Models · Feature backbones

SigLIP 2

Google's SigLIP 2 image encoders (B, L, So400m, g): image features aligned with text, at several resolutions.

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

Checkpoint

Paper 2025 Added in 0.2.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 1.43 ms 224x224, fp16
vs the faster original 4.3× original fp16: 6.11 ms
Peak memory 216 MiB original: 583 MiB
Accuracy vs original fp32 +0.02 kernels top-1 %, ImageNet-1k val (53.84 original)

Use it

pip install fastforks

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

from fastforks import features

model = features.load("siglip2-b16-256")  # kernels on a supported GPU, else portable PyTorch
result = model(image)
result.last_hidden_state  # also: pooler_output, hidden_states
result.backend  # what ran: mode, device, dtype, kernels
fastforks bench siglip2-b16-256

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.

BackenddtypeTime per call (ms)vs original fp16Peak memory
allocated / reserved
Load (s)First call (s)
Original fp32timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)fp32 6.290.97×394 MiB/438 MiB14.20.25
Original fp16torch.autocast(fp16) around the forward; timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)fp16 6.11–583 MiB/620 MiB14.20.18
Portable fp32fp32not measured ––––
Portablefp16 3.281.9×216 MiB/248 MiB–0.07
Kernelsfp16 1.434.3×216 MiB/248 MiB2.10.05
Portable + torch.compile 2.532.4×250 MiB/266 MiB7 (build)–
TensorRT fp16fp16 1.135.4×196 MiBenginenot comparable24 (build)–

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-10 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, 2026-11-13Originalfp3234.1 1,432 MiB process–224x224
Portablefp1615.8 1,563 MiB processmatches224x224
Portablefp3217.2 1,405 MiB processmatches224x224
Kernelsfp16not implemented–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_siglip_256.v2_webli.safetensors 372 MB· hosted on Hugging Face

fastforks downloads and checks these files on first use (fastforks download siglip2-b16-256).

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

kNN top-1 % (k = 20, 10000 labelled / 5000 test images)

Rundtypetop-1 % ↑Δ top-1 % vs fp32 originalimagesms / image
Original fp32fp3253.84 reference 5000 3.68
Original fp16fp1653.78 −0.06 5000 1.13
Portable fp32not measured
Portablefp1653.80 −0.04 5000 0.970
Kernelsfp1653.86 +0.02 5000 0.626

Two disjoint seeded subsets of the val folders; each image shorter side 256 (bicubic), centre crop 224, the CLS / pooled feature; cosine kNN, vote weighted by similarity. Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-11; portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 8bce3316, 2026-11-11.

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| <= 0.0001 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 6.5 % and p99 <= 65 % of the reference's median |value| (absolute error of the output last_hidden_state).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.0001 (relative)not recorded
cpu_rtol≤ 0.00002 (relative)not recorded
GPU default dtypes, mean |diff|≤ 0.065 (relative)0.02629 (relative)
GPU default dtypes, p99 |diff|≤ 0.65 (relative)0.13154 (relative)

Recorded for siglip2-b16-256 on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-13, with these kernels: cudnn_haccfvm_convfvm_extrafvm_mbfvm_r3

Details

SigLIP 2 (Tschannen et al. 2025, Google) image towers: ViT-B / L / So400m / g (fixed-resolution checkpoints), patch tokens + attention-pool embedding at any size; the native resolution is in the id.

Input / output
Input: one RGB image. Output: dense patch features and a pooled image feature.
Input size
any H x W >= 1, resized or padded inside the model; tested up to 3000x4000 (1,726 MiB allocated)
Options
  • fp16_accumulate=false
  • fp32_accumulate=false
Outputs
last_hidden_state, pooler_output, hidden_states
Checkpoint
arch="vit_base_patch16_siglip_256.v2_webli"
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 replays a CUDA graph per small input shape without hidden states (up to about 1M activations: ViT-B ~460 px, ViT-L ~300 px), captured on the shape's second call, at most 4 per model; `model.release_graphs()` frees them.
  • Small kernels calls (graph-sized, with or without `layers`) run GEMMs of at most 600 tokens on cuBLAS: faster in a graph replay (ViT-B/16 224 px: 1.26 against 1.52 ms), about 10 % slower eagerly on ViT-T / B; a replay equals the eager call bit for bit.
  • `fp32_accumulate = true` puts every GEMM on cuBLAS with fp32 accumulation (the portable precision; slower on GeForce).
Full reference, generated from the repository

