Models · Classification

EVA-02

EVA-02 (Fang et al. 2023): a vision transformer pre-trained with masked image modelling.

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

Checkpoint eva02_large_patch14_448.mim_m38m_ft_in22k_in1k

Paper 2023 Added in 0.1.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 14.1 ms 224x224, fp16
vs the faster original 2.3× original fp16: 32.1 ms
Peak memory 652 MiB original: 1,819 MiB
Accuracy vs original fp32 −0.12 kernels top-1 %, ImageNet-1k val (90.62 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("eva02_large_patch14_448.mim_m38m_ft_in22k_in1k")  # kernels on a supported GPU, else portable PyTorch
result = model(image)
result.logits
result.backend  # what ran: mode, device, dtype, kernels
fastforks bench eva02_large_patch14_448.mim_m38m_ft_in22k_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 83.70.38×1,260 MiB/1,284 MiB2.90.14
Original fp16torch.autocast(fp16) around the forwardfp16 32.1–1,819 MiB/1,864 MiB7.00.39
Portable fp32fp32not measured ––––
Portablefp16 18.21.8×648 MiB/664 MiB4.50.17
Kernelsfp16 14.12.3×652 MiB/786 MiB5.90.17
Portable + torch.compile 14.92.1×674 MiB/684 MiB21 (build)–
TensorRT fp16fp16 12.42.6×642 MiBenginenot comparable44 (build)–

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-03 to 2026-11-16. 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, 2026-11-12Originalfp32267 3,515 MiB processmatches224x224
Portablefp16235 4,094 MiB processmatches224x224
Portablefp32267 4,096 MiB processmatches224x224
Kernelsfp16not implemented–224x224
NVIDIA L4 CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp1625.3 624 MiB/644 MiB–224x224
Kernelsfp1619.6 1,204 MiB/1,346 MiB–224x224
NVIDIA L40S CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp1616.7 624 MiB/644 MiB–224x224
Kernelsfp166.82 1,204 MiB/1,346 MiB–224x224
NVIDIA A10 CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp1623.1 624 MiB/644 MiB–224x224
Kernelsfp1616.9 1,204 MiB/1,346 MiB–224x224
NVIDIA A100-SXM4-40GB CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp1616.3 624 MiB/644 MiB–224x224
Kernelsfp1610.3 1,204 MiB/1,346 MiB–224x224
NVIDIA RTX PRO 6000 CUDA, torch 2.14.1+cu130, 2026-11-08Portablefp166.80 651 MiB/664 MiB–224x224
Kernelsfp165.46 1,233 MiB/1,406 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
MIT source
Training data

Our reading of the licences, not legal advice.

Weights

eva02_large_patch14_448.mim_m38m_ft_in22k_in1k.safetensors 1.2 GB· hosted on Hugging Face

fastforks downloads and checks these files on first use (fastforks download eva02_large_patch14_448.mim_m38m_ft_in22k_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 fp32fp3290.6299.10 reference 5000 86.0
Original fp16fp1690.6299.10 ±0.00 5000 34.5
Portable fp32fp3290.4699.08 −0.16 5000 86.6
Portablefp1690.4699.08 −0.16 5000 21.4
Kernelsfp1690.5099.08 −0.12 5000 13.5

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 0.00154
GPU default dtypes, p99 |diff|≤ 0.08 0.00587

Recorded for eva02_large_patch14_448.mim_m38m_ft_in22k_in1k on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-04, with these kernels: cudnn_haccfvm_convfvm_extrafvm_mbfvm_r3

Details

EVA-02 (Fang et al. 2023): ViT with SwiGLU and rotary position embeddings, MIM pre-trained.

Input / output
Input: one RGB image. Output: ImageNet-1k class logits.
Input size
fixed 448x448: the input is resized (short side 448) and centre-cropped to it; tested up to 448x448 (666 MiB allocated)
Outputs
logits
Checkpoint
arch="eva02_large_patch14_448.mim_m38m_ft_in22k_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

Limitations

  • Kernels keep CUDA graphs: their private pool and static outputs stay reserved, so peak reserved is ~80 MiB above portable (peak allocated is not); `cuda_graphs=false` or `release_graphs()` frees it.
Full reference, generated from the repository

EVA-02 (Fang et al. 2023): a vision transformer pre-trained with masked image modelling.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
eva02_large_patch14_448.mim_m38m_ft_in22k_in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
83.7
32.1
18.2
14.1
14.9
12.4
1260 / 1284
1819 / 1864
648 / 664
652 / 786
674 / 684
1259 / 1286
Details
  • Original code: https://github.com/huggingface/pytorch-image-models at 0df212b369a5
  • Code licence: Apache-2.0 (permissive); files:
  • Weights licence: MIT
  • Licence evidence: https://huggingface.co/timm/eva02_large_patch14_448.mim_m38m_ft_in22k_in1k/blob/fae13da496eea422cb162dfdb693dc125b68f04d/README.md
  • Training data: Merged-38M (EVA-02 pre-training mix: ImageNet-22k, CC12M, CC3M, COCO, ADE20K, Objects365, OpenImages), 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: eva02_large_patch14_448.mim_m38m_ft_in22k_in1k (default of eva02_large_patch14_448)
  • Outputs: logits; label_set: imagenet-1k
  • Input size: fixed 448x448: the input is resized (short side 448) and centre-cropped to it
  • Tested up to (portable, fp16, RTX 5060 Ti 16 GB): eva02_large_patch14_448.mim_m38m_ft_in22k_in1k 448x448 (0.7 GiB peak allocated, 0.8 reserved)
  • eva02_large_patch14_448.mim_m38m_ft_in22k_in1k 224x224, first call: portable + torch.compile 21 s (cold Inductor cache), TensorRT fp16 44 s (ONNX + engine).
  • eva02_large_patch14_448.mim_m38m_ft_in22k_in1k original 224x224: torch.autocast(fp16) around the forward
  • eva02_large_patch14_448.mim_m38m_ft_in22k_in1k load / first call: original 6.98 s / 0.389 s, portable 4.54 s / 0.173 s, kernels 5.89 s / 0.165 s
  • Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5

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
eva02_large_patch14_448.mim_m38m_ft_in22k_in1kImageNet-1k val, 5000 random images, top-1 / top-5 %90.62 / 99.1090.62 / 99.1090.46 / 99.0890.46 / 99.0890.50 / 99.08

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
eva02_large_patch14_448.mim_m38m_ft_in22k_in1k0.00154 / 0.00587RTX 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

Limitations

  • Kernels keep CUDA graphs: their private pool and static outputs stay reserved, so peak reserved is ~80 MiB above portable (peak allocated is not); cuda_graphs=false or release_graphs() frees it.

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 73%
  2. tabby 2.7%
  3. tiger cat 2.5%
  4. lynx 0.5%
  5. snow leopard 0.2%
Chelsea the cat, input
Input · 100%

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