Add ExecuTorch mobile/edge programs and model-card usage

XNNPACK FP32/INT8, Core ML FP32, and Vulkan FP32 .pte programs for small, small_plus, and base, with manifest.json (sizes, SHA-256, delegation, held-out mIoU).

.gitattributes

@@ -33,3 +33,15 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+mobile/base/pyrafuse_base_coreml_fp32.pte filter=lfs diff=lfs merge=lfs -text
+mobile/base/pyrafuse_base_vulkan_fp32.pte filter=lfs diff=lfs merge=lfs -text
+mobile/base/pyrafuse_base_xnnpack_fp32.pte filter=lfs diff=lfs merge=lfs -text
+mobile/base/pyrafuse_base_xnnpack_int8.pte filter=lfs diff=lfs merge=lfs -text
+mobile/small/pyrafuse_small_coreml_fp32.pte filter=lfs diff=lfs merge=lfs -text
+mobile/small/pyrafuse_small_vulkan_fp32.pte filter=lfs diff=lfs merge=lfs -text
+mobile/small/pyrafuse_small_xnnpack_fp32.pte filter=lfs diff=lfs merge=lfs -text
+mobile/small/pyrafuse_small_xnnpack_int8.pte filter=lfs diff=lfs merge=lfs -text
+mobile/small_plus/pyrafuse_small_plus_coreml_fp32.pte filter=lfs diff=lfs merge=lfs -text
+mobile/small_plus/pyrafuse_small_plus_vulkan_fp32.pte filter=lfs diff=lfs merge=lfs -text
+mobile/small_plus/pyrafuse_small_plus_xnnpack_fp32.pte filter=lfs diff=lfs merge=lfs -text
+mobile/small_plus/pyrafuse_small_plus_xnnpack_int8.pte filter=lfs diff=lfs merge=lfs -text

README.md

@@ -11,6 +11,9 @@ tags:
 - fabric-segmentation
 - dinov3
 - tensorrt
+- executorch
+- on-device
+- mobile
 ---
 
 # PyraFuse: skin, fabric, and background segmentation
@@ -46,9 +49,76 @@ model = load_pretrained("base", revision="v1.0.0", device="cuda")
 Inputs must be RGB, ImageNet-normalised, and have height and width divisible by
 16. The default checkpoint configuration uses 448 × 448 inputs.
 
