mlboydaisuke/MoGe-2-ViT-S-ExecuTorch

moge2_vits — ExecuTorch

• Source: Ruicheng/moge-2-vits-normal
• License: MIT
• Input: [[1, 3, 518, 518]] — RGB, ImageNet norm, 518x518
• Output: points [1,H,W,3] metric point map, normal [1,H,W,3], mask [1,H,W] validity, metric_scale [1]

Variants

All variants take and return fp32 tensors — swap the .pte file, keep your app code.

build	file	size (MB)	parity vs fp32 eager (worst corr)	Mac median (ms)*
fp32	moge2_vits_xnnpack_fp32.pte	140.8	0.999998	750.9
fp16	moge2_vits_xnnpack_fp16.pte	96.6	0.434851 — see below	1714.8
int8 (dynamic)	moge2_vits_xnnpack_int8.pte	76.4	0.998758	746.5
Core ML (fp16, iOS)	moge2_vits_coreml_all.pte	73.2	0.541272 — see below	87.7

The Core ML build is the same graph lowered to Apple's Neural Engine instead of XNNPACK, which is CPU-only. Measured on an iPhone 17 Pro across seven models, it runs 3.5x to 13.9x faster (median 12x) at roughly half the file size — for example Depth-Anything-V2-Small at 500.8 ms against 42.7 ms, and MODNet at 81.7 ms against 5.9 ms. It computes in fp16 and is iOS-only; the XNNPACK files stay the portable option and are what runs on Android.

*Mac arm64, single process, median of 10 — a reference point for relative cost only, not a device number (torch eager fp32 on the same machine: 345.4 ms).

Checked in the task's own units

Correlation is a first filter. These are the numbers that decide:

• fp16 — correlation reads 0.43 on this build, and that number is an artifact: one of the four outputs is a near-binary validity mask whose raw logits correlate badly while the thresholded mask is identical. Measured properly against fp32 — mask IoU 1.0000, point map cosine 1.000000, normals cosine 1.000000, metric scale within 0.6% — the geometry is unchanged.
• int8 (dynamic) — measured in the units that matter for this model — cosine similarity of the point map and normals: median 1.0000 over 10 real images, worst 1.0000.
• Core ML (fp16, iOS) — measured in the units that matter for this model — cosine similarity of the point map and normals: median 1.0000 over 10 real images, worst 1.0000.

Verification (executorch 1.4.0, torch 2.13.0)

Parity is measured against the fp32 eager model on real image input; corr is the correlation over all elements of each output tensor.

output	shape	max_abs_diff	corr
0	[1, 518, 518, 3]	1.431e-06	1.000000
1	[1, 518, 518, 3]	1.878e-06	1.000000
2	[1, 518, 518]	1.192e-07	0.999998
3	[1]	5.960e-07	nan

XNNPACK delegate coverage (fp32): 53.7% (623/1160 ops); ops left on the portable kernels: aten.arange.start_step x90, aten.clamp.default x80, aten.index.Tensor x80, aten.expand_copy.default x65, aten.squeeze_copy.dims x38, aten.native_layer_norm.default x26, aten.mul.Scalar x24, aten.logical_not.default x24, aten.where.self x22, dim_order_ops._to_dim_order_copy.default x20, aten.eq.Scalar x12, aten.full_like.default x12, aten.any.dim x12, aten.lt.Scalar x10, aten.sub.Tensor x10, aten.unsqueeze_copy.default x5, aten.sum.dim_IntList x2, aten.pow.Tensor_Scalar x2, aten._upsample_bilinear2d_aa.default x1, aten.select_copy.int x1, aten.split_with_sizes_copy.default x1

Conversion

torch.export -> to_edge_transform_and_lower(partitioner) -> .pte (conversion scripts: executorch-models)