Models · Classification

EfficientFormer

EfficientFormer (Li et al. 2022, Snap Research; L1, L3, L7; distilled on ImageNet-1k): a hybrid of pooling blocks and transformer blocks built for phone latency. Apache-2.0.

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

Checkpoint

Paper 2022 Added in 0.2.0 Evaluated with 0.2.0

Portable, RTX 5060 Ti 16 GB 3.72 ms 224x224, fp16
vs the faster original 1.3× original fp32: 4.87 ms
Peak memory 93 MiB original: 87 MiB
Accuracy vs original fp32 ±0.00 portable top-1 %, ImageNet-1k val (81.14 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("efficientformer_l1.snap_dist_in1k")  # kernels on a supported GPU, else portable PyTorch
result = model(image)
result.logits
result.backend  # what ran: mode, device, dtype, kernels
fastforks bench efficientformer_l1.snap_dist_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 fp32Peak memory
allocated / reserved
Load (s)First call (s)
Original fp32fp32 4.87–87 MiB/98 MiB3.80.35
Original fp16torch.autocast(float16) around the callfp16 6.400.76×104 MiB/124 MiB0.20.22
Portable fp32fp32not measured ––––
Portablefp16 3.721.3×93 MiB/104 MiB1.70.07
Kernelsnot measured ––––
Portable + torch.compile 1.692.9×125 MiB/144 MiB9 (build)–
TensorRT fp16fp16 0.33515×38 MiBenginenot comparable35 (build)–

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-11. 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, 2026-11-14Originalfp3212.5 550 MiB process–224x224
Portablefp169.24 620 MiB processmatches224x224
Portablefp328.57 589 MiB processmatches224x224
Kernelsfp32not 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

efficientformer_l1.snap_dist_in1k.safetensors 49.3 MB· hosted on Hugging Face

fastforks downloads and checks these files on first use (fastforks download efficientformer_l1.snap_dist_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 fp32fp3281.1495.22 reference 5000 0.897
Original fp16fp1681.1695.24 +0.02 5000 0.612
Portable fp32fp3281.1495.14 ±0.00 5000 0.911
Portablefp1681.1495.12 ±0.00 5000 0.503
Kernelsnot measured

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, 2026-11-11; portable fp32: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 6f0a70e1, 2026-11-17; portable: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 7ba7307f, 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| <= 1.3e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.0046 and p99 <= 0.019 (absolute error of the output logits).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.000013 not recorded
GPU default dtypes, mean |diff|≤ 0.0046 0.00293
GPU default dtypes, p99 |diff|≤ 0.019 0.01165

Recorded for efficientformer_l1.snap_dist_in1k on RTX 5060 Ti, torch 2.10.0+cu130, 2026-11-17, with these kernels: cudnn_haccfvm_convfvm_extrafvm_mbfvm_r3

Details

EfficientFormer (Li et al. 2022, Snap Research): a hybrid classifier, pooling blocks on feature maps followed by transformer blocks on tokens.

Input / output
Input: one RGB image. Output: ImageNet-1k class logits.
Input size
fixed 224x224: the input is resized (short side 235) and centre-cropped to it; tested up to 224x224 (93 MiB allocated)
Outputs
logits
Checkpoint
arch="efficientformer_l1.snap_dist_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

