            "generation": program.generation,
             "continuation": program.continuation,
             "axes": list(program.axes),
         }
@@ -975,12 +980,12 @@ def validate_executable_closure(pkg: Any, metadata: dict[str, Any]) -> None:
                 f"but the consumer requires {target.dtype}/{target.rank}.",
                 "insert an explicit typed transform or remove the incompatible edge.",
             )
-        if edge.get("dtype") != source.dtype or edge.get("rank") != source.rank:
+        if edge.get("dtype") != source.dtype:
             raise _closure_error(
                 target_endpoint,
-                f"the edge declares dtype/rank {edge.get('dtype')}/{edge.get('rank')}, "
-                f"but the ONNX ports are {source.dtype}/{source.rank}.",
-                "derive edge dtype and rank directly from the matched graph ports.",
+                f"the edge declares dtype {edge.get('dtype')}, "
+                f"but the ONNX ports use {source.dtype}.",
+                "derive the edge dtype directly from the matched graph ports.",
             )
         incoming.setdefault(target_endpoint, []).append(edge)
 
@@ -1182,7 +1187,6 @@ def build_native_vlm_package_metadata(
                         "from": f"{source_name}.{source_port.name}",
                         "to": f"{target_name}.{target_port.name}",
                         "dtype": source_port.dtype,
-                        "rank": source_port.rank,
                         "device_transfer": False,
                     }
                 )
@@ -1314,17 +1318,24 @@ def build_native_vlm_package_metadata(
     vision_config = {key: value for key, value in vision_config.items() if value is not None}
 
     metadata = dict(decoder_metadata or {})
+    metadata.setdefault("schema_version", "v1")
     capabilities = list(metadata.get("required_capabilities", []))
     for capability in (
-        "multimodal_image_preprocessing",
+        "image_preprocessing_program",
         "autoregressive_every_step_components",
     ):
         if capability not in capabilities:
             capabilities.append(capability)
+    if len(preprocessing_outputs) > 1 and "packed_image_outputs" not in capabilities:
+        capabilities.append("packed_image_outputs")
+    if positions is not None and "position_program" not in capabilities:
+        capabilities.append("position_program")
     if positions is not None and positions["rank"] > 1:
-        capabilities.append("multiaxis_positions")
+        capabilities.append("multi_axis_positions")
     if decoder_io.get("state_pairs"):
-        capabilities.append("loop_state")
+        capabilities.append("loop_carried_state")
+    if decoder_io.get("token_input") and decoder_io.get("inputs_embeds_input"):
+        capabilities.append("dual_sequence_inputs")
     metadata["required_capabilities"] = capabilities
     metadata.setdefault("model", {})["io"] = decoder_io
     metadata["preprocessing"] = {

src/mobius/integrations/onnx_genai/inference_metadata_test.py

@@ -420,26 +420,30 @@ def test_gemma4_routes_all_embedding_outputs(self, tmp_path):
         validate_executable_closure(package, metadata)
         self._validate(metadata)
         _assert_all_graph_ports_declared(package, metadata)
+        assert metadata["schema_version"] == "v1"
+        assert {
+            "image_preprocessing_program",
+            "packed_image_outputs",
+            "position_program",
+            "dual_sequence_inputs",
+        } <= set(metadata["required_capabilities"])
         flow = metadata["pipeline"]["dataflow"]
         assert {
             "from": "embedding.inputs_embeds",
             "to": "decoder.inputs_embeds",
             "dtype": "fp32",
-            "rank": 3,
             "device_transfer": False,
         } in flow
         assert {
             "from": "embedding.per_layer_inputs",
             "to": "decoder.per_layer_inputs",
             "dtype": "fp32",
-            "rank": 3,
             "device_transfer": False,
         } in flow
         assert {
             "from": "audio_encoder.audio_features",
             "to": "embedding.audio_features",
             "dtype": "fp32",
-            "rank": 2,
             "device_transfer": False,
         } in flow
         embedding_audio = next(
@@ -621,12 +625,19 @@ def test_qwen_packed_grid_rank3_positions_sparse_and_fixed_state(self, tmp_path)
         )
         self._validate(metadata)
         _assert_all_graph_ports_declared(package, metadata)
+        assert {
+            "position_program",
+            "multi_axis_positions",
+            "loop_carried_state",
+        } <= set(metadata["required_capabilities"])
         positions = metadata["pipeline"]["positions"]
         assert positions == {
             "input": "position_ids",
             "rank": 3,
+            "tensor_rank": 3,
             "dtype": "int64",
-            "continuation": "from_grid",
+            "generation": "processor_coordinates",
+            "continuation": "carry_max",
             "axes": ["temporal", "height", "width"],
             "sections": [16, 24, 24],
             "processor_summaries": ["vision_encoder.image_grid_thw"],
@@ -764,7 +775,6 @@ def test_phi_routes_both_modality_gates_and_mask_processor(self, tmp_path):
                 "from": f"embedding.{gate}",
                 "to": f"decoder.{gate}",
                 "dtype": "fp32",
-                "rank": 0,
                 "device_transfer": False,
             } in flow
         outputs = metadata["preprocessing"]["image"]["outputs"]