bnn-lab-wrap-demo

bnn-lab wrap-demo .bnnpack (canary)

Canary / lab demo — not ImageNet SOTA. Packed binary Linear weights for the committed wrap_demo shape (wide Sequential, hidden=4096, layers 3 and 5). Format: .bnnpack v2 (weights_only load, content hashes).

What 32× means

32× is uint64 pack compression of the replaced Linear weights, not a GPU speedup from sign() / STE. Wall-clock is a separate metric.

Dual-metric AND-gate (same shape — not this file's bytes)

This Hub file is a shape/codec canary (PTQ pack of seed-0 random-init middles). It is not the QAT checkpoint behind the published JSON.

On wrap_demo hidden=4096, after MSE + fold_α QAT (200 steps), committed results/wrap_demo.json:

PTQ-only cosine on this shape is bounded by cosine_max_without_qat: 0.5 (legacy ~0.31). Do not quote PTQ as drop-in.

Not this pack — Ultra TinyBlock still REFUSE

results/ultra_wrap.json hybrid/binary TinyBlock: cosine ~0.70 (0.6988), status REFUSE_DROP_IN_CLAIM. Ternary on that suite meets cosine but loses wall-clock (~0.73×). Ternary 0.991 / 0.73× does not count as the AND-gate.

Load

from huggingface_hub import hf_hub_download
from bnn.codec import decode_file, packed_module_fp_err

path = hf_hub_download("KanakMalpani/bnn-lab-wrap-demo", filename="model.bnnpack")
modules, meta = decode_file(path)
mod = modules["3"]  # PackedBinaryXNORLinear, in=4096 out=4096
print(meta["canary"])
print("GEMM err", packed_module_fp_err(mod))  # 0

Needs pip install bnn-lab huggingface_hub (or pip install -e ".[hf]" from the repo). Tutorial: docs/tutorials/08_HF_OPTIMISER.md.

Offline encode (no Hub):

python scripts/encode_hf_canaries.py --only wrap-demo --out-dir results/hf_canaries

Files

• model.bnnpack — packed layers 3 and 5 (binary XNOR, ~4 MiB packed vs 128 MiB FP32 for those two Linears).
• This card.

License

MIT (same as bnn-lab).