Feat/sam3 onnx hub support (issue #324)

.gitignore

@@ -203,6 +203,7 @@ $RECYCLE.BIN/
 # Temporary files
 temp/
 tmp/
+out/
 *.tmp
 *.temp
 
@@ -223,6 +224,17 @@ tmp/
 *.graphml
 *.onnxdata
 
+# ONNX Runtime external-data sidecars dropped at CWD (UUID-named *.data files)
+# from sessions that did not pin the external-data directory
+/*.data
+
+# Quark / quantizer side-effect outputs
+/quantized_info.csv
+
+# Local debug / scratch scripts (convention: leading underscore = private)
+scripts/_*.py
+scripts/_*.json
+
 # Logs
 *.log
 logs/

README.md

@@ -22,6 +22,7 @@ Purpose-built for Windows hardware diversity, the CLI handles conversion, graph
 - **All Windows ML EPs supported.** Every [supported execution provider](https://microsoft.github.io/winml-cli/latest/concepts/eps-and-devices/#eps-winml-cli-supports) is available behind the same commands.
 - **Curated model catalog.** A [verified set of models](https://microsoft.github.io/winml-cli/latest/reference/supported-models/) that run across all Windows ML EPs - a reliable starting point.
 - **Bring your own ONNX.** Not only for converting from PyTorch - bring an [existing ONNX model](https://microsoft.github.io/winml-cli/latest/tutorials/build-from-onnx/) to get operator-compatibility insights and optimize it based on the analysis.
+- **Hub-hosted ONNX models.** Reference pre-exported ONNX files on the Hugging Face Hub as `<org>/<repo>/<path/to/file>.onnx` (e.g. `onnx-community/sam3-tracker-ONNX/onnx/vision_encoder_int8.onnx`). Supported by `winml config`, `winml build`, `winml run`, `winml serve`, `winml perf`, and `winml eval`.
 
 ---
 

scripts/e2e_eval/build_registry.py

@@ -128,19 +128,35 @@ def load_optimum_types() -> set[str]:
 
 
 def load_curated_entries(curated_path: Path) -> list[dict]:
-    """Load curated entries (hf_id + task + group + priority) from source JSON."""
+    """Load curated entries (hf_id + task + group + priority) from source JSON.
+
+    Optional fields recognised by downstream consumers but not enforced here:
+
+    * ``composite_onnx`` -- ``{role: hub_onnx_ref}`` map for models that
+      ship as several role-tagged ONNX graphs (e.g. SAM-family mask
+      generators with ``image-encoder`` + ``prompt-decoder``).  Old
+      consumers ignore this field; new consumers (e.g.
+      ``mask-generation`` evaluator dispatch) read it to discover the
+      per-role files.  ``hf_id`` is still required and should point at
+      the canonical repo (typically the encoder's repo).
+    """
     with curated_path.open(encoding="utf-8") as f:
         entries = json.load(f)
-    return [
-        {
+    loaded: list[dict] = []
+    for e in entries:
+        if "hf_id" not in e:
+            continue
+        item = {
             "hf_id": e["hf_id"],
             "task": e.get("task") or "",
             "group": e.get("group", "P0"),
             "priority": e.get("priority", "P0"),
         }
-        for e in entries
-        if "hf_id" in e
-    ]
+        # Pass-through additive fields so they survive into the built registry.
+        if "composite_onnx" in e:
+            item["composite_onnx"] = e["composite_onnx"]
+        loaded.append(item)
+    return loaded
 
 
 def print_stats(registry_path: Path) -> None:
@@ -380,6 +396,10 @@ def build_registry(
                     existing["priority"] = priority
                     existing["group"] = group
                     safe_print(f"    [{priority}] {model_id} / {task} — updated (group={group})")
+                # Carry curated ``composite_onnx`` onto an existing entry so
+                # downstream consumers always see the canonical role map.
+                if "composite_onnx" in c and "composite_onnx" not in existing:
+                    existing["composite_onnx"] = c["composite_onnx"]
                 continue
 
             # New curated entry — fetch metadata if not already loaded
@@ -399,6 +419,8 @@ def build_registry(
                 "last_update_time": metadata["last_modified"],
                 "optimum_supported": is_optimum,
             }
+            if "composite_onnx" in c:
+                entry["composite_onnx"] = c["composite_onnx"]
 
             seen.add(key)
             entry_lookup[key] = entry

scripts/e2e_eval/testsets/models_with_acc.json

@@ -1927,5 +1927,47 @@
         "depth_column": "depth_map"
       }
     }
+  },
+  {
+    "hf_id": "onnx-community/sam3-tracker-ONNX",
+    "task": "mask-generation",
+    "model_type": "sam3_tracker",
+    "group": "Top200",
+    "priority": "P2",
+    "composite_onnx": {
+      "image-encoder": "onnx-community/sam3-tracker-ONNX/onnx/vision_encoder_int8.onnx",
+      "prompt-decoder": "onnx-community/sam3-tracker-ONNX/onnx/prompt_encoder_mask_decoder_int8.onnx"
+    },
+    "dataset_config": {
+      "path": "mattmdjaga/human_parsing_dataset",
+      "split": "train",
+      "metric": "mIoU",
+      "samples": 10,
+      "columns_mapping": {
+        "input_column": "image",
+        "mask_column": "mask"
+      }
+    }
+  },
+  {
+    "hf_id": "onnx-community/sam2.1-hiera-small-ONNX",
+    "task": "mask-generation",
+    "model_type": "sam2",
+    "group": "Top200",
+    "priority": "P2",
+    "composite_onnx": {
+      "image-encoder": "onnx-community/sam2.1-hiera-small-ONNX/onnx/vision_encoder_int8.onnx",
+      "prompt-decoder": "onnx-community/sam2.1-hiera-small-ONNX/onnx/prompt_encoder_mask_decoder_int8.onnx"
+    },
+    "dataset_config": {
+      "path": "mattmdjaga/human_parsing_dataset",
+      "split": "train",
+      "metric": "mIoU",
+      "samples": 10,
+      "columns_mapping": {
+        "input_column": "image",
+        "mask_column": "mask"
+      }
+    }
   }
 ]