| --- |
| license: cc-by-nc-4.0 |
| pretty_name: jepa.cpp parity fixtures |
| size_categories: |
| - n<1K |
| tags: |
| - jepa |
| - jepa.cpp |
| - ggml |
| - gguf |
| - parity |
| - golden-references |
| - regression-testing |
| --- |
| |
| # jepa.cpp parity fixtures |
|
|
| PyTorch golden reference dumps for [**jepa.cpp**](https://github.com/aselimc/jepa.cpp), a ggml-based C/C++ inference engine for the |
| JEPA family. `tests/test-parity` and `tests/test-predictor` replay these tensors through the engine and |
| gate per-token cosine, pooled outputs and classifier top-1/top-5 against per-family thresholds. |
|
|
| Generated from jepa.cpp `main` @ [`00bfd4e`](https://github.com/aselimc/jepa.cpp/commit/00bfd4e) with |
| `scripts/dump_reference.py --model all`, in float32 eval mode on 32 CPU threads, no autocast. |
|
|
| ## Contents |
|
|
| | directory | model | files | size | |
| |---|---|---:|---:| |
| | `ref/ijepa-vith14-1k/` | I-JEPA ViT-H/14 (IN1k) | 25 | 15 MB | |
| | `ref/lejepa-vits16/` | LeJEPA ViT-S/16 (IN1k) | 41 | 7 MB | |
| | `ref/levjepa-vitl16/` | LeVJEPA ViT-L/16 (VideoMix) | 21 | 136 MB | |
| | `ref/lewm-pusht/` | LeWorldModel Push-T | 20 | 3 MB | |
| | `ref/vjepa2-ac-vitg/` | V-JEPA 2-AC ViT-g (action-conditioned world model) | 45 | 56 MB | |
| | `ref/vjepa2-vitg-fpc64-256/` | V-JEPA 2 ViT-g/16 (fpc64, 256) | 21 | 429 MB | |
| | `ref/vjepa2-vitl-fpc16-256-ssv2/` | V-JEPA 2 ViT-L/16 SSv2 classifier | 13 | 56 MB | |
| | `ref/vjepa2-vitl-fpc64-256/` | V-JEPA 2 ViT-L/16 (fpc64, 256) | 21 | 366 MB | |
| | `ref/vjepa2_1-vitb-384/` | V-JEPA 2.1 ViT-B/16 @384 | 15 | 107 MB | |
|
|
| 222 files, 1176 MB in total. |
|
|
| ## Layout |
|
|
| One directory per model, each with a `manifest.json` and one `.npy` per tensor per sample, named |
| `<sample>.<tensor>.npy`. All arrays are float32 C-order except `frames_u8` (uint8) and `top5_idx` |
| (int64). Shapes carry no batch dimension except `input`, which is stored exactly as fed to the model. |
|
|
| The manifest records the model id, the hyper-parameters, the preprocessing pipeline that produced |
| `input`, the PyTorch forward wall time per sample (`timing_s.forward_s`, the baseline for the speed |
| tables of the jepa.cpp docs), the frame indices sampled from each clip, and the label strings for the |
| classifier. Per-model tensor lists and the exact preprocessing recipe: |
| [**docs/fixtures.md**](https://aselimc.github.io/jepa.cpp/fixtures/). |
|
|
| ## How jepa.cpp consumes it |
|
|
| ```bash |
| 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 |
| |
| scripts/download_fixtures.sh # this dataset -> tests/fixtures/ref, plus the media it needs |
| scripts/download_models.sh small # GGUFs from https://huggingface.co/jepacpp |
| |
| cmake -S . -B build && ctest --test-dir build # the parity suites register at configure time |
| ``` |
|
|
| `test-parity` runs two passes per file: first the stored `input` tensor (bypassing preprocessing, so a |
| graph bug shows up alone), then jepa.cpp's own preprocessor on the source media (so a preprocessing |
| mismatch shows up separately). The second pass needs `tests/fixtures/media/`, which is **not** part of |
| this dataset — see below. |
|
|
| ## Input provenance |
|
|
| The dumps are model outputs computed on two public research sets. Neither the source images nor the |
| source videos are redistributed here; only the reference activations and the decoded frame tensors the |
| tests replay are. |
|
|
| | input | source | fetched by | |
| |---|---|---| |
| | 8 images `coco_*.jpg` | [COCO val2017](https://cocodataset.org/), images subject to their original Flickr terms | tracked in the jepa.cpp git repo | |
| | 6 short clips | [`nateraw/kinetics-mini`](https://huggingface.co/datasets/nateraw/kinetics-mini), a small sample of Kinetics-400 | `scripts/download_fixtures.sh` | |
|
|
| The `input` and `frames_u8` arrays inside `ref/` are preprocessed pixels of those images and clips, kept |
| because the parity tests must feed the network exactly the tensor PyTorch saw. If you hold rights in any |
| of the underlying material and want it removed, open an issue on [github.com/aselimc/jepa.cpp](https://github.com/aselimc/jepa.cpp/issues). |
|
|
| ## Licence |
|
|
| `cc-by-nc-4.0`, the most restrictive licence among the checkpoints whose outputs are stored here: the |
| dumps of I-JEPA ViT-H/14 (IN1k) and LeVJEPA ViT-L/16 (VideoMix) are outputs of CC BY-NC 4.0 models, the rest of MIT / Apache-2.0 ones. |
| Research and non-commercial use. |
|
|
| ## Regenerating instead of downloading |
|
|
| ```bash |
| scripts/download_fixtures.sh media # the source media |
| scripts/download_models.sh --convert all # the source checkpoints |
| .venv/bin/python scripts/dump_reference.py --model all # ~1 min on 32 cores |
| ``` |
|
|
| ## Links |
|
|
| - Code: <https://github.com/aselimc/jepa.cpp> |
| - Documentation: <https://aselimc.github.io/jepa.cpp/> |
| - GGUF models: <https://huggingface.co/jepacpp> |
|
|
