Support 24 more published ImageNet classifiers through YAML inheritance

.claude/skills/mblt-vision/SKILL.md

@@ -28,6 +28,14 @@ description: >-
   Use file_cfg.filename for MXQ and derive the same-stem ONNX artifact unless
   onnx_filename is required.
 - Every post_cfg declares dataset; resolve output taxonomy from the dataset/task pair.
+- Additional published EfficientViT/LeViT/ViT classifiers use `bases/imagenet_classifier.yaml`.
+  Preserve the recorded ops transforms checked by `tests/test_published_classifiers.py`,
+  including the TF-normalized `ViT_Large_Patch14_CLIP_224` exception. Their MXQ-only
+  repositories use `onnx_local_artifact_only: true`: require trusted local ONNX and
+  retain automatic MXQ loading; metadata validation skips only the marked framework.
+- New pending family siblings use `file_cfg.artifact_status: pending_hf_family_sibling`
+  and require trusted local artifacts for runtime/compile. Report them as pending without
+  network; retain byte-identical existing sibling YAMLs.
 - For Hub linkage changes, run `tests/test_hub_validation.py` as described in `tests/TEST.md`.
   The metadata checker follows runtime board paths without downloading weights; live commits
   are diagnostics, not substitutes for missing historical approval provenance.

AGENTS.md

@@ -72,6 +72,15 @@ The current ownership boundary is deliberate:
   floors, and rounding differs by a pixel at 224/0.9 and 224/0.95. Several names
   (ConvNext_Base) resolve in timm yet carry torchvision's transform here, so read the
   expected value from that model's source.yaml provenance rather than from the name.
+- Published EfficientViT, non-convolutional LeViT and additional ViT classifiers inherit
+  `bases/imagenet_classifier.yaml`. Their exact recorded preprocessing and dataset connection
+  are pinned by `tests/test_published_classifiers.py`; preserve TF normalization for
+  `ViT_Large_Patch14_CLIP_224` rather than inferring normalization from its name.
+  These 24 repositories are MXQ-only: their base uses `onnx_local_artifact_only: true`.
+  Keep automatic MXQ loading enabled, require trusted local ONNX, and test both guard paths.
+  Newly added EfficientViT B0/B1/B2 sibling configs carry
+  `artifact_status: pending_hf_family_sibling`: verify as pending without network and
+  require trusted local runtime/compile inputs. Existing sibling configs stay unchanged.
 - To share processing defaults without changing legacy models, opt new family YAMLs into
   `extends: bases/family.yaml#DEFAULT`. See the configuration-inheritance section of
   `mblt_vision/README.md` for ordering and merge rules. Keep legacy `update` shallow and

mblt_vision/README.md

@@ -100,6 +100,41 @@ This section lists the publicly pre-trained models supported by the vision frame
 
 ### Image Classification
 
+Additional published ImageNet classifiers are available through the same task
+subpackage, top-level lazy exports and `MBLT_Engine`:
+
+- EfficientViT: `EfficientViT_B3`, `EfficientViT_L1`, `EfficientViT_L2`, `EfficientViT_L3`.
+- Non-convolutional LeViT: `LeViT_128`, `LeViT_128S`, `LeViT_192`, `LeViT_256`, `LeViT_384`.
+- ViT RPN, relative-position, SRelPos, GAP and CLIP families: discover exact names with
+  `list_models("image_classification")`.
+
+These configurations extend `models/bases/imagenet_classifier.yaml` and override
+only recorded processing differences. The source is each model's
+`models/image_classification/<Model>/_source/predict_float.input.json` and matching
+`distribution.yaml` in mblt-model-ops at `6fdbf471a3bd6ec22035a42c8730fe88a68881e4`.
+Input crop/resize and normalization are model-specific: in particular,
+`ViT_Large_Patch14_CLIP_224` uses TF normalization despite its CLIP name.
+All 24 additions retain the existing `revision: main` convention because their
+historical approved artifact HF commits are not recorded; current Hub tips and
+upstream source-model commits are not substituted as approval pins. This is a
+configuration/CPU processing parity addition, not a new inference accuracy claim.
+All 24 Hub repositories expose MXQ only, with no published ONNX file. Their base
+sets `file_cfg.onnx_local_artifact_only: true`: MXQ Hub loading remains enabled,
+but ONNX requires a trusted local `model_path` or `onnx_path`. Automatic ONNX
+loading fails with a specific message, and metadata verification skips only that
+framework. Existing configurations without this marker retain their behavior.
+
+`EfficientViT_B0`, `EfficientViT_B1` and `EfficientViT_B2` are intentional family
+sibling configurations using official timm 1.0.14 `r224_in1k` defaults (224 crop,
+bicubic resize to `floor(224 / 0.95) = 235`, ImageNet normalization). They are
+**pending HF (family sibling)**, not published-model claims. Their
+`file_cfg.artifact_status: pending_hf_family_sibling` makes metadata verification
+skip both frameworks without network access and allows runtime/compile use only
+with trusted local artifacts. Existing sibling configurations are not rewritten.
+EfficientViT M0–M5 is a distinct architecture and is not added here. Alternative
+ViT CLIP/GAP resolutions require exact upstream squash-mode preprocessing support;
+this change does not approximate them with center cropping.
+
 | Model | Input Size<br>(H,W,C) | Acc<sup>Top1</sup><br>(NPU) | Acc<sup>Top1</sup><br>(GPU) | FLOPs (B) | params (M) | Source | Note |
 | --- | --- | --- | --- | --- | --- | --- | --- |
 | AlexNet | (224,224,3) | 56.084 | 56.556 | 1.43 | 61.10 | [Link](https://docs.pytorch.org/vision/stable/models/generated/torchvision.models.alexnet.html) |  |

