Skip to content

GPU lane: INT4 / FP8 (not classic BNN)

On commodity NVIDIA servers, use industry low-bit stacks — not CUDA XNOR BNNs.

Goal Tooling
Weight-only INT4 AWQ / GPTQ / bitsandbytes / torchao
FP8 train/serve TransformerEngine / torchao FP8 + vLLM / TensorRT-LLM
Throughput serve vLLM / SGLang / TensorRT

Explicit non-goal

Classic binary networks beating Tensor Core FP8/INT8 on datacenter GPUs. See arXiv:1911.04477 and ADR docs/08.

This machine

Often CPU-only PyTorch — run GPU recipes elsewhere; keep docs honest.

CLI (preferred)

bnn bridge list
bnn bridge gpu --probe
bnn bridge torchao --probe

Concrete recipe script

python scripts\bridges\torchao_int4_recipe.py --probe

Writes results/bridge_gpu_torchao.json (torchao INT4 + AWQ pointers). Does not install CUDA deps by default.

bnn recommend --goal gpu-server