Google's SigLIP 2 image encoders (B, L, So400m, g): image features aligned with text, at several resolutions.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
siglip2-b16-256synthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
6.29
6.11
3.28
1.43
2.53
1.13
394 / 438
583 / 620
216 / 248
216 / 248
250 / 266
407 / 440
siglip2-b16-512synthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
6.32
5.99
3.31
1.44
2.75
1.14
396 / 438
583 / 618
219 / 266
217 / 268
251 / 286
408 / 460
siglip2-l16-256synthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
17.4
11.1
5.71
3.36
5.15
2.93
1249 / 1270
1843 / 1890
643 / 668
643 / 670
679 / 708
1253 / 1282
siglip2-l16-512synthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
17.5
11.1
5.70
3.35
5.38
2.91
1251 / 1268
1846 / 1888
647 / 666
644 / 670
681 / 706
1252 / 1278
siglip2-so400m14-224synthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
23.9
12.4
6.38
4.32
6.71
3.77
1700 / 1812
2515 / 2632
858 / 868
858 / 892
887 / 908
1659 / 1674
siglip2-so400m14-378synthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
23.9
12.5
6.38
4.33
6.98
3.79
1702 / 1812
2516 / 2632
861 / 908
860 / 912
888 / 926
1660 / 1712
siglip2-so400m16-256synthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
22.8
12.6
6.42
4.21
6.62
3.60
1696 / 1816
2515 / 2636
857 / 868
857 / 890
894 / 908
1657 / 1672
siglip2-so400m16-512synthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
22.8
12.4
6.40
4.21
7.02
3.61
1700 / 1816
2518 / 2636
861 / 906
858 / 910
896 / 928
1658 / 1712
siglip2-g16-256synthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
59.8
30.9
14.7
10.7
15.3
9.99
4528 / 4568
6733 / 6854
2261 / 2288
2262 / 2312
2307 / 2348
4441 / 4474
siglip2-g16-384synthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
59.9
31.0
14.7
10.7
16.9
10.2
4529 / 4568
6735 / 6854
2262 / 2288
2263 / 2312
2308 / 2348
4442 / 4474
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_siglip_256.v2_webli/blob/ac4cc431026674dcac29e34a00d3bd2048390eb3/README.md
  • Training data: WebLI (Google web image-text data, not released) (WebLI image-text pairs (multilingual), sigmoid contrastive loss plus captioning / self-distillation / masked prediction (SigLIP 2 paper); not released)
  • Options: fp16_accumulate = False (precision only), fp32_accumulate = False (precision only)
  • Kernel libraries: cudnn_hacc, fvm_conv, fvm_extra, fvm_mb, fvm_r3
  • Variants: siglip2-b16-256, siglip2-b16-512, siglip2-l16-256, siglip2-l16-512, siglip2-so400m14-224, siglip2-so400m14-378, siglip2-so400m16-256, siglip2-so400m16-512, siglip2-g16-256, siglip2-g16-384
  • Outputs: last_hidden_state, pooler_output, hidden_states
  • Input size: any H x W >= 1, resized or padded inside the model
  • Tested up to (portable, fp16, RTX 5060 Ti 16 GB): siglip2-b16-256 3000x4000 (1.7 GiB peak allocated, 1.9 reserved); siglip2-b16-512 3000x4000 (1.7 GiB peak allocated, 1.8 reserved); siglip2-l16-256 3000x4000 (2.5 GiB peak allocated, 2.8 reserved); siglip2-l16-512 3000x4000 (2.5 GiB peak allocated, 2.7 reserved); siglip2-so400m14-224 3000x4000 (3.2 GiB peak allocated, 3.4 reserved); siglip2-so400m14-378 3000x4000 (3.2 GiB peak allocated, 3.4 reserved); siglip2-so400m16-256 3000x4000 (2.8 GiB peak allocated, 3.1 reserved); siglip2-so400m16-512 3000x4000 (2.8 GiB peak allocated, 3.1 reserved); siglip2-g16-256 3000x4000 (4.7 GiB peak allocated, 4.9 reserved); siglip2-g16-384 3000x4000 (4.7 GiB peak allocated, 4.9 reserved)
  • siglip2-b16-256 224x224, first call: portable + torch.compile 7 s (cold Inductor cache), TensorRT fp16 24 s (ONNX + engine).
  • siglip2-b16-512 224x224, first call: portable + torch.compile 7 s (cold Inductor cache), TensorRT fp16 24 s (ONNX + engine).
  • siglip2-l16-256 224x224, first call: portable + torch.compile 11 s (cold Inductor cache), TensorRT fp16 32 s (ONNX + engine).
  • siglip2-l16-512 224x224, first call: portable + torch.compile 11 s (cold Inductor cache), TensorRT fp16 28 s (ONNX + engine).
  • siglip2-so400m14-224 224x224, first call: portable + torch.compile 12 s (cold Inductor cache), TensorRT fp16 35 s (ONNX + engine).
  • siglip2-so400m14-378 224x224, first call: portable + torch.compile 12 s (cold Inductor cache), TensorRT fp16 32 s (ONNX + engine).
