RelateAnything GGUF adapter test fixtures
This repository currently publishes deterministic relation_pair_linear_v1
GGUF test fixtures in F32, F16, and Q8_0. They validate conversion, GGML
CPU/CUDA/Vulkan execution, and raw-logit comparisons. They contain none of
the released relsgg-vits16, relsgg-vits16plus, or relsgg-vitb16 DINOv3
backbones or dynamic relation heads. Do not use their speed or predictions as
a full RelateAnything deployment result.
| Archived file in this repository | Size | Use |
|---|---|---|
relation_pair_linear_v1-demo-f32.gguf |
5,696 B | adapter numerical reference |
relation_pair_linear_v1-demo-f16.gguf |
3,328 B | half-precision adapter check |
relation_pair_linear_v1-demo-q8_0.gguf |
2,272 B | quantized adapter check |
Local copies now live under cpp_ggml/test_data/fixtures/ with -fixture-
names. The old demo-f32 and demo-f32-v2 variants were deleted locally;
the old demo-f32-v3, demo-f16, and demo-q8_0 aliases are byte-identical
to the three rows above. These archive names are not official model variants.
The official PyTorch checkpoints and their intended use are listed by upstream. Source code, SHA256 hashes, actual benchmark results, and the full-model conversion status are in the RelateAnything-GGML model cards.
- Downloads last month
- 162
8-bit
16-bit
32-bit