+## Mobile and edge: ExecuTorch programs
+
+The `mobile/` folder contains ready-to-run [ExecuTorch](https://executorch.ai)
+programs for Android, iOS, React Native, Flutter, and embedded Linux. The same
+`.pte` file runs in every framework that embeds the ExecuTorch runtime.
+
+```text
+mobile/
+├── manifest.json                       # sizes, SHA-256, delegation, held-out mIoU
+├── small/pyrafuse_small_<target>.pte
+├── small_plus/pyrafuse_small_plus_<target>.pte
+└── base/pyrafuse_base_<target>.pte
+```
+
+| Target | Runs on | Precision |
+| --- | --- | --- |
+| `xnnpack_fp32` | CPU: Android, iOS, macOS, Linux/ARM boards | FP32 |
+| `xnnpack_int8` | same CPUs, about 4× smaller | INT8 weights, dynamic INT8 activations |
+| `coreml_fp32` | iOS 17+ / macOS 14+ GPU and CPU via Core ML | FP32 |
+| `vulkan_fp32` | Android GPU via Vulkan | FP32 |
+
+| Variant | Program | Size | mIoU | Agreement with PyTorch |
+| --- | --- | ---: | ---: | ---: |
+| `small` | PyTorch reference | | 0.907 | |
+| | `xnnpack_fp32` | 91 MB | 0.907 | 100.00% |
+| | `xnnpack_int8` | 23 MB | 0.906 | 99.78% |
+| | `coreml_fp32` | 92 MB | on device | on device |
+| | `vulkan_fp32` | 91 MB | on device | on device |
+| `small_plus` | PyTorch reference | | 0.908 | |
+| | `xnnpack_fp32` | 119 MB | 0.908 | 100.00% |
+| | `xnnpack_int8` | 31 MB | 0.907 | 99.74% |
+| | `coreml_fp32` | 121 MB | on device | on device |
+| | `vulkan_fp32` | 119 MB | on device | on device |
+| `base` | PyTorch reference | | 0.912 | |
+| | `xnnpack_fp32` | 348 MB | 0.912 | 100.00% |
+| | `xnnpack_int8` | 88 MB | 0.912 | 99.80% |
+| | `coreml_fp32` | 349 MB | on device | on device |
+| | `vulkan_fp32` | 348 MB | on device | on device |
+
+mIoU is measured on the first 200 images of the held-out test split at 448 px by executing each program with the ExecuTorch runtime; the Core ML and Vulkan programs need Apple or Android hardware and are not yet scored.
+
+**Contract.** Input `float32 [1, 3, 448, 448]`: RGB, stretch-resized, divided by
+255 (ImageNet normalisation is inside the program). Output `float32 [1, 3, 448, 448]`
+logits for `background`, `fabric`, `skin`; take `argmax` over dim 1.
+
+```python
+from executorch.runtime import Runtime
+from pyrafuse import download_mobile_model  # pip install "pyrafuse[mobile]"
+
+forward = Runtime.get().load_program(
+    download_mobile_model("small", "xnnpack_int8")).load_method("forward")
+logits = forward.execute([image])[0]  # image: float32 [1, 3, 448, 448] in [0, 1]
+```
+
+React Native (`react-native-executorch` 0.10+) loads these files directly with
+`useSemanticSegmenter` using labels `['background', 'fabric', 'skin']`,
+`resizeMode: 'stretch'`, and `normalizeOpts: { alpha: 1 / 255, beta: 0 }`.
+Copy-paste snippets for React Native, Flutter, Kotlin, Swift, and C++ are in the
+[code repository README](https://github.com/jamal-saeedi/PyraFuse#mobile-and-edge-deployment).
+
+There are no FP16 programs: DINOv3's first block produces attention logits up
+to about 2.5 × 10⁶, beyond the FP16 range, so FP16 backends (including the Core
+ML Neural Engine) output NaNs. Core ML and Vulkan programs are validated at
+export but have not yet been benchmarked on devices.
+
 ## Release policy
 
-These eager PyTorch checkpoint bundles are the portable source of truth.
+These eager PyTorch checkpoint bundles are the portable source of truth; the
+`mobile/` ExecuTorch programs are exported from them with the repository's
+`scripts/export_mobile.py`.
 TensorRT engines are deliberately not distributed here because they are tied to
 the target GPU, CUDA, TensorRT version, precision, and optimization profile.
 Build and validate TensorRT engines on the deployment host from the PyTorch

mobile/base/pyrafuse_base_coreml_fp32.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0cf0c3e2b71c476a55f035aacb74dfffe163cf60c4112601246830245baf94c6
+size 349313802

mobile/base/pyrafuse_base_vulkan_fp32.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:74c3cdea6c86b836fa7f5f3daf1f2fad86d3eaa4152ba20a043dca7149b8b49f
+size 347829004

mobile/base/pyrafuse_base_xnnpack_fp32.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb349ab8f49b46e710edfddc791d9c250ffcc29938fba831e80dcd6e012c862e
+size 347851148

mobile/base/pyrafuse_base_xnnpack_int8.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c27a182b6e66284ac7d4332e945d1b329abf07fa7aaa63b228a6b6a90868df1
+size 87957004