  • siglip2-so400m16-256 224x224, first call: portable + torch.compile 13 s (cold Inductor cache), TensorRT fp16 30 s (ONNX + engine).
  • siglip2-so400m16-512 224x224, first call: portable + torch.compile 13 s (cold Inductor cache), TensorRT fp16 30 s (ONNX + engine).
  • siglip2-g16-256 224x224, first call: portable + torch.compile 18 s (cold Inductor cache), TensorRT fp16 51 s (ONNX + engine).
  • siglip2-g16-384 224x224, first call: portable + torch.compile 19 s (cold Inductor cache), TensorRT fp16 54 s (ONNX + engine).
  • siglip2-b16-256 original (fp32) 224x224: timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)
  • siglip2-b16-256 original 224x224: torch.autocast(fp16) around the forward; timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)
  • siglip2-b16-512 original (fp32) 224x224: timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)
  • siglip2-b16-512 original 224x224: torch.autocast(fp16) around the forward; timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)
  • siglip2-l16-256 original (fp32) 224x224: timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)
  • siglip2-l16-256 original 224x224: torch.autocast(fp16) around the forward; timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)
  • siglip2-l16-512 original (fp32) 224x224: timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)
  • siglip2-l16-512 original 224x224: torch.autocast(fp16) around the forward; timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)
  • siglip2-so400m14-224 original (fp32) 224x224: timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)
  • siglip2-so400m14-224 original 224x224: torch.autocast(fp16) around the forward; timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)
  • siglip2-so400m14-378 original (fp32) 224x224: timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)
  • siglip2-so400m14-378 original 224x224: torch.autocast(fp16) around the forward; timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)
  • siglip2-so400m16-256 original (fp32) 224x224: timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)
  • siglip2-so400m16-256 original 224x224: torch.autocast(fp16) around the forward; timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)
  • siglip2-so400m16-512 original (fp32) 224x224: timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)
  • siglip2-so400m16-512 original 224x224: torch.autocast(fp16) around the forward; timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)
  • siglip2-g16-256 original (fp32) 224x224: timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)
  • siglip2-g16-256 original 224x224: torch.autocast(fp16) around the forward; timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)
  • siglip2-g16-384 original (fp32) 224x224: timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)
  • siglip2-g16-384 original 224x224: torch.autocast(fp16) around the forward; timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)
  • siglip2-b16-256 load / first call: original 14.2 s / 0.182 s, portable – s / 0.075 s, kernels – s / 0.054 s
  • siglip2-b16-512 load / first call: original 14.17 s / 0.178 s, portable – s / 0.07 s, kernels – s / 0.051 s
  • siglip2-l16-256 load / first call: original 40.91 s / 0.189 s, portable – s / 0.054 s, kernels – s / 0.031 s
  • siglip2-l16-512 load / first call: original 41.54 s / 0.185 s, portable – s / 0.052 s, kernels – s / 0.032 s
  • siglip2-so400m14-224 load / first call: original 54.4 s / 0.201 s, portable – s / 0.083 s, kernels – s / 0.073 s
  • siglip2-so400m14-378 load / first call: original 54.79 s / 0.198 s, portable – s / 0.082 s, kernels – s / 0.069 s
  • siglip2-so400m16-256 load / first call: original 54.53 s / 0.2 s, portable – s / 0.084 s, kernels – s / 0.07 s
  • siglip2-so400m16-512 load / first call: original 54.67 s / 0.196 s, portable – s / 0.079 s, kernels – s / 0.067 s
  • siglip2-g16-256 load / first call: original 140.56 s / 0.218 s, portable 13.75 s / 0.111 s, kernels 10.31 s / 0.131 s
  • siglip2-g16-384 load / first call: original 139.99 s / 0.217 s, portable 9.63 s / 0.092 s, kernels 9.84 s / 0.08 s
  • Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5

