| { |
| "format": "tflite", |
| "source": { |
| "repo_id": "google/vit-base-patch16-224", |
| "revision": "3f49326eb077187dfe1c2a2bb15fbd74e6ab91e3", |
| "url": "https://huggingface.co/google/vit-base-patch16-224", |
| "license": "apache-2.0" |
| }, |
| "artifacts": [ |
| { |
| "file": "model_fp32.tflite", |
| "precision": "fp32", |
| "sha256": "34678efe841c83480730c5c62b206bb6e597bc3b0d463ac3691cf698aec2295b", |
| "bytes": 346369792, |
| "inputs": [ |
| { |
| "name": "g0::pixel_values", |
| "shape": [ |
| 1, |
| 3, |
| 224, |
| 224 |
| ], |
| "dtype": "float32", |
| "layout": "NCHW" |
| } |
| ], |
| "outputs": [ |
| { |
| "name": "g0::linear_72", |
| "shape": [ |
| 1, |
| 1000 |
| ], |
| "dtype": "float32", |
| "semantics": "raw classifier logits" |
| } |
| ], |
| "quantization": null, |
| "layer_norm": { |
| "operator": "STABLEHLO_COMPOSITE", |
| "name": "odml.group_norm", |
| "count": 25, |
| "decomposition_dtype": "float32", |
| "converter_default": true, |
| "input_output_dtype": "float32" |
| }, |
| "graph": { |
| "main_operators": 346, |
| "main_conversion_boundaries": {}, |
| "dead_operators": 0, |
| "unused_constant_tensors": 0 |
| } |
| }, |
| { |
| "file": "model_a16w8.tflite", |
| "precision": "a16w8", |
| "sha256": "a0021a9b5f70422695639cc7343e8810b703fd6c788fcad4dff806ba4073e5b8", |
| "bytes": 91658499, |
| "inputs": [ |
| { |
| "name": "g0::pixel_values", |
| "shape": [ |
| 1, |
| 3, |
| 224, |
| 224 |
| ], |
| "dtype": "float32", |
| "layout": "NCHW" |
| } |
| ], |
| "outputs": [ |
| { |
| "name": "g0::static_linear_72", |
| "shape": [ |
| 1, |
| 1000 |
| ], |
| "dtype": "float32", |
| "semantics": "raw classifier logits" |
| } |
| ], |
| "quantization": { |
| "quantizer": "torchao", |
| "mode": "static", |
| "export": "native_affine_transformer_regions", |
| "scope": "all_fully_connected_layers_and_transformer_activation_regions", |
| "fully_connected_count": 73, |
| "activations": { |
| "granularity": "per_tensor", |
| "fully_connected_dtype": "int16", |
| "attention_dtype": "int16", |
| "residual_dtype": "int16", |
| "layer_norm_interface_dtype": "int16" |
| }, |
| "weights": { |
| "dtype": "int8", |
| "granularity": "per_output_channel", |
| "symmetric": true |
| }, |
| "bias_dtype": "int64", |
| "model_input_dtype": "float32", |
| "model_output_dtype": "float32", |
| "accumulator_guard": { |
| "explicit_input_bounds": false, |
| "bounded_layers": 0 |
| }, |
| "floating_regions": [ |
| "patch_embedding", |
| "public_input_output", |
| "layer_norm_fallback_decompositions" |
| ], |
| "integer_regions": [ |
| "fully_connected", |
| "attention_batch_matmul", |
| "residual_add", |
| "softmax", |
| "gelu", |
| "reshape", |
| "transpose" |
| ] |
| }, |
| "layer_norm": { |
| "operator": "STABLEHLO_COMPOSITE", |
| "name": "odml.group_norm", |
| "count": 25, |
| "decomposition_dtype": "float32", |
| "converter_default": true, |
| "input_output_dtype": "int16" |
| }, |
| "deployment_note": "Primary quantized option for the tested NPUs. No explicit accumulator input clipping; qualify accumulator behavior on other backends. CPU fallback can be slow; tested GPU IR importer lacks INT16 quantization.", |
| "graph": { |
| "main_operators": 275, |
| "main_conversion_boundaries": { |
| "float_to_integer": 1, |
| "integer_to_float": 1 |
| }, |
| "dead_operators": 0, |
| "unused_constant_tensors": 0 |
| } |
| }, |
| { |
| "file": "model_a8w8.tflite", |
| "precision": "a8w8", |
| "sha256": "58a49e8c32a90224cbe13b8ce2eaebfa38eb6ec545cee8736126ec907b4c98d4", |
| "bytes": 91335728, |
| "inputs": [ |
| { |
| "name": "g0::pixel_values", |
| "shape": [ |
| 1, |
| 3, |
| 224, |
| 224 |
| ], |
| "dtype": "float32", |
| "layout": "NCHW" |
| } |
| ], |
| "outputs": [ |
| { |
| "name": "g0::static_linear_72", |
| "shape": [ |
| 1, |
| 1000 |
| ], |
| "dtype": "float32", |
| "semantics": "raw classifier logits" |
| } |
| ], |
| "quantization": { |
| "quantizer": "torchao", |
| "mode": "static", |
| "export": "native_affine_transformer_regions", |
| "scope": "all_fully_connected_layers_and_transformer_activation_regions", |
| "fully_connected_count": 73, |
| "activations": { |
| "granularity": "per_tensor", |
| "fully_connected_dtype": "int8", |
| "attention_dtype": "int16", |
| "residual_dtype": "int16", |
| "layer_norm_interface_dtype": "int16" |
| }, |
| "weights": { |
| "dtype": "int8", |
| "granularity": "per_output_channel", |
| "symmetric": true |
| }, |
| "bias_dtype": "int32", |
| "model_input_dtype": "float32", |
| "model_output_dtype": "float32", |
| "floating_regions": [ |
| "patch_embedding", |
| "public_input_output", |
| "layer_norm_fallback_decompositions" |
| ], |
| "integer_regions": [ |
| "fully_connected", |
| "attention_batch_matmul", |
| "residual_add", |
| "softmax", |
| "gelu", |
| "reshape", |
| "transpose" |
| ], |
| "accumulator_guard": { |
| "explicit_input_bounds": false, |
| "bounded_layers": 0 |
| } |
| }, |
| "layer_norm": { |
| "operator": "STABLEHLO_COMPOSITE", |
| "name": "odml.group_norm", |
| "count": 25, |
| "decomposition_dtype": "float32", |
| "converter_default": true, |
| "input_output_dtype": "int16" |
| }, |
| "deployment_note": "Reference mixed INT8/INT16 model that may suit devices limited to 32-bit integer accumulation in fully connected kernels. Has a substantial accuracy tradeoff and retains INT16 regions; backend support must be verified.", |
| "graph": { |
| "main_operators": 372, |
| "main_conversion_boundaries": { |
| "float_to_integer": 1, |
| "integer_to_integer": 97, |
| "integer_to_float": 1 |
| }, |
| "dead_operators": 0, |
| "unused_constant_tensors": 0 |
| } |
| } |
| ], |
| "preprocessing": { |
| "config_file": "preprocessor_config.json", |
| "library": "transformers", |
| "backend": "pil" |
| }, |
| "labels_file": "labels.json", |
| "gpu_compat": false |
| } |
|
|
