jburtoft/boltz-neuron-cache

boltz-neuron prebuilt NEFF cache (Inferentia2)

Prebuilt Neuron compile cache (NEFF) for running Boltz-2 on AWS Inferentia2 with boltz-neuron.

torch.compile(backend="neuron") compiles once per distinct input shape, which takes minutes and needs a large-RAM host. Restoring this cache lets a small inf2.xlarge ($0.76/hr) run inference fully warm with zero recompilation — no need to compile on a bigger instance first.

What's in it

Shapes covered	8eil example, padded N = 192, diffusion_samples ∈ {1, 8, 25}
Config	bf16 + compile_structure + compile_safe + pad_n 64, fk_steering off
Target	inf2 (Inferentia2 / NeuronCore-v2)
Stack	PyTorch Native Beta 5 — torch 2.12.1, torch-neuronx 2.12.3, neuronx-cc 2.27.2878, nki 0.6.0
Size	73 MB packed / 85 MB unpacked, 1633 NEFF entries

Usage

# after running boltz-neuron's scripts/setup_inf2.sh on an inf2 instance
huggingface-cli download jburtoft/boltz-neuron-cache boltz_neuron_cache_ds1-8-25.tgz \
    --local-dir ~/
bash scripts/restore_cache.sh ~/boltz_neuron_cache_ds1-8-25.tgz

python -m boltz_neuron predict input.yaml --out_dir out \
    --recycling_steps 3 --sampling_steps 50 --diffusion_samples 1 --num_workers 0

Validity — read this

Cache keys depend on both the input shape and the compiler build. This cache only avoids compilation when all of the following hold:

• target is inf2 (not trn2)
• the padded token count is 192 (i.e. real N in 129–192, since pad_n=64 rounds up)
• diffusion_samples is 1, 8, or 25
• the stack is PyTorch Native Beta 5 with neuronx-cc 2.27.2878

Anything outside that recompiles. Note an inf2.xlarge's ~15 GB RAM is too small for the larger graphs, so pre-warm every shape you need on an inf2.8xlarge (scripts/build_cache.sh) and ship the resulting cache.

Verified

Shipped to a fresh inf2.xlarge and run end to end: 0 new NEFFs compiled, valid mmCIF produced (8eil, 1380 atoms), exit 0, 44 s wall, 0 GB swap used.

License

MIT (this cache + the wrapper). Boltz-2 is MIT-licensed by its authors.