Measured

Two disjoint seeded subsets of the val folders; each image shorter side 256 (bicubic), centre crop 224, the CLS / pooled feature; cosine kNN, vote weighted by similarity. Original = the original code named under Details, same checkpoint and protocol.

variantdatasetoriginal fp32original fp16portablekernelsnote
siglip2-b16-256ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images)53.8453.7853.8053.86
siglip2-b16-512ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images)26.2426.0626.2826.52
siglip2-l16-256ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images)64.7464.8464.6864.62
siglip2-l16-512ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images)65.1865.1265.1465.10
siglip2-so400m14-224ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images)66.1866.1266.1066.10
siglip2-so400m14-378ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images)59.9059.9460.0059.92
siglip2-so400m16-256ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images)59.5259.4659.3459.50
siglip2-so400m16-512ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images)67.1467.1667.0467.20
siglip2-g16-256ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images)67.9267.8467.8068.00
siglip2-g16-384ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images)69.0269.0669.1269.10

Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 8bce3316, 2026-11-11 to 2026-11-12.

Parity

CPU, float32: max |diff| <= 0.0001 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 6.5 % and p99 <= 65 % of the reference's median |value| (absolute error of the output last_hidden_state).

variantkernels error vs original, mean / p99GPU, torchkernels used
siglip2-b16-2560.02629 / 0.13154RTX 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
siglip2-b16-5120.0262 / 0.10061RTX 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
siglip2-l16-2560.01865 / 0.10027RTX 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
siglip2-l16-5120.03071 / 0.21762RTX 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
siglip2-so400m14-2240.01771 / 0.10073RTX 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
siglip2-so400m14-3780.02225 / 0.13434RTX 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
siglip2-so400m16-2560.02378 / 0.14412RTX 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
siglip2-so400m16-5120.02515 / 0.1561RTX 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
siglip2-g16-2560.03118 / 0.21238RTX 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
siglip2-g16-3840.04305 / 0.42891RTX 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 replays a CUDA graph per small input shape without hidden states (up to about 1M activations: ViT-B ~460 px, ViT-L ~300 px), captured on the shape's second call, at most 4 per model; model.release_graphs() frees them.
  • Small kernels calls (graph-sized, with or without layers) run GEMMs of at most 600 tokens on cuBLAS: faster in a graph replay (ViT-B/16 224 px: 1.26 against 1.52 ms), about 10 % slower eagerly on ViT-T / B; a replay equals the eager call bit for bit.
  • fp32_accumulate = true puts every GEMM on cuBLAS with fp32 accumulation (the portable precision; slower on GeForce).

Example output

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

Input
Output
Chelsea the cat, input
Input · 100%

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