mblt_vision/_hub_validation.py

@@ -71,7 +71,21 @@ def check_hub_artifacts(
         paths.append(
             (framework, f"{board}/{filename}" if framework == "mxq" else filename)
         )
-    if files.get("local_artifact_only"):
+
+    if files.get("artifact_status") == "pending_hf_family_sibling":
+        return tuple(
+            HubArtifactCheck(
+                repo_id, revision, None, framework, path, "pending_hf_family_sibling"
+            )
+            for framework, path in paths
+        )
+
+    def is_local_only(framework: str) -> bool:
+        return bool(files.get("local_artifact_only")) or (
+            framework == "onnx" and bool(files.get("onnx_local_artifact_only"))
+        )
+
+    if all(is_local_only(framework) for framework, _ in paths):
         return tuple(
             HubArtifactCheck(
                 repo_id, revision, None, framework, path, "local_artifact_only"
@@ -91,10 +105,14 @@ def check_hub_artifacts(
         HubArtifactCheck(
             repo_id,
             revision,
-            info.sha,
+            None if is_local_only(framework) else info.sha,
             framework,
             path,
-            "resolved" if path in available else "missing",
+            "local_artifact_only"
+            if is_local_only(framework)
+            else "resolved"
+            if path in available
+            else "missing",
         )
         for framework, path in paths
     )

mblt_vision/compile/vision.py

@@ -786,8 +786,10 @@ def resolve_quantization_values(
         _validate_ratio("topk_ratio", topk_ratio) if topk_ratio is not None else None
     )
     hosted_config: dict[str, Any] | None = None
-    if (resolved_percentile is None or resolved_topk is None) and not file_cfg.get(
-        "local_artifact_only"
+    if (
+        (resolved_percentile is None or resolved_topk is None)
+        and not file_cfg.get("local_artifact_only")
+        and file_cfg.get("artifact_status") != "pending_hf_family_sibling"
     ):
         repo_id = file_cfg.get("repo_id")
         revision = file_cfg.get("revision", "main")
@@ -884,13 +886,24 @@ def _resolve_compile_onnx_path(
     repo_id = file_cfg.get("repo_id")
     revision = file_cfg.get("revision")
     filename = _configured_onnx_filename(file_cfg)
+    if file_cfg.get("artifact_status") == "pending_hf_family_sibling":
+        raise RuntimeError(
+            "This configuration is pending HF (family sibling). Supply a trusted local "
+            "model_path or onnx_path for compilation; automatic downloads are disabled."
+        )
     if file_cfg.get("local_artifact_only"):
         repository = repo_id if isinstance(repo_id, str) and repo_id else "this model"
         raise RuntimeError(
             f"Remote artifacts for '{repository}' are disabled because this model does not "
             "have a repository-pinned immutable revision and SHA-256 digest. Supply a "
             "trusted local model_path or onnx_path instead."
         )
+    if file_cfg.get("onnx_local_artifact_only"):
+        repository = repo_id if isinstance(repo_id, str) and repo_id else "this model"
+        raise RuntimeError(
+            f"'{repository}' publishes MXQ artifacts only. ONNX compilation requires a trusted "
+            "local model_path or onnx_path; automatic ONNX downloads are disabled."
+        )
     if isinstance(repo_id, str) and isinstance(revision, str) and filename:
         downloaded_path = Path(
             hf_hub_download(

