vjepa2-vitl-fpc64-256-GGUF

V-JEPA 2 ViT-L/16 (fpc64, 256) — GGUF for jepa.cpp

Meta's V-JEPA 2 ViT-L/16 video encoder with its masked latent predictor, converted to GGUF for jepa.cpp — a ggml C/C++ engine that runs it on a plain CPU with no Python and no PyTorch. Tubelets of two frames and 3-D RoPE in Meta's tiled layout; a whole clip goes through one graph.

326 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-fpc64-256-GGUF vjepa2-vitl-fpc64-256-f16.gguf --local-dir models/gguf

# a clip (THWC uint8 .npy, written by scripts/video_frames.py) -> a pooled feature
build/jepa-embed -m vjepa2-vitl-fpc64-256-f16.gguf --frames-npy clip.npy --pool mean -t 32 -o feat.npy

--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.

Full checksums:

3cfa5734f46140876aa4a49e914bf9f050e481253c4362e57ef4529a0820bd4d  vjepa2-vitl-fpc64-256-f32.gguf
7ee7643dd2ef8dcee6fddf8c5a13f9115c134030e724caec8993144b5acd00ed  vjepa2-vitl-fpc64-256-f16.gguf
2503b98662e7867afcf6b5ada294c76c1f8f8d72900b0a2ba5ae4ce9a5b48169  vjepa2-vitl-fpc64-256-q8_0.gguf
2d76e582432dae2f1b71a5a7f271e058d3b9d98094dc6968ffffa6f82a918600  vjepa2-vitl-fpc64-256-q4_0.gguf
c6c402e55e5508a062c697703c4660b0017da17432b2ff66a287f6f071cb219a  vjepa2-vitl-fpc64-256-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-fpc64-256-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.

UCF-101 k-NN — 10 classes, 105 query clips (val+test) against a gallery of 300, 16 frames per clip, k = 20 cosine vote over frozen features. Nothing is trained.

Speed — the encoder graph at f16 on 32 threads (AMD Ryzen Threadripper PRO 7995WX 96-Cores): 821 ms per 16-frame clip; 6388 ms per 64-frame clip. The same shape on NVIDIA RTX 4500 Ada Generation: 46.5 ms. Peak RSS at f16: 1034 MiB.

Use f32 if you consume individual tokens of this model. Its activation range contains a degenerate low-norm token cluster that the F16 activation rounding inside ggml's mul_mat collapses — a property of the checkpoint, reproduced in numpy, not an engine defect, and the reason its f16 and q8_0 worst-token columns above read so much lower than its pooled ones. Everything pooled is unaffected. The mechanism is worked through in accuracy.

Source, licence and attribution

Converted from facebook/vjepa2-vitl-fpc64-256.

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 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
python scripts/convert.py --family vjepa2 --src models/facebook/vjepa2-vitl-fpc64-256 --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-fpc64-256-f16.gguf \
      models/gguf/vjepa2-vitl-fpc64-256-$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