mobile/manifest.json

@@ -0,0 +1,403 @@
+{
+  "programs": [
+    {
+      "file": "small/pyrafuse_small_xnnpack_fp32.pte",
+      "target": "xnnpack_fp32",
+      "backend": "xnnpack",
+      "precision": "fp32",
+      "runs_on": "CPU on Android, iOS, macOS, Linux/ARM edge",
+      "size_bytes": 90837004,
+      "sha256": "b601c8d2a6dbd6d719aa22366cd1d45dbbe5ece43f422b338e981ef41ff17e55",
+      "delegate_calls": 72,
+      "non_delegated_ops": [
+        "aten._adaptive_avg_pool2d.default",
+        "aten.arange.start_step",
+        "aten.as_strided_copy.default",
+        "aten.expand_copy.default",
+        "aten.native_group_norm.default",
+        "aten.native_layer_norm.default",
+        "aten.repeat.default",
+        "aten.split_with_sizes_copy.default",
+        "aten.unsqueeze_copy.default",
+        "dim_order_ops._to_dim_order_copy.default"
+      ],
+      "build_time_s": 49.6,
+      "image_size": 448,
+      "variant": "small",
+      "eval": {
+        "miou": 0.907,
+        "iou": {
+          "background": 0.9816,
+          "fabric": 0.934,
+          "skin": 0.8053
+        },
+        "pixel_acc": 0.9802,
+        "median_ms_host_cpu": 468.0,
+        "evaluated_with": "ExecuTorch runtime, host CPU",
+        "pixel_agreement_vs_pytorch": 1.0
+      }
+    },
+    {
+      "file": "small/pyrafuse_small_xnnpack_int8.pte",
+      "target": "xnnpack_int8",
+      "backend": "xnnpack",
+      "precision": "int8",
+      "runs_on": "CPU on Android, iOS, macOS, Linux/ARM edge",
+      "size_bytes": 23352972,
+      "sha256": "aeb044d6347fbed419c417f6e6775f4315169faaf742dee4bd08055893f99c93",
+      "delegate_calls": 72,
+      "non_delegated_ops": [
+        "aten._adaptive_avg_pool2d.default",
+        "aten.arange.start_step",
+        "aten.as_strided_copy.default",
+        "aten.expand_copy.default",
+        "aten.native_group_norm.default",
+        "aten.native_layer_norm.default",
+        "aten.repeat.default",
+        "aten.split_with_sizes_copy.default",
+        "aten.unsqueeze_copy.default",
+        "dim_order_ops._to_dim_order_copy.default"
+      ],
+      "build_time_s": 70.0,
+      "image_size": 448,
+      "variant": "small",
+      "eval": {
+        "miou": 0.9064,
+        "iou": {
+          "background": 0.9815,
+          "fabric": 0.9333,
+          "skin": 0.8045
+        },
+        "pixel_acc": 0.98,
+        "median_ms_host_cpu": 461.8,
+        "evaluated_with": "ExecuTorch runtime, host CPU",
+        "pixel_agreement_vs_pytorch": 0.99775
+      }
+    },
+    {
+      "file": "small/pyrafuse_small_coreml_fp32.pte",
+      "target": "coreml_fp32",
+      "backend": "coreml",
+      "precision": "fp32",
+      "runs_on": "Apple GPU / CPU via Core ML (iOS 17+, macOS 14+)",
+      "size_bytes": 92299382,
+      "sha256": "fe2eba0f1470eb0c210f7ef1866952ce805c8a42284dd39e0af98cbf50214d7a",
+      "delegate_calls": 1,
+      "non_delegated_ops": [],
+      "build_time_s": 22.8,
+      "image_size": 448,
+      "variant": "small",
+      "eval": null
+    },
+    {
+      "file": "small/pyrafuse_small_vulkan_fp32.pte",
+      "target": "vulkan_fp32",
+      "backend": "vulkan",
+      "precision": "fp32",
+      "runs_on": "Android GPU via Vulkan",
+      "size_bytes": 90814732,
+      "sha256": "f1a98a0ac86bdb80430cf1260476ab28bfaf5dfcdb02d6dc023c248d8983e421",
+      "delegate_calls": 2,
+      "non_delegated_ops": [
+        "aten._adaptive_avg_pool2d.default",
+        "aten.as_strided_copy.default"
+      ],
+      "build_time_s": 41.1,
+      "image_size": 448,
+      "variant": "small",
+      "eval": null
+    },
+    {
+      "file": "small_plus/pyrafuse_small_plus_xnnpack_fp32.pte",
+      "target": "xnnpack_fp32",
+      "backend": "xnnpack",
+      "precision": "fp32",
+      "runs_on": "CPU on Android, iOS, macOS, Linux/ARM edge",
+      "size_bytes": 119232652,
+      "sha256": "2a451f94525d62435b6d2fc859e709d0d73b4e402f62b2f19212077587428dda",
+      "delegate_calls": 72,
+      "non_delegated_ops": [
+        "aten._adaptive_avg_pool2d.default",
+        "aten.arange.start_step",
+        "aten.as_strided_copy.default",