mblt_vision/image_classification/__init__.py

@@ -5,6 +5,14 @@
 from .._compat import create_model_class
 
 __all__: list[str] = [
+    "VGG11",
+    "VGG11_BN",
+    "VGG13",
+    "VGG13_BN",
+    "VGG16",
+    "VGG16_BN",
+    "VGG19",
+    "VGG19_BN",
     "AlexNet",
     "CAFormer_B36",
     "CAFormer_M36",
@@ -21,17 +29,17 @@
     "ConvNeXt_Large",
     "ConvNeXt_Small",
     "ConvNeXt_Tiny",
-    "DeiT_Base_Patch16_224",
-    "DeiT_Base_Patch16_384",
-    "DeiT_Small_Patch16_224",
-    "DeiT_Tiny_Patch16_224",
     "DeiT3_Base_Patch16_224",
     "DeiT3_Base_Patch16_384",
     "DeiT3_Large_Patch16_224",
     "DeiT3_Large_Patch16_384",
     "DeiT3_Medium_Patch16_224",
     "DeiT3_Small_Patch16_224",
     "DeiT3_Small_Patch16_384",
+    "DeiT_Base_Patch16_224",
+    "DeiT_Base_Patch16_384",
+    "DeiT_Small_Patch16_224",
+    "DeiT_Tiny_Patch16_224",
     "DenseNet121",
     "DenseNet161",
     "DenseNet169",
@@ -50,11 +58,23 @@
     "EfficientNet_V2_L",
     "EfficientNet_V2_M",
     "EfficientNet_V2_S",
+    "EfficientViT_B0",
+    "EfficientViT_B1",
+    "EfficientViT_B2",
+    "EfficientViT_B3",
+    "EfficientViT_L1",
+    "EfficientViT_L2",
+    "EfficientViT_L3",
     "FlexiViT_Base",
     "FlexiViT_Large",
     "FlexiViT_Small",
     "GoogLeNet",
     "Inception_V3",
+    "LeViT_128",
+    "LeViT_128S",
+    "LeViT_192",
+    "LeViT_256",
+    "LeViT_384",
     "LeViT_Conv_128",
     "LeViT_Conv_128S",
     "LeViT_Conv_192",
@@ -89,14 +109,14 @@
     "RepViT_M2",
     "RepViT_M2_3",
     "RepViT_M3",
+    "ResNeXt50_32x4d",
+    "ResNeXt101_32x8d",
+    "ResNeXt101_64x4d",
     "ResNet18",
     "ResNet34",
     "ResNet50",
     "ResNet101",
     "ResNet152",
-    "ResNeXt50_32x4d",
-    "ResNeXt101_32x8d",
-    "ResNeXt101_64x4d",
     "ShuffleNet_V2_X0_5",
     "ShuffleNet_V2_X1_0",
     "ShuffleNet_V2_X1_5",
@@ -106,34 +126,41 @@
     "Swin_B",
     "Swin_S",
     "Swin_T",
-    "VGG11",
-    "VGG11_BN",
-    "VGG13",
-    "VGG13_BN",
-    "VGG16",
-    "VGG16_BN",
-    "VGG19",
-    "VGG19_BN",
-    "VisFormer_Small",
-    "VisFormer_Tiny",
     "ViT_B_16",
     "ViT_B_32",
     "ViT_Base_Patch8_224",
     "ViT_Base_Patch16_224",
     "ViT_Base_Patch16_384",
+    "ViT_Base_Patch16_CLIP_224",
+    "ViT_Base_Patch16_RPN_224",
     "ViT_Base_Patch32_224",
     "ViT_Base_Patch32_384",
+    "ViT_Base_Patch32_CLIP_224",
+    "ViT_Base_Patch32_CLIP_384",
+    "ViT_Base_Patch32_CLIP_448",
+    "ViT_Huge_Patch14_CLIP_224",
     "ViT_L_16",
     "ViT_L_32",
+    "ViT_Large_Patch14_CLIP_224",
     "ViT_Large_Patch16_224",
     "ViT_Large_Patch16_384",
     "ViT_Large_Patch32_384",
+    "ViT_Medium_Patch16_GAP_256",
+    "ViT_RelPos_Base_Patch16_224",
+    "ViT_RelPos_Base_Patch32_Plus_RPN_256",
+    "ViT_RelPos_Medium_Patch16_224",
+    "ViT_RelPos_Medium_Patch16_RPN_224",
+    "ViT_RelPos_Small_Patch16_224",
+    "ViT_SRelPos_Medium_Patch16_224",
+    "ViT_SRelPos_Small_Patch16_224",
     "ViT_Small_Patch16_224",
     "ViT_Small_Patch16_384",
     "ViT_Small_Patch32_224",
     "ViT_Small_Patch32_384",
     "ViT_Tiny_Patch16_224",
     "ViT_Tiny_Patch16_384",
+    "VisFormer_Small",
+    "VisFormer_Tiny",
     "Wide_ResNet50_2",
     "Wide_ResNet101_2",
     "YOLO11lCls",
@@ -158,6 +185,33 @@
     "YOLOv8xCls",
 ]
 
