Models · Classification

SigLIP (fine-tuned)

SigLIP So400m/14 (Zhai et al. 2023) fine-tuned on ImageNet-1k.

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

Checkpoint vit_so400m_patch14_siglip_378.webli_ft_in1k

Paper 2023 Added in 0.1.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 12.9 ms 224x224, fp16
vs the faster original 2.0× original fp16: 25.8 ms
Peak memory 876 MiB original: 2,495 MiB
Accuracy vs original fp32 −0.04 kernels top-1 %, ImageNet-1k val (90.04 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_so400m_patch14_siglip_378.webli_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_so400m_patch14_siglip_378.webli_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 76.90.34×1,710 MiB/1,744 MiB5.00.16
Original fp16torch.autocast(fp16) around the forwardfp16 25.8–2,495 MiB/2,564 MiB7.50.35
Portable fp32fp32not measured ––––
Portablefp16 13.12.0×879 MiB/1,708 MiB7.30.07
Kernelsfp16 12.92.0×876 MiB/926 MiB6.00.10
Portable + torch.compile 13.31.9×909 MiB/944 MiB14 (build)–
TensorRT fp16fp16 11.42.3×848 MiBenginenot comparable61 (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, 2026-11-12Originalfp32215 3,973 MiB processmatches224x224
Portablefp16197 5,547 MiB processmatches224x224
Portablefp32214 5,547 MiB processmatches224x224
Kernelsfp16not implemented–224x224
NVIDIA L4 CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp1616.4 853 MiB/864 MiB–224x224
Kernelsfp1615.8 1,174 MiB/1,266 MiB–224x224
NVIDIA L40S CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp168.01 866 MiB/904 MiB–224x224
Kernelsfp166.99 1,190 MiB/1,286 MiB–224x224
NVIDIA A10 CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp1616.4 853 MiB/864 MiB–224x224
Kernelsfp1616.0 1,174 MiB/1,266 MiB–224x224
NVIDIA A100-SXM4-40GB CUDA, torch 2.14.1+cu130, 2026-11-08Portablefp167.96 853 MiB/864 MiB–224x224
Kernelsfp168.73 1,174 MiB/1,266 MiB–224x224
NVIDIA RTX PRO 6000 CUDA, torch 2.14.1+cu130, 2026-11-08Portablefp164.80 877 MiB/904 MiB–224x224
Kernelsfp165.17 1,214 MiB/1,346 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_so400m_patch14_siglip_378.webli_ft_in1k.safetensors 1.7 GB· hosted on Hugging Face

fastforks downloads and checks these files on first use (fastforks download vit_so400m_patch14_siglip_378.webli_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 fp32fp3290.0498.90 reference 5000 72.8
Original fp16fp1690.0298.88 −0.02 5000 21.3
Portable fp32fp3290.0298.90 −0.02 5000 71.3
Portablefp1689.9898.94 −0.06 5000 12.7
Kernelsfp1690.0098.92 −0.04 5000 11.3

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 0.00232
GPU default dtypes, p99 |diff|≤ 0.08 0.00745

Recorded for vit_so400m_patch14_siglip_378.webli_ft_in1k on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-04, with these kernels: cudnn_haccfvm_convfvm_extrafvm_mbfvm_r3

Details

SigLIP So400m/14 image tower (Zhai et al. 2023) fine-tuned on ImageNet-1k with an attention-pool head.

Input / output
Input: one RGB image. Output: ImageNet-1k class logits.
Input size
fixed 378x378: the input is resized (short side 378) and centre-cropped to it; tested up to 378x378 (901 MiB allocated)
Outputs
logits
Checkpoint
arch="vit_so400m_patch14_siglip_378.webli_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

SigLIP So400m/14 (Zhai et al. 2023) fine-tuned on ImageNet-1k.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
vit_so400m_patch14_siglip_378.webli_ft_in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
76.9
25.8
13.1
12.9
13.3
11.4
1710 / 1744
2495 / 2564
879 / 1708
876 / 926
909 / 944
1701 / 1752
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_so400m_patch14_siglip_378.webli_ft_in1k/blob/efb07b5711bda4e3eb24630db9879e0832615ad4/README.md
  • Training data: WebLI (Google web image-text data, not released), ImageNet-1k (ILSVRC 2012 classification)
  • Kernel libraries: cudnn_hacc, fvm_conv, fvm_extra, fvm_mb, fvm_r3
  • Variants: vit_so400m_patch14_siglip_378.webli_ft_in1k (default of vit_so400m_patch14_siglip_378)
  • Outputs: logits; label_set: imagenet-1k
  • Input size: fixed 378x378: the input is resized (short side 378) and centre-cropped to it
  • Tested up to (portable, fp16, RTX 5060 Ti 16 GB): vit_so400m_patch14_siglip_378.webli_ft_in1k 378x378 (0.9 GiB peak allocated, 1.0 reserved)
  • vit_so400m_patch14_siglip_378.webli_ft_in1k 224x224, first call: portable + torch.compile 14 s (cold Inductor cache), TensorRT fp16 61 s (ONNX + engine).
  • vit_so400m_patch14_siglip_378.webli_ft_in1k original 224x224: torch.autocast(fp16) around the forward
  • vit_so400m_patch14_siglip_378.webli_ft_in1k load / first call: original 7.52 s / 0.346 s, portable 7.29 s / 0.069 s, kernels 5.82 s / 0.093 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_so400m_patch14_siglip_378.webli_ft_in1kImageNet-1k val, 5000 random images, top-1 / top-5 %90.04 / 98.9090.02 / 98.8889.98 / 98.9490.02 / 98.9090.00 / 98.92

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_so400m_patch14_siglip_378.webli_ft_in1k0.00232 / 0.00745RTX 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 (portable · float16 · NVIDIA GeForce RTX 5060 Ti).

Input

Top 5

  1. Egyptian cat 86%
  2. tiger cat 3.0%
  3. tabby 2.6%
  4. lynx 0.2%
  5. Ibizan hound <0.1%
Chelsea the cat, input
Input · 100%

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