Models · Classification

Vision Transformer

Vision Transformer (Dosovitskiy et al. 2020): a transformer classifier on 16x16 image patches.

stable Fork of huggingface/pytorch-image-models@0df212b · backends: portable, kernels

Checkpoint vit_base_patch16_224.augreg2_in21k_ft_in1k

Paper 2020 Added in 0.1.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 1.66 ms 224x224, fp16
vs the faster original 3.3× original fp16: 5.48 ms
Peak memory 200 MiB original: 531 MiB
Accuracy vs original fp32 ±0.00 kernels top-1 %, ImageNet-1k val (85.88 original)

Use it

pip install fastforks

Lean install with only what this model needs: pip install "fastforks-core[timm]"

from fastforks import classification

model = classification.load("vit_base_patch16_224.augreg2_in21k_ft_in1k")  # kernels on a supported GPU, else portable PyTorch
result = model(image)
result.logits
result.backend  # what ran: mode, device, dtype, kernels
fastforks bench vit_base_patch16_224.augreg2_in21k_ft_in1k

Times the model on this machine. load(…, mode="portable") or mode="kernels" picks a backend; usage guide.

Speed and memory

Every backend on one GPU (RTX 5060 Ti 16 GB) and the same input, the authors’ code included. Kernels runs fp16. Portable runs fp16 on CUDA.

Input: synthetic, 224x224 (the task’s standard size)

BackenddtypeTime per call (ms)vs original fp16Peak memory
allocated / reserved
Load (s)First call (s)
Original fp32fp32 6.330.87×371 MiB/386 MiB0.80.15
Original fp16torch.autocast(fp16) around the forwardfp16 5.48–531 MiB/548 MiB2.50.28
Portable fp32fp32not measured ––––
Portablefp16 3.021.8×205 MiB/390 MiB0.70.00
Kernelsfp16 1.663.3×200 MiB/248 MiB2.20.07
Portable + torch.compile 2.102.6×238 MiB/266 MiB7 (build)–
TensorRT fp16fp16 1.144.8×196 MiBenginenot comparable29 (build)–

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-03 to 2026-11-15. First call, 3 warm-up, peak reset, median of 3 rounds x median of 5. Ratios divide two times of this table (same GPU, same input). Original fp16 is the authors’ own precision option, else torch.autocast. torch.compile and TensorRT: build time instead of load time. TensorRT memory (engine, not comparable) is the GPU memory the engine took, read from the driver because TensorRT allocates outside PyTorch, so it can't be compared with the PyTorch peaks of the other rows.

Other devices

DeviceBackenddtypems / callPeak memoryOutput vs referenceInput
Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-14Originalfp3217.4 1,258 MiB processmatches224x224
Portablefp1618.7 1,497 MiB processmatches224x224
Portablefp3216.9 1,314 MiB processmatches224x224
Kernelsfp16not implemented–224x224
NVIDIA L4 CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp164.92 179 MiB/206 MiB–224x224
Kernelsfp161.92 252 MiB/308 MiB–224x224
NVIDIA L40S CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp163.51 179 MiB/206 MiB–224x224
Kernelsfp161.22 252 MiB/308 MiB–224x224
NVIDIA A10 CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp162.85 179 MiB/206 MiB–224x224
Kernelsfp161.77 252 MiB/308 MiB–224x224
NVIDIA A100-SXM4-40GB CUDA, torch 2.14.1+cu130, 2026-11-08Portablefp163.52 179 MiB/206 MiB–224x224
Kernelsfp162.21 252 MiB/308 MiB–224x224
NVIDIA RTX PRO 6000 CUDA, torch 2.14.1+cu130, 2026-11-08Portablefp161.16 203 MiB/226 MiB–224x224
Kernelsfp161.02 300 MiB/368 MiB–224x224

CUDA memory is torch’s peak allocated / reserved; Apple has no such counter, so the Mac rows show the process footprint (weights, Python and torch included). “Output vs reference” compares the device’s output with the stored golden output under the card’s tolerance; – = not checked on that device. Kernels “not implemented” = no kernels for this model on that backend yet: it runs portable there.

Licence and weights

Commercial use allowed

Code
Apache-2.0
Weights
Apache-2.0 source
Training data

Our reading of the licences, not legal advice.

Weights

vit_base_patch16_224.augreg2_in21k_ft_in1k.safetensors 346 MB· hosted on Hugging Face

fastforks downloads and checks these files on first use (fastforks download vit_base_patch16_224.augreg2_in21k_ft_in1k).

Accuracy

The test set through the authors’ code and through ours: same checkpoint, same protocol. The distance between the original’s fp32 and fp16 runs is what half precision alone costs; our runs should stay within it.

ImageNet-1k val

5000 random images, top-1 / top-5 %

Rundtypetop-1 % ↑top-5 % ↑Δ top-1 % vs fp32 originalimagesms / image
Original fp32fp3285.8897.70 reference 5000 3.72
Original fp16fp1685.9097.70 +0.02 5000 1.25
Portable fp32fp3285.8697.68 −0.02 5000 3.89
Portablefp1685.8497.70 −0.04 5000 0.933
Kernelsfp1685.8897.68 ±0.00 5000 0.809

fastforks: the model's own resize / crop on the GPU; original: the library's pretrained loader and its PIL eval transform. Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-06; portable fp32: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 7c91993, 2026-10-31; portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 939adfe8, 2026-11-06.