+EfficientViT_B0 = create_model_class("EfficientViT_B0", __name__)
+EfficientViT_B1 = create_model_class("EfficientViT_B1", __name__)
+EfficientViT_B2 = create_model_class("EfficientViT_B2", __name__)
+EfficientViT_B3 = create_model_class("EfficientViT_B3", __name__)
+EfficientViT_L1 = create_model_class("EfficientViT_L1", __name__)
+EfficientViT_L2 = create_model_class("EfficientViT_L2", __name__)
+EfficientViT_L3 = create_model_class("EfficientViT_L3", __name__)
+LeViT_128 = create_model_class("LeViT_128", __name__)
+LeViT_128S = create_model_class("LeViT_128S", __name__)
+LeViT_192 = create_model_class("LeViT_192", __name__)
+LeViT_256 = create_model_class("LeViT_256", __name__)
+LeViT_384 = create_model_class("LeViT_384", __name__)
+ViT_RelPos_Medium_Patch16_RPN_224 = create_model_class("ViT_RelPos_Medium_Patch16_RPN_224", __name__)
+ViT_RelPos_Base_Patch32_Plus_RPN_256 = create_model_class("ViT_RelPos_Base_Patch32_Plus_RPN_256", __name__)
+ViT_Base_Patch16_RPN_224 = create_model_class("ViT_Base_Patch16_RPN_224", __name__)
+ViT_Base_Patch16_CLIP_224 = create_model_class("ViT_Base_Patch16_CLIP_224", __name__)
+ViT_Base_Patch32_CLIP_448 = create_model_class("ViT_Base_Patch32_CLIP_448", __name__)
+ViT_Medium_Patch16_GAP_256 = create_model_class("ViT_Medium_Patch16_GAP_256", __name__)
+ViT_Large_Patch14_CLIP_224 = create_model_class("ViT_Large_Patch14_CLIP_224", __name__)
+ViT_Base_Patch32_CLIP_224 = create_model_class("ViT_Base_Patch32_CLIP_224", __name__)
+ViT_RelPos_Small_Patch16_224 = create_model_class("ViT_RelPos_Small_Patch16_224", __name__)
+ViT_Base_Patch32_CLIP_384 = create_model_class("ViT_Base_Patch32_CLIP_384", __name__)
+ViT_SRelPos_Medium_Patch16_224 = create_model_class("ViT_SRelPos_Medium_Patch16_224", __name__)
+ViT_Huge_Patch14_CLIP_224 = create_model_class("ViT_Huge_Patch14_CLIP_224", __name__)
+ViT_RelPos_Base_Patch16_224 = create_model_class("ViT_RelPos_Base_Patch16_224", __name__)
+ViT_RelPos_Medium_Patch16_224 = create_model_class("ViT_RelPos_Medium_Patch16_224", __name__)
+ViT_SRelPos_Small_Patch16_224 = create_model_class("ViT_SRelPos_Small_Patch16_224", __name__)
 AlexNet = create_model_class("AlexNet", __name__)
 CAFormer_B36 = create_model_class("CAFormer_B36", __name__)
 CAFormer_M36 = create_model_class("CAFormer_M36", __name__)

