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
- Paper SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features arXiv · first released 20 Feb 2025
- Authors’ code google-research/big_vision (JAX), not the code forked here: The authors' code and checkpoints are JAX (big_vision); timm's PyTorch port carries the same weights.
Paper 2025 Added in 0.2.0 Evaluated with 0.2.0
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.
| Backend | dtype | Time per call (ms) | vs original fp16 | Peak 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 | 6.29 | 0.97× | 394 MiB/438 MiB | 14.2 | 0.25 |
| 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 MiB | 14.2 | 0.18 |
| Portable fp32 | fp32 | not measured | – | – | – | – |
| Portable | fp16 | 3.28 | 1.9× | 216 MiB/248 MiB | – | 0.07 |
| Kernels | fp16 | 1.43 | 4.3× | 216 MiB/248 MiB | 2.1 | 0.05 |
| Portable + torch.compile | 2.53 | 2.4× | 250 MiB/266 MiB | 7 (build) | – | |
| TensorRT fp16 | fp16 | 1.13 | 5.4× | 196 MiBenginenot comparable | 24 (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
| Device | Backend | dtype | ms / call | Peak memory | Output vs reference | Input |
|---|---|---|---|---|---|---|
| Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-13 | Original | fp32 | 34.1 | 1,432 MiB process | – | 224x224 |
| Portable | fp16 | 15.8 | 1,563 MiB process | matches | 224x224 | |
| Portable | fp32 | 17.2 | 1,405 MiB process | matches | 224x224 | |
| Kernels | fp16 | not 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 Facefastforks 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)
| Run | dtype | top-1 % ↑ | Δ top-1 % vs fp32 original | images | ms / image |
|---|---|---|---|---|---|
| Original fp32 | fp32 | 53.84 | reference | 5000 | 3.68 |
| Original fp16 | fp16 | 53.78 | −0.06 | 5000 | 1.13 |
| Portable fp32 | not measured | ||||
| Portable | fp16 | 53.80 | −0.04 | 5000 | 0.970 |
| Kernels | fp16 | 53.86 | +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).
| Check | Tolerance | Measured |
|---|---|---|
| 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=falsefp32_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 classifiersfvm_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.
| variant | input | backend | time (ms) | peak memory (MiB, allocated / reserved) |
|---|---|---|---|---|
siglip2-b16-256 | synthetic (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-b16-512 | synthetic (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-l16-256 | synthetic (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-l16-512 | synthetic (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-so400m14-224 | synthetic (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-so400m14-378 | synthetic (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-so400m16-256 | synthetic (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-so400m16-512 | synthetic (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-g16-256 | synthetic (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-g16-384 | synthetic (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-b16-2563000x4000 (1.7 GiB peak allocated, 1.9 reserved);siglip2-b16-5123000x4000 (1.7 GiB peak allocated, 1.8 reserved);siglip2-l16-2563000x4000 (2.5 GiB peak allocated, 2.8 reserved);siglip2-l16-5123000x4000 (2.5 GiB peak allocated, 2.7 reserved);siglip2-so400m14-2243000x4000 (3.2 GiB peak allocated, 3.4 reserved);siglip2-so400m14-3783000x4000 (3.2 GiB peak allocated, 3.4 reserved);siglip2-so400m16-2563000x4000 (2.8 GiB peak allocated, 3.1 reserved);siglip2-so400m16-5123000x4000 (2.8 GiB peak allocated, 3.1 reserved);siglip2-g16-2563000x4000 (4.7 GiB peak allocated, 4.9 reserved);siglip2-g16-3843000x4000 (4.7 GiB peak allocated, 4.9 reserved) siglip2-b16-256224x224, first call: portable + torch.compile 7 s (cold Inductor cache), TensorRT fp16 24 s (ONNX + engine).siglip2-b16-512224x224, first call: portable + torch.compile 7 s (cold Inductor cache), TensorRT fp16 24 s (ONNX + engine).siglip2-l16-256224x224, first call: portable + torch.compile 11 s (cold Inductor cache), TensorRT fp16 32 s (ONNX + engine).siglip2-l16-512224x224, first call: portable + torch.compile 11 s (cold Inductor cache), TensorRT fp16 28 s (ONNX + engine).siglip2-so400m14-224224x224, first call: portable + torch.compile 12 s (cold Inductor cache), TensorRT fp16 35 s (ONNX + engine).siglip2-so400m14-378224x224, first call: portable + torch.compile 12 s (cold Inductor cache), TensorRT fp16 32 s (ONNX + engine).siglip2-so400m16-256224x224, first call: portable + torch.compile 13 s (cold Inductor cache), TensorRT fp16 30 s (ONNX + engine).siglip2-so400m16-512224x224, first call: portable + torch.compile 13 s (cold Inductor cache), TensorRT fp16 30 s (ONNX + engine).siglip2-g16-256224x224, first call: portable + torch.compile 18 s (cold Inductor cache), TensorRT fp16 51 s (ONNX + engine).siglip2-g16-384224x224, first call: portable + torch.compile 19 s (cold Inductor cache), TensorRT fp16 54 s (ONNX + engine).siglip2-b16-256original (fp32) 224x224: timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)siglip2-b16-256original 224x224: torch.autocast(fp16) around the forward; timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)siglip2-b16-512original (fp32) 224x224: timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)siglip2-b16-512original 224x224: torch.autocast(fp16) around the forward; timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)siglip2-l16-256original (fp32) 224x224: timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)siglip2-l16-256original 224x224: torch.autocast(fp16) around the forward; timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)siglip2-l16-512original (fp32) 224x224: timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)siglip2-l16-512original 224x224: torch.autocast(fp16) around the forward; timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)siglip2-so400m14-224original (fp32) 224x224: timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)siglip2-so400m14-224original 224x224: torch.autocast(fp16) around the forward; timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)siglip2-so400m14-378original (fp32) 224x224: timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)siglip2-so400m14-378original 224x224: torch.autocast(fp16) around the forward; timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)siglip2-so400m16-256original (fp32) 224x224: timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)siglip2-so400m16-256original 224x224: torch.autocast(fp16) around the forward; timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)siglip2-so400m16-512original (fp32) 224x224: timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)siglip2-so400m16-512original 224x224: torch.autocast(fp16) around the forward; timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)siglip2-g16-256original (fp32) 224x224: timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)siglip2-g16-256original 224x224: torch.autocast(fp16) around the forward; timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)siglip2-g16-384original (fp32) 224x224: timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)siglip2-g16-384original 224x224: torch.autocast(fp16) around the forward; timm 1.0.30 create_model(dynamic_img_size=True), forward_head(pre_logits=True)siglip2-b16-256load / first call: original 14.2 s / 0.182 s, portable – s / 0.075 s, kernels – s / 0.054 ssiglip2-b16-512load / first call: original 14.17 s / 0.178 s, portable – s / 0.07 s, kernels – s / 0.051 ssiglip2-l16-256load / first call: original 40.91 s / 0.189 s, portable – s / 0.054 s, kernels – s / 0.031 ssiglip2-l16-512load / first call: original 41.54 s / 0.185 s, portable – s / 0.052 s, kernels – s / 0.032 ssiglip2-so400m14-224load / first call: original 54.4 s / 0.201 s, portable – s / 0.083 s, kernels – s / 0.073 ssiglip2-so400m14-378load / first call: original 54.79 s / 0.198 s, portable – s / 0.082 s, kernels – s / 0.069 ssiglip2-so400m16-256load / first call: original 54.53 s / 0.2 s, portable – s / 0.084 s, kernels – s / 0.07 ssiglip2-so400m16-512load / first call: original 54.67 s / 0.196 s, portable – s / 0.079 s, kernels – s / 0.067 ssiglip2-g16-256load / first call: original 140.56 s / 0.218 s, portable 13.75 s / 0.111 s, kernels 10.31 s / 0.131 ssiglip2-g16-384load / 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.
| variant | dataset | original fp32 | original fp16 | portable | kernels | note |
|---|---|---|---|---|---|---|
siglip2-b16-256 | ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images) | 53.84 | 53.78 | 53.80 | 53.86 | |
siglip2-b16-512 | ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images) | 26.24 | 26.06 | 26.28 | 26.52 | |
siglip2-l16-256 | ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images) | 64.74 | 64.84 | 64.68 | 64.62 | |
siglip2-l16-512 | ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images) | 65.18 | 65.12 | 65.14 | 65.10 | |
siglip2-so400m14-224 | ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images) | 66.18 | 66.12 | 66.10 | 66.10 | |
siglip2-so400m14-378 | ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images) | 59.90 | 59.94 | 60.00 | 59.92 | |
siglip2-so400m16-256 | ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images) | 59.52 | 59.46 | 59.34 | 59.50 | |
siglip2-so400m16-512 | ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images) | 67.14 | 67.16 | 67.04 | 67.20 | |
siglip2-g16-256 | ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images) | 67.92 | 67.84 | 67.80 | 68.00 | |
siglip2-g16-384 | ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images) | 69.02 | 69.06 | 69.12 | 69.10 |
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).
| variant | kernels error vs original, mean / p99 | GPU, torch | kernels used |
|---|---|---|---|
siglip2-b16-256 | 0.02629 / 0.13154 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120 |
siglip2-b16-512 | 0.0262 / 0.10061 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120 |
siglip2-l16-256 | 0.01865 / 0.10027 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120 |
siglip2-l16-512 | 0.03071 / 0.21762 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120 |
siglip2-so400m14-224 | 0.01771 / 0.10073 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120 |
siglip2-so400m14-378 | 0.02225 / 0.13434 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120 |
siglip2-so400m16-256 | 0.02378 / 0.14412 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120 |
siglip2-so400m16-512 | 0.02515 / 0.1561 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120 |
siglip2-g16-256 | 0.03118 / 0.21238 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120 |
siglip2-g16-384 | 0.04305 / 0.42891 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_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 = trueputs 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: Chelsea the cat, Stéfan van der Walt (CC0).