+        "aten.expand_copy.default",
+        "aten.native_group_norm.default",
+        "aten.native_layer_norm.default",
+        "aten.repeat.default",
+        "aten.split_with_sizes_copy.default",
+        "aten.unsqueeze_copy.default",
+        "dim_order_ops._to_dim_order_copy.default"
+      ],
+      "build_time_s": 47.1,
+      "image_size": 448,
+      "variant": "small_plus",
+      "eval": {
+        "miou": 0.9079,
+        "iou": {
+          "background": 0.9814,
+          "fabric": 0.935,
+          "skin": 0.8072
+        },
+        "pixel_acc": 0.9802,
+        "median_ms_host_cpu": 492.7,
+        "evaluated_with": "ExecuTorch runtime, host CPU",
+        "pixel_agreement_vs_pytorch": 1.0
+      }
+    },
+    {
+      "file": "small_plus/pyrafuse_small_plus_xnnpack_int8.pte",
+      "target": "xnnpack_int8",
+      "backend": "xnnpack",
+      "precision": "int8",
+      "runs_on": "CPU on Android, iOS, macOS, Linux/ARM edge",
+      "size_bytes": 30601100,
+      "sha256": "03583837dbece044eb089c07aa9264568cf67f79a6c8b335c395d5276979c6a2",
+      "delegate_calls": 84,
+      "non_delegated_ops": [
+        "aten._adaptive_avg_pool2d.default",
+        "aten.arange.start_step",
+        "aten.as_strided_copy.default",
+        "aten.expand_copy.default",
+        "aten.native_group_norm.default",
+        "aten.native_layer_norm.default",
+        "aten.repeat.default",
+        "aten.split_with_sizes_copy.default",
+        "aten.unsqueeze_copy.default",
+        "dim_order_ops._to_dim_order_copy.default"
+      ],
+      "build_time_s": 77.4,
+      "image_size": 448,
+      "variant": "small_plus",
+      "eval": {
+        "miou": 0.9069,
+        "iou": {
+          "background": 0.9811,
+          "fabric": 0.9338,
+          "skin": 0.8059
+        },
+        "pixel_acc": 0.9799,
+        "median_ms_host_cpu": 497.3,
+        "evaluated_with": "ExecuTorch runtime, host CPU",
+        "pixel_agreement_vs_pytorch": 0.99739
+      }
+    },
+    {
+      "file": "small_plus/pyrafuse_small_plus_coreml_fp32.pte",
+      "target": "coreml_fp32",
+      "backend": "coreml",
+      "precision": "fp32",
+      "runs_on": "Apple GPU / CPU via Core ML (iOS 17+, macOS 14+)",
+      "size_bytes": 120703686,
+      "sha256": "0f6ceb623cd2f4590840a452a7194c22afdb26b9afa442d836f979fad667c38f",
+      "delegate_calls": 1,
+      "non_delegated_ops": [],
+      "build_time_s": 23.6,
+      "image_size": 448,
+      "variant": "small_plus",
+      "eval": null
+    },
+    {
+      "file": "small_plus/pyrafuse_small_plus_vulkan_fp32.pte",
+      "target": "vulkan_fp32",
+      "backend": "vulkan",
+      "precision": "fp32",
+      "runs_on": "Android GPU via Vulkan",
+      "size_bytes": 119215372,
+      "sha256": "8b5c2a7715a52de7ce678ecacbd9dd778f2590283534132940dd26019a6ca3c4",
+      "delegate_calls": 2,
+      "non_delegated_ops": [
+        "aten._adaptive_avg_pool2d.default",
+        "aten.as_strided_copy.default"
+      ],
+      "build_time_s": 44.9,
+      "image_size": 448,
+      "variant": "small_plus",
+      "eval": null
+    },
+    {
+      "file": "base/pyrafuse_base_xnnpack_fp32.pte",
+      "target": "xnnpack_fp32",
+      "backend": "xnnpack",
+      "precision": "fp32",
+      "runs_on": "CPU on Android, iOS, macOS, Linux/ARM edge",
+      "size_bytes": 347851148,
+      "sha256": "eb349ab8f49b46e710edfddc791d9c250ffcc29938fba831e80dcd6e012c862e",
+      "delegate_calls": 72,
+      "non_delegated_ops": [
+        "aten._adaptive_avg_pool2d.default",
+        "aten.arange.start_step",
+        "aten.as_strided_copy.default",
+        "aten.expand_copy.default",
+        "aten.native_group_norm.default",
+        "aten.native_layer_norm.default",
+        "aten.repeat.default",
+        "aten.split_with_sizes_copy.default",
+        "aten.unsqueeze_copy.default",
+        "dim_order_ops._to_dim_order_copy.default"
+      ],
+      "build_time_s": 51.4,
+      "image_size": 448,
+      "variant": "base",
+      "eval": {
+        "miou": 0.9125,
+        "iou": {
+          "background": 0.9825,
+          "fabric": 0.9365,