mblt_vision/models/EfficientViT_B0.yaml

@@ -0,0 +1,13 @@
+# Official timm 1.0.14 efficientvit_b0.r224_in1k pretrained_cfg:
+# 224 crop, bicubic, crop_pct 0.95 -> floor(224 / 0.95) = 235, ImageNet torch stats.
+# Family sibling configuration; no published artifact or approval is claimed.
+DEFAULT:
+  extends: bases/imagenet_classifier.yaml
+  file_cfg:
+    repo_id: mobilint/EfficientViT_B0
+    filename: efficientvit_b0.mxq
+    revision: main
+    artifact_status: pending_hf_family_sibling
+  pre_cfg:
+    Resize:
+      size: 235

mblt_vision/models/EfficientViT_B1.yaml

@@ -0,0 +1,13 @@
+# Official timm 1.0.14 efficientvit_b1.r224_in1k pretrained_cfg:
+# 224 crop, bicubic, crop_pct 0.95 -> floor(224 / 0.95) = 235, ImageNet torch stats.
+# Family sibling configuration; no published artifact or approval is claimed.
+DEFAULT:
+  extends: bases/imagenet_classifier.yaml
+  file_cfg:
+    repo_id: mobilint/EfficientViT_B1
+    filename: efficientvit_b1.mxq
+    revision: main
+    artifact_status: pending_hf_family_sibling
+  pre_cfg:
+    Resize:
+      size: 235

mblt_vision/models/EfficientViT_B2.yaml

@@ -0,0 +1,13 @@
+# Official timm 1.0.14 efficientvit_b2.r224_in1k pretrained_cfg:
+# 224 crop, bicubic, crop_pct 0.95 -> floor(224 / 0.95) = 235, ImageNet torch stats.
+# Family sibling configuration; no published artifact or approval is claimed.
+DEFAULT:
+  extends: bases/imagenet_classifier.yaml
+  file_cfg:
+    repo_id: mobilint/EfficientViT_B2
+    filename: efficientvit_b2.mxq
+    revision: main
+    artifact_status: pending_hf_family_sibling
+  pre_cfg:
+    Resize:
+      size: 235

mblt_vision/models/EfficientViT_B3.yaml

@@ -0,0 +1,12 @@
+# Approval HF commit: not recorded. Do not substitute upstream source-model commits.
+# Preprocessing: mblt-model-ops models/image_classification/EfficientViT_B3/_source/predict_float.input.json
+DEFAULT:
+  extends: bases/imagenet_classifier.yaml
+  file_cfg:
+    repo_id: mobilint/EfficientViT_B3
+    filename: efficientvit_b3.mxq
+    revision: main
+  pre_cfg:
+    Resize:
+      interpolation: bicubic
+      size: 235

mblt_vision/models/EfficientViT_L1.yaml

@@ -0,0 +1,8 @@
+# Approval HF commit: not recorded. Do not substitute upstream source-model commits.
+# Preprocessing: mblt-model-ops models/image_classification/EfficientViT_L1/_source/predict_float.input.json
+DEFAULT:
+  extends: bases/imagenet_classifier.yaml
+  file_cfg:
+    repo_id: mobilint/EfficientViT_L1
+    filename: efficientvit_l1.mxq
+    revision: main

mblt_vision/models/EfficientViT_L2.yaml

@@ -0,0 +1,8 @@
+# Approval HF commit: not recorded. Do not substitute upstream source-model commits.
+# Preprocessing: mblt-model-ops models/image_classification/EfficientViT_L2/_source/predict_float.input.json
+DEFAULT:
+  extends: bases/imagenet_classifier.yaml
+  file_cfg:
+    repo_id: mobilint/EfficientViT_L2
+    filename: efficientvit_l2.mxq
+    revision: main

mblt_vision/models/EfficientViT_L3.yaml

@@ -0,0 +1,8 @@
+# Approval HF commit: not recorded. Do not substitute upstream source-model commits.
+# Preprocessing: mblt-model-ops models/image_classification/EfficientViT_L3/_source/predict_float.input.json
+DEFAULT:
+  extends: bases/imagenet_classifier.yaml
+  file_cfg:
+    repo_id: mobilint/EfficientViT_L3
+    filename: efficientvit_l3.mxq
+    revision: main