jepacpp/vjepa2-vitl-fpc16-256-ssv2-GGUF

V-JEPA 2 ViT-L/16 SSv2 classifier — GGUF for jepa.cpp

Meta's V-JEPA 2 ViT-L/16 encoder with its Something-Something-v2 attentive-pooler head — 174 action classes from one clip — converted to GGUF for jepa.cpp, a ggml C/C++ engine that runs it on a plain CPU with no Python and no PyTorch. The 174 label strings travel inside the GGUF, so no side file is needed.

375 M parameters; D = 1024, 24 layers, 16 heads, patch 16, tubelet 2, 256x256. Everything the engine needs — dimensions, positional scheme, preprocessing recipe, and class labels where there are any — travels inside the file, so inference needs one binary and one GGUF and nothing else.

Run it

git clone --recursive https://github.com/aselimc/jepa.cpp && cd jepa.cpp
cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release && cmake --build build -j
hf download jepacpp/vjepa2-vitl-fpc16-256-ssv2-GGUF vjepa2-vitl-fpc16-256-ssv2-f16.gguf --local-dir models/gguf

# a 16-frame clip -> the top 5 of 174 actions, or the pooled feature
build/jepa-classify -m vjepa2-vitl-fpc16-256-ssv2-f16.gguf --frames-npy clip.npy -k 5 -t 32
build/jepa-embed    -m vjepa2-vitl-fpc16-256-ssv2-f16.gguf --frames-npy clip.npy --pool mean -t 32

--pool selects mean, cls, lewm or none (the full token map); -o writes a .npy. scripts/download_models.sh fetches whole sets at once. The C API is one header, `include/jepa.h` — full reference on the C API page.

Files

_{ᵖ tests/test-parity on the CPU backend, stored reference input, 32 threads, worst sample — docs/parity.md. ᵈ scripts/gguf_dequant_selftest.py: the dequantized weights through the numpy reference graph at f32 activations, so the figure is the weight error alone — docs/quantization.md. cos mean is the mean per-token cosine of last_hidden_state, worst its single worst token.}

Tiers. exact — reproduces the PyTorch reference to the printed precision on the CPU. parity — passes its family's test-parity thresholds. advisory — below 8 bits per weight, which is not a parity configuration: the results are reported, only the derived tensors and the top-1 are gated. Which file to ship: Accuracy → which dtype. vjepa2-vitl-fpc16-256-ssv2-q4_k.gguf misses even the advisory derived-tensor bar (logits and pooled), on the CPU and on CUDA alike.

Full checksums:

4eb87e2691f68bcdc2ab5f2640031cfa1c1b1996e8ec54ed8b5ff400a83c4f52  vjepa2-vitl-fpc16-256-ssv2-f32.gguf
9c85a9abe92fa8c416f76528dcc26bfefed63920d38e5f6cdaf34f9d84575141  vjepa2-vitl-fpc16-256-ssv2-f16.gguf
a0fa64702b6a8dbc5efb163f9135242db622cb81ef5f74520f5bd9152a16f481  vjepa2-vitl-fpc16-256-ssv2-q8_0.gguf
a004c349542957fd64c53af2d97d6b840f8c498ea36305146c6e4d26d7a67cd9  vjepa2-vitl-fpc16-256-ssv2-q4_0.gguf
2221adb97d4a3541343a641fb45bd0d1a6a96ca90c15af9a4d7e7c32923cc113  vjepa2-vitl-fpc16-256-ssv2-q4_k.gguf

Verify a download with sha256sum -c. The other types jepa-quantize can produce (q4_1, q5_0, q5_1, q5_k, q6_k, measured in quantization) are not published here; make them locally with build/jepa-quantize vjepa2-vitl-fpc16-256-ssv2-f16.gguf out.gguf q6_k -t 32.

Measured

Every figure below is read from a committed artifact of jepa.cpp `00bfd4e` by scripts/hf_publish.py — parity, quantization, accuracy, performance and tests/results/*.json.

Something-Something-v2 validation, all 24,777 clips, 174 classes, one view per clip and no test-time augmentation. Both engines read the same uniformly sampled frames and each runs the preprocessing itself.

On the CPU, over a fixed 2,478-clip subset (every 10th clip of the validation order):

Speed — the encoder graph at f16 on 32 threads (AMD Ryzen Threadripper PRO 7995WX 96-Cores): 823 ms per 16-frame clip. Peak RSS at f16: 808 MiB.

f16 is the recommendation for classifier work. Quantization moves decisions without moving the score: read the agreement column, not the top-1 delta, as the cost of a low-bit file. The full argument, with the per-clip breakdown, is in accuracy.

Source, licence and attribution

Converted from `facebook/vjepa2-vitl-fpc16-256-ssv2`.

MIT. The source checkpoint is published by Meta AI (FAIR): license: mit on the model card and LICENSE in facebookresearch/vjepa2 (Copyright (c) Meta Platforms, Inc. and affiliates). There is no separate weights licence, no gating and no acceptable-use policy. These GGUF files are the same weights re-serialised into the GGUF container, quantized where the file name says so. Cite the V-JEPA 2 paper. The head was fine-tuned on Something-Something v2, whose dataset terms are a separate matter from this weight licence.

The licence travels inside every GGUF as general.license and the origin as general.source_url; build/jepa-info <file> --kv prints them. jepa.cpp's own code is MIT.

Conversion

Produced by jepa.cpp `00bfd4e`:

scripts/download_models.sh --convert vjepa2-ssv2
python scripts/convert.py --family vjepa2 --src models/facebook/vjepa2-vitl-fpc16-256-ssv2 --ftype f16
#   ... and again with --ftype f32 for the f32 file

for q in q8_0 q4_0 q4_k; do
  build/jepa-quantize models/gguf/vjepa2-vitl-fpc16-256-ssv2-f16.gguf \
      models/gguf/vjepa2-vitl-fpc16-256-ssv2-$q.gguf $q -t 32
done

jepa-quantize re-types only the 2-D attention / FFN / projection / classifier matrices; patch embeddings, position tables, norms and biases keep the source type. The rules are in docs/gguf-schema.md.

Links

• —Code: <https://github.com/aselimc/jepa.cpp>
• —Documentation: <https://aselimc.github.io/jepa.cpp/>
• —Parity fixtures: <https://huggingface.co/datasets/jepacpp/jepa.cpp-fixtures>
• —All jepa.cpp GGUFs: <https://huggingface.co/jepacpp>