EfficientFormer (Li et al. 2022, Snap Research; L1, L3, L7; distilled on ImageNet-1k): a hybrid of pooling blocks and transformer blocks built for phone latency. Apache-2.0.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
efficientformer_l1.snap_dist_in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
4.87
6.40
3.72
–
1.69
0.335
87 / 98
104 / 124
93 / 104
–
125 / 144
128 / 142
efficientformer_l3.snap_dist_in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
8.39
11.0
6.37
–
2.57
0.750
164 / 172
214 / 236
161 / 216
–
193 / 216
227 / 286
efficientformer_l7.snap_dist_in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
12.3
15.9
9.65
–
3.69
1.40
363 / 378
505 / 522
300 / 324
–
332 / 364
460 / 490
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/efficientformer_l1.snap_dist_in1k/blob/8738a58190800aaf5995f453b05f643df453a565/README.md
  • Training data: ImageNet-1k (ILSVRC 2012 classification)
  • Kernel libraries: cudnn_hacc, fvm_conv, fvm_extra, fvm_mb, fvm_r3
  • Variants: efficientformer_l1.snap_dist_in1k (default of efficientformer_l1), efficientformer_l3.snap_dist_in1k (default of efficientformer_l3), efficientformer_l7.snap_dist_in1k (default of efficientformer_l7)
  • Outputs: logits; label_set: imagenet-1k
  • Input size: fixed 224x224: the input is resized (short side 235) and centre-cropped to it
  • Tested up to (portable, fp16, RTX 5060 Ti 16 GB): efficientformer_l1.snap_dist_in1k 224x224 (0.1 GiB peak allocated, 0.1 reserved); efficientformer_l3.snap_dist_in1k 224x224 (0.2 GiB peak allocated, 0.2 reserved); efficientformer_l7.snap_dist_in1k 224x224 (0.3 GiB peak allocated, 0.3 reserved)
  • efficientformer_l1.snap_dist_in1k 224x224, first call: portable + torch.compile 9 s (cold Inductor cache), TensorRT fp16 35 s (ONNX + engine).
  • efficientformer_l3.snap_dist_in1k 224x224, first call: portable + torch.compile 12 s (cold Inductor cache), TensorRT fp16 39 s (ONNX + engine).
  • efficientformer_l7.snap_dist_in1k 224x224, first call: portable + torch.compile 14 s (cold Inductor cache), TensorRT fp16 40 s (ONNX + engine).
  • efficientformer_l1.snap_dist_in1k original 224x224: torch.autocast(float16) around the call
  • efficientformer_l3.snap_dist_in1k original 224x224: torch.autocast(float16) around the call
  • efficientformer_l7.snap_dist_in1k original 224x224: torch.autocast(float16) around the call
  • efficientformer_l1.snap_dist_in1k load / first call: original 0.22 s / 0.219 s, portable 1.71 s / 0.071 s
  • efficientformer_l3.snap_dist_in1k load / first call: original 0.37 s / 0.029 s, portable 1.86 s / 0.076 s
  • efficientformer_l7.snap_dist_in1k load / first call: original 0.93 s / 0.03 s, portable 2.27 s / 0.074 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 fp32note
efficientformer_l1.snap_dist_in1kImageNet-1k val, 5000 random images, top-1 / top-5 %81.14 / 95.2281.16 / 95.2481.14 / 95.1281.14 / 95.14
efficientformer_l3.snap_dist_in1kImageNet-1k val, 5000 random images, top-1 / top-5 %83.12 / 96.3283.14 / 96.3283.16 / 96.3283.16 / 96.32
efficientformer_l7.snap_dist_in1kImageNet-1k val, 5000 random images, top-1 / top-5 %84.44 / 96.6884.44 / 96.6684.38 / 96.6884.38 / 96.66

Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 7ba7307f, 2026-11-11; portable32: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 6f0a70e1, 2026-11-17.

Parity

CPU, float32: max |diff| <= 1.3e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.0046 and p99 <= 0.019 (absolute error of the output logits).

variantkernels error vs original, mean / p99GPU, torchkernels used
efficientformer_l1.snap_dist_in1k0.00293 / 0.01165RTX 5060 Ti, torch 2.10.0+cu130cudnn_hacc@cudnn91501, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
efficientformer_l3.snap_dist_in1k0.00344 / 0.01248RTX 5060 Ti, torch 2.10.0+cu130cudnn_hacc@cudnn91501, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
efficientformer_l7.snap_dist_in1k0.00193 / 0.00768RTX 5060 Ti, torch 2.10.0+cu130cudnn_hacc@cudnn91501, 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 9.8%
  3. tabby 2.1%
  4. lynx 0.1%
  5. kit fox <0.1%
Chelsea the cat, input
Input · 100%

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