+          "skin": 0.8187
+        },
+        "pixel_acc": 0.9812,
+        "median_ms_host_cpu": 767.1,
+        "evaluated_with": "ExecuTorch runtime, host CPU",
+        "pixel_agreement_vs_pytorch": 1.0
+      }
+    },
+    {
+      "file": "base/pyrafuse_base_xnnpack_int8.pte",
+      "target": "xnnpack_int8",
+      "backend": "xnnpack",
+      "precision": "int8",
+      "runs_on": "CPU on Android, iOS, macOS, Linux/ARM edge",
+      "size_bytes": 87957004,
+      "sha256": "8c27a182b6e66284ac7d4332e945d1b329abf07fa7aaa63b228a6b6a90868df1",
+      "delegate_calls": 72,
+      "non_delegated_ops": [
+        "aten._adaptive_avg_pool2d.default",
+        "aten.arange.start_step",
+        "aten.as_strided_copy.default",
+        "aten.expand_copy.default",
+        "aten.native_group_norm.default",
+        "aten.native_layer_norm.default",
+        "aten.repeat.default",
+        "aten.split_with_sizes_copy.default",
+        "aten.unsqueeze_copy.default",
+        "dim_order_ops._to_dim_order_copy.default"
+      ],
+      "build_time_s": 72.2,
+      "image_size": 448,
+      "variant": "base",
+      "eval": {
+        "miou": 0.9116,
+        "iou": {
+          "background": 0.9822,
+          "fabric": 0.9352,
+          "skin": 0.8173
+        },
+        "pixel_acc": 0.9809,
+        "median_ms_host_cpu": 713.0,
+        "evaluated_with": "ExecuTorch runtime, host CPU",
+        "pixel_agreement_vs_pytorch": 0.99802
+      }
+    },
+    {
+      "file": "base/pyrafuse_base_coreml_fp32.pte",
+      "target": "coreml_fp32",
+      "backend": "coreml",
+      "precision": "fp32",
+      "runs_on": "Apple GPU / CPU via Core ML (iOS 17+, macOS 14+)",
+      "size_bytes": 349313802,
+      "sha256": "0cf0c3e2b71c476a55f035aacb74dfffe163cf60c4112601246830245baf94c6",
+      "delegate_calls": 1,
+      "non_delegated_ops": [],
+      "build_time_s": 25.9,
+      "image_size": 448,
+      "variant": "base",
+      "eval": null
+    },
+    {
+      "file": "base/pyrafuse_base_vulkan_fp32.pte",
+      "target": "vulkan_fp32",
+      "backend": "vulkan",
+      "precision": "fp32",
+      "runs_on": "Android GPU via Vulkan",
+      "size_bytes": 347829004,
+      "sha256": "74c3cdea6c86b836fa7f5f3daf1f2fad86d3eaa4152ba20a043dca7149b8b49f",
+      "delegate_calls": 2,
+      "non_delegated_ops": [
+        "aten._adaptive_avg_pool2d.default",
+        "aten.as_strided_copy.default"
+      ],
+      "build_time_s": 45.2,
+      "image_size": 448,
+      "variant": "base",
+      "eval": null
+    }
+  ],
+  "format": "ExecuTorch .pte",
+  "input": {
+    "name": "image",
+    "dtype": "float32",
+    "shape": [
+      1,
+      3,
+      448,
+      448
+    ],
+    "layout": "NCHW RGB",
+    "range": "[0, 1] (pixel / 255)",
+    "normalization": "ImageNet mean/std applied inside the graph",
+    "mean": [
+      0.485,
+      0.456,
+      0.406
+    ],
+    "std": [
+      0.229,
+      0.224,
+      0.225
+    ]
+  },
+  "output": {
+    "name": "logits",
+    "dtype": "float32",
+    "shape": [
+      1,
+      3,
+      448,
+      448
+    ],
+    "classes": {
+      "0": "background",
+      "1": "fabric",
+      "2": "skin"
+    },
+    "postprocess": "argmax over dim 1"
+  },
+  "toolchain": {
+    "executorch": "1.5.1",
+    "torch": "2.14.0+cpu",
+    "coremltools": "9.0"
+  },
+  "eval_protocol": "first 200 images of the held-out test split (seed 42), 448 px stretch-resize; host CPU x86_64",
+  "pytorch_reference": {
+    "small": {
+      "miou": 0.907,
+      "iou": {
+        "background": 0.9816,
+        "fabric": 0.934,
+        "skin": 0.8053
+      },
+      "pixel_acc": 0.9802,
+      "median_ms_host_cpu": 382.0
+    },
+    "small_plus": {
+      "miou": 0.9079,
+      "iou": {
+        "background": 0.9814,
+        "fabric": 0.935,
+        "skin": 0.8072
+      },
+      "pixel_acc": 0.9802,
+      "median_ms_host_cpu": 423.6
+    },
+    "base": {
+      "miou": 0.9125,
+      "iou": {
+        "background": 0.9825,
+        "fabric": 0.9365,
+        "skin": 0.8187
+      },
+      "pixel_acc": 0.9812,
+      "median_ms_host_cpu": 716.0
+    }
+  }
+}

