RAM-GGUF

RAM family GGUF exports

Current downloads: GitHub RAM release and
Hugging Face Asher-1/RAM_GGUF.

Nine GGUF models converted from the repository's three official checkpoints:
RAM Swin-L, RAM++ Swin-L and Tag2Text Swin-B, each in F32, F16 and Q8_0.
The release asset names preserve the source checkpoint stem. GGML is pinned
to v0.21.0. The repository's C++ runner supports tagging on CUDA/Vulkan/CPU
and Tag2Text caption generation with a 3-beam BERT decoder. No cuDNN is used.

Current GGUF files contain weights plus tag names, per-tag thresholds,
Tag2Text delete indices, and the BERT uncased WordPiece vocabulary as GGUF
metadata. Tag2Text exports also retain the official threshold array and a
calibrated deployment array; only large and some use 0.6795 instead of
0.68 to account for measured backend decision-boundary drift. The checked-in
files remain fallbacks for older exports. RAM and
RAM++ emit 4585 tags; Tag2Text emits 3429 tags and can generate captions.
On the documented RTX 3060 run, all six CUDA/Vulkan Tag2Text caption variants
matched Python sentence-for-sentence on 50 retained GT images and 18 repository
images. Caption latency was lower for 6/6 variants; the tag latency gate was
met for 16/18 combinations. RAM and RAM++ F32 remain slower on Vulkan in the
latest matrix, so a universal speedup is not claimed.
Consult cpp_ggml/models/MODEL_CARD.md and
cpp_ggml/benchmarks/ in the recognize-anything-ggml repository for
download links, SHA-256 checksums, accuracy and pipeline-speed measurements.

These exports supersede the old ram-f32.gguf, ram-f16.gguf and
ram-q8_0.gguf assets. Existing consumers should switch to the new
ram_swin_large_14m-* names and use the current runtime metadata.