DINOv3 ViT · fastforks

Meta's DINOv3 vision transformers: self-supervised image features with register tokens and RoPE. Needs the gated DINOv3 licence.

Commercial use: yes.

variant	input	backend	time (ms)	peak memory (MiB, allocated / reserved)
`dinov3-vits16`	synthetic (224x224)	original (fp32) original portable kernels portable + torch.compile TensorRT fp16	6.47 9.16 4.67 0.903 2.18 0.564	119 / 132 158 / 166 77 / 80 77 / 106 109 / 138 64 (engine)
`dinov3-vitb16`	synthetic (224x224)	original (fp32) original portable kernels	6.95 9.33 4.86 1.40	367 / 390 527 / 550 201 / 206 201 / 252
`dinov3-vitl16`	synthetic (224x224)	original (fp32) original portable kernels	19.2 17.7 9.13 3.28	1199 / 1212 1772 / 1812 617 / 646 617 / 674
`dinov3-vith16plus`	synthetic (224x224)	original (fp32) original portable kernels	48.0 25.9 13.3 9.30	3255 / 3276 4849 / 4876 1645 / 1668 1644 / 1694

Details

Original code: https://github.com/huggingface/transformers at v5.19.0
Code licence: LicenseRef-DINOv3-License (permissive); files: LICENSE.md
Weights licence: LicenseRef-DINOv3-License
Licence evidence: https://huggingface.co/facebook/dinov3-vits16-pretrain-lvd1689m/blob/114c1379950215c8b35dfcd4e90a5c251dde0d32/LICENSE.md (DINOv3 License, the same file in every model repo; gated: accept on Hugging Face to download)
Training data: LVD-1689M: 1.689B images curated from 17B public Instagram images by retrieval / clustering, self-supervised (no labels), per the paper; not released
Licence note: Code and weights use the DINOv3 License (Meta: attribution and conditions apply); read it before commercial use.
Options: fp16_accumulate = False (precision only)
Kernel libraries: fastbb
Variants: dinov3-vits16, dinov3-vitb16, dinov3-vitl16, dinov3-vith16plus
Outputs: last_hidden_state, pooler_output, hidden_states
Input size: any H x W >= 1, resized or padded inside the model
Tested up to (portable, fp16, RTX 5060 Ti 16 GB): dinov3-vits16 3000x4000 (1.0 GiB peak allocated, 1.2 reserved); dinov3-vitb16 3000x4000 (1.6 GiB peak allocated, 2.0 reserved); dinov3-vitl16 3000x4000 (2.4 GiB peak allocated, 2.9 reserved); dinov3-vith16plus 3000x4000 (4.2 GiB peak allocated, 4.9 reserved)
dinov3-vits16 224x224, first call: portable + torch.compile 9 s (cold Inductor cache), TensorRT fp16 25 s (ONNX + engine).
dinov3-vits16 original 224x224: torch.autocast(fp16) around the forward
dinov3-vitb16 original 224x224: torch.autocast(fp16) around the forward
dinov3-vitl16 original 224x224: torch.autocast(fp16) around the forward
dinov3-vith16plus original 224x224: torch.autocast(fp16) around the forward
dinov3-vits16 load / first call: original 0.48 s / 0.947 s, portable 0.07 s / 0.271 s, kernels 0.06 s / 0.259 s
dinov3-vitb16 load / first call: original 0.68 s / 0.012 s, portable 0.38 s / 0.198 s, kernels 0.16 s / 0.291 s
dinov3-vitl16 load / first call: original 2.37 s / 0.023 s, portable 0.54 s / 0.18 s, kernels 0.52 s / 0.246 s
dinov3-vith16plus load / first call: original 5.95 s / 0.046 s, portable 2.02 s / 0.099 s, kernels 1.4 s / 0.202 s
Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5

Measured

Two disjoint seeded subsets of the val folders; each image shorter side 256 (bicubic), centre crop 224, the CLS / pooled feature; cosine kNN, vote weighted by similarity. Original = the original code named under Details, same checkpoint and protocol.

variant	dataset	original fp32	original fp16	portable	kernels
`dinov3-vits16`	ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images)	55.38	55.42	55.42	55.36
`dinov3-vitb16`	ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images)	66.54	66.54	66.52	66.54
`dinov3-vitl16`	ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images)	75.94	75.92	75.80	75.88
`dinov3-vith16plus`	ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images)	79.50	79.50	79.52	79.52

Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 939adfe8, 2026-11-06.

Parity

variant	kernels error vs original, mean / p99	GPU, torch	kernels used
`dinov3-vits16`	0.00418 / 0.01399	RTX 5060 Ti, torch 2.14.1+cu130	`fastbb@7d058322634d:sm_120`
`dinov3-vitb16`	0.00562 / 0.01956	RTX 5060 Ti, torch 2.14.1+cu130	`fastbb@7d058322634d:sm_120`
`dinov3-vitl16`	0.00773 / 0.02994	RTX 5060 Ti, torch 2.14.1+cu130	`fastbb@7d058322634d:sm_120`
`dinov3-vith16plus`	0.00567 / 0.0234	RTX 5060 Ti, torch 2.14.1+cu130	`fastbb@7d058322634d:sm_120`

Limitations

kernels replays a CUDA graph per small input shape and layer set (up to 512 px, ViT-S 1024 px), captured on the shape's second call; at most 8 per model holding at most 256 MB of outputs (a full cache replaces only graphs unused for 64 calls, other new shapes run eagerly).
The graphs' shared memory pool (intermediates) comes on top of that limit and stays reserved: up to about 470 MB more PyTorch cache than eager for ViT-H+ at 512 px with all layers, usually under 75 MB; model.release_graphs() frees it on the device.