mobile/small/pyrafuse_small_coreml_fp32.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fe2eba0f1470eb0c210f7ef1866952ce805c8a42284dd39e0af98cbf50214d7a
+size 92299382

mobile/small/pyrafuse_small_vulkan_fp32.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f1a98a0ac86bdb80430cf1260476ab28bfaf5dfcdb02d6dc023c248d8983e421
+size 90814732

mobile/small/pyrafuse_small_xnnpack_fp32.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b601c8d2a6dbd6d719aa22366cd1d45dbbe5ece43f422b338e981ef41ff17e55
+size 90837004

mobile/small/pyrafuse_small_xnnpack_int8.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aeb044d6347fbed419c417f6e6775f4315169faaf742dee4bd08055893f99c93
+size 23352972

mobile/small_plus/pyrafuse_small_plus_coreml_fp32.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f6ceb623cd2f4590840a452a7194c22afdb26b9afa442d836f979fad667c38f
+size 120703686

mobile/small_plus/pyrafuse_small_plus_vulkan_fp32.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8b5c2a7715a52de7ce678ecacbd9dd778f2590283534132940dd26019a6ca3c4
+size 119215372

mobile/small_plus/pyrafuse_small_plus_xnnpack_fp32.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2a451f94525d62435b6d2fc859e709d0d73b4e402f62b2f19212077587428dda
+size 119232652

mobile/small_plus/pyrafuse_small_plus_xnnpack_int8.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:03583837dbece044eb089c07aa9264568cf67f79a6c8b335c395d5276979c6a2
+size 30601100