README.md

metadata

library_name: fastforks
pipeline_tag: image-segmentation
license: other
license_name: mixed-see-variant-table
license_link: >-
  https://huggingface.co/tue-mps/ade20k_semantic_eomt_large_512/blob/aeb262f5a0bc609236d27f1277c3173ce57aeb27/README.md
tags:
  - segmentation
  - semantic-segmentation
  - cuda
  - pytorch
  - fastforks

EoMT (fastforks)

EoMT (Kerssies et al. 2025): encoder-only mask transformer, a plain ViT-L/16 whose last blocks also carry the mask queries; DINOv2 or DINOv3 pre-training, semantic ADE20K checkpoints.

This repo is the model card only: it holds no weights. fastforks downloads the authors' weights from the links below and checks them against the SHA-256 in the model card. Code: https://github.com/kleinicke/fastforks. Kernels for the kernels backend: https://huggingface.co/fastforks/kernels. Original code: https://github.com/huggingface/transformers at c4c4605f9618.

Author weight repos: https://huggingface.co/tue-mps/ade20k_semantic_eomt_large_512, https://huggingface.co/tue-mps/eomt-dinov3-ade-semantic-large-512.

Task: segmentation. Tier: stable. Outputs: logits. Backends: portable, kernels.

Input: one RGB image. Output: per-pixel class logits.

Variants

Weights are the authors' files (not hosted here). SHA-256 shortened to 12 hex digits; the full hashes are in card.toml.

variant	weights (author link)	SHA-256	weights licence	commercial use
eomt-l-dinov2-ade20k@1	eomt-l-dinov2-ade20k.config.json eomt-l-dinov2-ade20k.safetensors	5b11b6c2cc74 715f374b2bc9	MIT	yes
eomt-l-dinov3-ade20k@1	eomt-l-dinov3-ade20k.config.json eomt-l-dinov3-ade20k.safetensors	a06117034714 319c72ee001f	MIT AND LicenseRef-DINOv3-License	yes

Licence

Original-code strings, recorded as they are (research-use package; check them before any other use). Code: Apache-2.0. Files: .

Training data: LVD-142M (DINOv2's curated web-image set, not released), ADE20K (SceneParsing, 150 classes).

• Licence note: The eomt-l-dinov3-ade20k checkpoint is fine-tuned from DINOv3, so it also carries the DINOv3 License (redistribution must include its text); that licence does not exclude commercial use.

Licence evidence: https://huggingface.co/tue-mps/ade20k_semantic_eomt_large_512/blob/aeb262f5a0bc609236d27f1277c3173ce57aeb27/README.md

Input size

any H x W >= 1, resized or padded inside the model Tested up to (portable, fp16, RTX 5060 Ti 16 GB): eomt-l-dinov2-ade20k 3000x4000 (14.5 GiB peak allocated, 14.7 reserved); eomt-l-dinov3-ade20k 3000x4000 (14.5 GiB peak allocated, 14.7 reserved)

Speed and memory

Segmentation, synthetic, 1024x2048

model	time (ms): original fp32	original	portable	kernels	kernels speed-up: vs original	vs portable	peak memory (MiB, allocated / reserved): original	portable	kernels
eomt-l-dinov2-ade20k	177.6	63.73	50.76	34.99	1.8x	1.5x	3827 / 4086	1515 / 1584	1515 / 1584
eomt-l-dinov3-ade20k	185.68	72.08	54.585	38.8	1.9x	1.4x	3822 / 4086	1514 / 1584	1514 / 1584

Startup and first call

model	load (s): original	portable	kernels	first call (s): original	portable	kernels
eomt-l-dinov2-ade20k	4.80 (512x512)	3.13	3.27†	0.54 (512x512)	0.05	0.04†
eomt-l-dinov3-ade20k	1.83 (512x512)	3.20	3.24†	0.07 (512x512)	0.08	0.04†

torch.compile and TensorRT

model	input	portable ms	compile ms	compile s (cold)	TensorRT fp16 ms	ONNX + build s	kernels ms	peak MiB (alloc / reserved): portable	compile	TensorRT (device)
eomt-l-dinov2-ade20k	1024x2048	50.76	50.8	0.3	fails: TensorRT: antialiased resize not supported	–	34.99	1515 / 1584	1512 / 1604	–

Fixed-size speed and peak memory, all models

model	mode	640x480	1280x720	1920x1080	12 MP 4000x3000
eomt-l-dinov2-ade20k	portable	50.3 ms / 1917 / 2484	55.4 ms / 2412 / 3506	63.2 ms / 3417 / 4834	129.2 ms / 14890 / 15016
eomt-l-dinov2-ade20k	kernels	38.5 ms / 1917 / 2484	43.8 ms / 2412 / 3506	51.5 ms / 3417 / 4834	117.4 ms / 14890 / 15016
eomt-l-dinov3-ade20k	portable	54.2 ms / 1916 / 2486	59.2 ms / 2411 / 3502	67.0 ms / 3415 / 4836	133.1 ms / 14889 / 15016
eomt-l-dinov3-ade20k	kernels	42.5 ms / 1916 / 2486	47.7 ms / 2411 / 3502	55.3 ms / 3415 / 4836	121.2 ms / 14889 / 15016

Fixed-size cells: time (ms) / peak allocated / peak reserved memory (MiB).

Hardware, protocol and caveats: benchmarks (RTX 5060 Ti 16 GB, sm_120). Original = the authors' code at the precision of portable (fp16, or (fp32) where fp16 is unsupported or not finite); original fp32 = its default fp32 run. Peak memory is the torch allocator's.

Accuracy

Parity with the original code

Every variant is checked against golden outputs of the original code at its pinned commit. CPU, float32: max |diff| <= 0.002 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.0002 and p99 <= 0.003 (absolute error of the output logits).

Usage

from fastforks import segmentation

model = segmentation.load("eomt-l-dinov2-ade20k")  # mode="auto": the preferred backend from recorded measurements
result = model("photo.jpg")
result.labels  # [H, W] per-pixel class index
result.backend  # what ran, e.g. Backend(mode='portable', device='cuda:0', dtype='float16')

Install with pip install fastforks (see https://github.com/kleinicke/fastforks). Other variants: eomt-l-dinov2-ade20k, eomt-l-dinov3-ade20k.

Kernels backend

segmentation.load("eomt-l-dinov2-ade20k", mode="kernels") runs our CUDA kernels (fastbb, seg_up). It needs a prebuilt kernel for your GPU architecture (downloaded from https://huggingface.co/fastforks/kernels on first use) and a recent NVIDIA driver; if there is none it raises BackendUnavailableError with the reason and never falls back silently. mode="auto" picks kernels when it can run and logs the choice.