feat(smollm3): add native SmolLM3 model family

benchmarks/performance/release.yaml

@@ -34,6 +34,10 @@ excluded_profiles:
     reason: >-
       The pinned Diffusers reference for MiniMax-H3 has not yet been integrated
       into the release performance runner.
+  - model: smollm3-3b
+    reason: >-
+      Dense SmolLM3 functional and reference-parity qualification is present, but
+      this change does not add a matching release-performance workload or receipt.
 
 entries:
   - id: albert.encode

python/tensorrt_model_connect/families/smollm3/MODEL.toml

@@ -0,0 +1,22 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+id = "smollm3"
+plugin = "smollm3"
+module = "plugin"
+capabilities = ["split_decoder_roles", "debug_layer_outputs"]
+aliases = [
+  "smollm3",
+  "SmolLM3",
+]
+prefixes = [
+  "smollm3",
+]
+architecture_patterns = [
+  "SmolLM3ForCausalLM",
+]
+debug_runner = "debug_runner.py|runner_from_bundle"
+default_build_route = "build_routing.py|prefer_native_default"
+debug_runtime_strategies = [
+  "smollm3_decoder_kv_cache",
+]

python/tensorrt_model_connect/families/smollm3/__init__.py

@@ -0,0 +1,64 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+import importlib
+import sys
+import types
+from typing import Any
+
+_plugin = None
+
+
+def _load_plugin_module():
+    global _plugin
+    if _plugin is None:
+        _plugin = importlib.import_module(f"{__name__}.plugin")
+        globals().update({
+            _name: _value
+            for _name, _value in vars(_plugin).items()
+            if not _name.startswith("__")
+        })
+    return _plugin
+
+
+def __getattr__(name: str) -> Any:
+    if name.startswith("__"):
+        raise AttributeError(name)
+    if name in {"graph_blocks", "graph_ops"}:
+        return importlib.import_module(f"{__name__}.{name}")
+    plugin_module = _load_plugin_module()
+    if name == "plugin":
+        return getattr(plugin_module, "plugin")
+    try:
+        return getattr(plugin_module, name)
+    except AttributeError:
+        raise AttributeError(name) from None
+
+
+def __dir__() -> list[str]:
+    plugin_module = _load_plugin_module()
+    return sorted(set(globals()) | {
+        _name for _name in vars(plugin_module) if not _name.startswith("__")
+    })
+
+
+class _FamilyModule(types.ModuleType):
+    def __setattr__(self, name, value):
+        # Importlib publishes a directly imported plugin submodule on its parent.
+        # Keep the public package attribute bound to the FamilyPlugin instance.
+        if name == "plugin" and isinstance(value, types.ModuleType):
+            super().__setattr__("_plugin", value)
+            super().__setattr__("plugin", value.plugin)
+            return
+        super().__setattr__(name, value)
+        if (
+            not name.startswith("__")
+            and name not in {"_plugin", "plugin"}
+            and not isinstance(value, types.ModuleType)
+        ):
+            setattr(_load_plugin_module(), name, value)
+
+
+sys.modules[__name__].__class__ = _FamilyModule