efficientvit-sam / README.md
Flor's picture
card: update from fastforks 1cbf1671
d090e8c verified
|
Raw History Blame Contribute Delete
9.8 kB
metadata
library_name: fastforks
pipeline_tag: mask-generation
license: apache-2.0
tags:
  - promptable_segmentation
  - segment-anything
  - promptable-segmentation
  - cuda
  - pytorch
  - fastforks
  - fastforks-card
base_model:
  - mit-han-lab/efficientvit-sam

EfficientViT-SAM (fastforks)

EfficientViT-SAM (Zhang, Cai et al. 2024, MIT Han lab): Segment Anything with the ViT-H image encoder replaced by an EfficientViT (conv stages + ReLU linear attention) distilled from it; SAM's prompt encoder and mask decoder, so points, boxes and mask logits work the same. The portable module() is the image encoder only.

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/mit-han-lab/efficientvit at de7d7733cc03.

Author weight repos: https://huggingface.co/mit-han-lab/efficientvit-sam.

Task: promptable_segmentation. Tier: stable. Outputs: pred_masks, iou_scores. Backends: portable, kernels.

Input: one image plus a point, box or text prompt. Output: masks with quality scores.

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
efficientvit-sam-l0@1 efficientvit_sam_l0.pt c4f994b01a16 Apache-2.0 yes
efficientvit-sam-l1@1 efficientvit_sam_l1.pt fa151df2b9b9 Apache-2.0 yes
efficientvit-sam-l2@1 efficientvit_sam_l2.pt d4bfd842224c Apache-2.0 yes
efficientvit-sam-xl0@1 efficientvit_sam_xl0.pt 9bc2f87bd23d Apache-2.0 yes
efficientvit-sam-xl1@1 efficientvit_sam_xl1.pt 10f422776794 Apache-2.0 yes

Licence

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

Training data: distilled from SAM ViT-H, then trained on SA-1B (11M licensed images, 1.1B automatically generated masks) per the paper; the dataset itself is under the SA-1B research licence.

Licence evidence: https://huggingface.co/mit-han-lab/efficientvit-sam/blob/a2f0c5929196a13ef1ffc9338d3f9e482e1e0e68/README.md (license: apache-2.0, with LICENSE.txt); code: https://github.com/mit-han-lab/efficientvit/blob/de7d7733cc0329f391b33f1f459271562ec27bd5/LICENSE

Input size

any H x W >= 1, resized inside to the fixed encoder input 1024x1024 (the default export size) Tested up to (portable, fp16, RTX 5060 Ti 16 GB): efficientvit-sam-l0 1080x1920 (0.2 GiB peak allocated, 0.2 reserved); efficientvit-sam-l1 1080x1920 (0.2 GiB peak allocated, 0.2 reserved); efficientvit-sam-l2 1080x1920 (0.2 GiB peak allocated, 0.3 reserved); efficientvit-sam-xl0 1080x1920 (0.4 GiB peak allocated, 0.5 reserved); efficientvit-sam-xl1 1080x1920 (0.6 GiB peak allocated, 0.7 reserved)

Speed and memory

Promptable segmentation, synthetic, 1024x1024

model time (ms): original fp32 original portable kernels kernels speed-up: vs original vs portable peak memory (MiB, allocated / reserved): original portable kernels
efficientvit-sam-l0 26.9 26.3 12.7 4.45 5.9x 2.9x 211 / 232 171 / 210 176 / 256
efficientvit-sam-l1 30.1 29.9 14.4 5.79 5.2x 2.5x 260 / 278 195 / 234 203 / 284
efficientvit-sam-l2 34.4 34.4 17.8 6.83 5.0x 2.6x 313 / 344 221 / 260 232 / 312
efficientvit-sam-xl0 46.9 28.2 19.0 9.75 2.9x 1.9x 654 / 670 433 / 506 401 / 492
efficientvit-sam-xl1 72.6 40.9 28.9 15.3 2.7x 1.9x 984 / 1022 598 / 676 582 / 728