Closeness to the original

Beyond the test-set scores: the raw output of ours against the original’s fp32 output on the same input, pixel by pixel. CPU, float32: max |diff| <= 0.002 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.02 and p99 <= 0.08 (absolute error of the output logits).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.002 not recorded
GPU default dtypes, mean |diff|≤ 0.02 0.00273
GPU default dtypes, p99 |diff|≤ 0.08 0.01018

Recorded for vit_base_patch16_224.augreg2_in21k_ft_in1k on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-04, with these kernels: cudnn_haccfvm_convfvm_extrafvm_mbfvm_r3

Details

Vision Transformer (Dosovitskiy et al. 2020): 16x16 patches, pre-norm transformer encoder, class token.

Input / output
Input: one RGB image. Output: ImageNet-1k class logits.
Input size
fixed 224x224: the input is resized (short side 248) and centre-cropped to it; tested up to 224x224 (214 MiB allocated)
Outputs
logits
Checkpoint
arch="vit_base_patch16_224.augreg2_in21k_ft_in1k"
CUDA kernels
  • cudnn_haccconvolutions with fp16 accumulation (cuDNN)
  • fvm_convfused convolutions for image classifiers
  • fvm_extrafused transformer row operations (rotary split, SwiGLU with LayerNorm)
  • fvm_mbfused MBConv blocks (MobileNet and EfficientNet families)
  • fvm_r3further fused classifier blocks
Apple (MPS)
fp32, fp16
Full reference, generated from the repository

Vision Transformer (Dosovitskiy et al. 2020): a transformer classifier on 16x16 image patches.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
vit_base_patch16_224.augreg2_in21k_ft_in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
6.33
5.48
3.02
1.66
2.10
1.14
371 / 386
531 / 548
205 / 390
200 / 248
238 / 266
394 / 420
Details
  • Original code: https://github.com/huggingface/pytorch-image-models at 0df212b369a5
  • Code licence: Apache-2.0 (permissive); files:
  • Weights licence: Apache-2.0
  • Licence evidence: https://huggingface.co/timm/vit_base_patch16_224.augreg2_in21k_ft_in1k/blob/063c6c38a5d8510b2e57df480445e94b231dad2c/README.md
  • Training data: ImageNet-21k / ImageNet-22k (full ImageNet, fall 2011 release), ImageNet-1k (ILSVRC 2012 classification)
  • Kernel libraries: cudnn_hacc, fvm_conv, fvm_extra, fvm_mb, fvm_r3
  • Variants: vit_base_patch16_224.augreg2_in21k_ft_in1k (default of vit_base_patch16_224)
  • Outputs: logits; label_set: imagenet-1k
  • Input size: fixed 224x224: the input is resized (short side 248) and centre-cropped to it
  • Tested up to (portable, fp16, RTX 5060 Ti 16 GB): vit_base_patch16_224.augreg2_in21k_ft_in1k 224x224 (0.2 GiB peak allocated, 0.3 reserved)
  • vit_base_patch16_224.augreg2_in21k_ft_in1k 224x224, first call: portable + torch.compile 7 s (cold Inductor cache), TensorRT fp16 29 s (ONNX + engine).
  • vit_base_patch16_224.augreg2_in21k_ft_in1k original 224x224: torch.autocast(fp16) around the forward
  • vit_base_patch16_224.augreg2_in21k_ft_in1k load / first call: original 2.49 s / 0.283 s, portable 0.73 s / 0.004 s, kernels 2.05 s / 0.069 s
  • Protocol, median of 5 after 3 warm-up, uint8 CUDA input (synthetic for bench_inputs sizes), card defaults; gc before each load; peak = max over the first call and the timed calls

Measured

fastforks: the model's own resize / crop on the GPU; original: the library's pretrained loader and its PIL eval transform. Original = the original code named under Details, same checkpoint and protocol.

variantdatasetoriginal fp32original fp16portableportable fp32kernelsnote
vit_base_patch16_224.augreg2_in21k_ft_in1kImageNet-1k val, 5000 random images, top-1 / top-5 %85.88 / 97.7085.90 / 97.7085.84 / 97.7085.86 / 97.6885.88 / 97.68

Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 939adfe8, 2026-11-06; portable32: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 7c91993, 2026-10-31; kernels: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 939adfe8, 2026-11-06.

Parity

CPU, float32: max |diff| <= 0.002 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.02 and p99 <= 0.08 (absolute error of the output logits).

variantkernels error vs original, mean / p99GPU, torchkernels used
vit_base_patch16_224.augreg2_in21k_ft_in1k0.00273 / 0.01018RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120

Example output

One fixed input through this checkpoint, the same for every model of the task (kernels · float16 · NVIDIA GeForce RTX 5060 Ti).

Input

Top 5

  1. Egyptian cat 63%
  2. tiger cat 9.5%
  3. tabby 7.7%
  4. lynx 0.6%
  5. lens cap 0.1%
Chelsea the cat, input
Input · 100%

Input: Chelsea the cat, Stéfan van der Walt (CC0).