Startup and first call

model load (s): original portable kernels first call (s): original portable kernels
efficientvit-sam-l0 0.35 0.83 1.63 0.82 0.72 0.25
efficientvit-sam-l1 0.31 0.76 1.64 0.03 0.69 0.25
efficientvit-sam-l2 0.39 0.93 1.78 0.04 0.66 0.25
efficientvit-sam-xl0 0.62 2.67 2.14 0.59 0.52 0.26
efficientvit-sam-xl1 1.06 2.32 2.51 0.04 0.53 0.24

torch.compile and TensorRT

model input time: portable compile TensorRT fp16 kernels preparation: compile (cold first call) ONNX + engine build peak MiB (alloc / reserved): portable compile kernels TensorRT (engine + torch)
efficientvit-sam-l0 1024x1024 12.7 ms 6.44 ms 5.56 ms, fp32 engine (image encoder only) 4.45 ms 25.2 s 19.9 s 171 / 210 188 / 228 176 / 256 404 / 428
efficientvit-sam-l1 1024x1024 14.4 ms 11.3 ms 7.30 ms, fp32 engine (image encoder only) 5.79 ms 24.8 s 22.0 s 195 / 234 212 / 252 203 / 284 517 / 578
efficientvit-sam-l2 1024x1024 17.8 ms 10.8 ms 9.65 ms, fp32 engine (image encoder only) 6.83 ms 26.1 s 24.1 s 221 / 260 238 / 278 232 / 312 633 / 690
efficientvit-sam-xl0 1024x1024 19.0 ms 14.2 ms 22.7 ms, fp32 engine (image encoder only) 9.75 ms 24.0 s 29.8 s 433 / 506 397 / 424 401 / 492 1288 / 1328
efficientvit-sam-xl1 1024x1024 28.9 ms 22.9 ms 38.4 ms, fp32 engine (image encoder only) 15.3 ms 26.7 s 36.6 s 598 / 676 562 / 616 582 / 728 1894 / 1960

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 original code named above 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

On public datasets

One point per ground-truth instance (the interior point farthest from the border); IoU of the best-scored of 3 masks / of the single mask. Original = the original code named under Details, same checkpoint and protocol.

variant dataset original fp32 original fp16 portable kernels note
efficientvit-sam-l0 COCO val2017, every 10th image, point prompts, mIoU best of 3 / single mask % 54.45 / 50.57 54.60 / 50.50 54.58 / 50.54 54.59 / 50.52
efficientvit-sam-l1 COCO val2017, every 10th image, point prompts, mIoU best of 3 / single mask % 55.48 / 51.26 55.38 / 51.06 55.37 / 51.01 55.42 / 51.08
efficientvit-sam-l2 COCO val2017, every 10th image, point prompts, mIoU best of 3 / single mask % 56.50 / 52.23 56.58 / 52.29 56.44 / 52.25 56.34 / 52.22
efficientvit-sam-xl0 COCO val2017, every 10th image, point prompts, mIoU best of 3 / single mask % 59.00 / 54.47 58.90 / 54.48 58.99 / 54.51 58.90 / 54.51
efficientvit-sam-xl1 COCO val2017, every 10th image, point prompts, mIoU best of 3 / single mask % 59.38 / 54.84 59.33 / 54.84 59.40 / 54.85 59.43 / 54.85

Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-10.

Parity with the original code

Every variant is checked against reference 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| <= 2.5 % and p99 <= 7.3 % of the reference's median |value| (absolute error of the output pred_masks).

Usage

from fastforks import promptable_segmentation

model = promptable_segmentation.load("efficientvit-sam-l0")  # mode="auto": the preferred backend from recorded measurements
result = model("photo.jpg", input_points=[[x, y]], input_labels=[1])
result.masks  # [M, H, W] bool
# SAM 3 (text prompts): model("photo.jpg", text="coin") -> boxes, scores, masks
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: efficientvit-sam-l0, efficientvit-sam-l1, efficientvit-sam-l2, efficientvit-sam-xl0, efficientvit-sam-xl1.

Kernels backend

promptable_segmentation.load("efficientvit-sam-l0", mode="kernels") runs our CUDA kernels (fvm_conv, fvm_extra, fvm_mb, fvm_r3). 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.