jburtoft/boltz-neuron-cache

boltz-neuron prebuilt caches (Inferentia2, Trainium2)

Prebuilt compile caches for running Boltz-2 on AWS Inferentia2 and Trainium2 with boltz-neuron.

torch.compile(backend="neuron") compiles once per input shape, which takes minutes and about 40 GB of host RAM per graph. Restoring a cache lets an inf2.xlarge run without compiling.

Which file

file	Neuron stack	use with
boltz_neuron_cache_v8_sdk233_ds1-8-25.tgz (52 MB)	Inferentia2, public Neuron SDK 2.33 torch-neuronx DLAMI (torch 2.14.0, torch-neuronx 2.14.3, neuronx-cc 2.28.7631, nki 0.7.0)	boltz-neuron master on inf2
boltz_neuron_cache_v8_sdk233_trn2_ds1-8-25.tgz (53 MB)	Trainium2, same SDK 2.33 stack, built at LNC=1 (the package default on trn2)	boltz-neuron master on trn2
boltz_neuron_cache_v7_ds1-8-25.tgz	PyTorch Native Beta 5 (torch 2.12.1, torch-neuronx 2.12.3, neuronx-cc 2.27.2878)	the earlier Beta 5 build only

Older v1-v6 files are superseded; do not use them.

Shapes covered

The repository's example input (examples/8eil.yaml): 186 tokens padded to 192, 1,380 atoms padded to 1,408, --diffusion_samples 1, 8 and 25, with --recycling_steps 3 --sampling_steps 50. A compiled shape is fixed by the padded token count, the padded atom count and --diffusion_samples, so a different protein usually needs its own compile. Build caches for your inputs with scripts/build_cache.sh on an inf2.8xlarge.

The v8 archive holds the NEFF cache and the AOTAutograd cache (no dynamo precompile package: it does not load on torch 2.14). The cache key includes a hash of the installed torch_neuronx sources, so it only matches an SDK 2.33 DLAMI set up by boltz-neuron's scripts/setup_inf2.sh.

Usage

# on an inf2 or trn2 instance from the SDK 2.33 torch-neuronx DLAMI, after boltz-neuron's scripts/setup_inf2.sh
source /opt/aws_neuronx_venv_torch_neuronx_2_14_0/bin/activate
CACHE=boltz_neuron_cache_v8_sdk233_ds1-8-25.tgz          # inf2
# CACHE=boltz_neuron_cache_v8_sdk233_trn2_ds1-8-25.tgz   # trn2
hf download jburtoft/boltz-neuron-cache $CACHE --repo-type dataset --local-dir ~/
bash scripts/restore_cache.sh ~/$CACHE

export NEURON_RT_VISIBLE_CORES=0
python -m boltz_neuron predict examples/8eil.yaml --out_dir out \
    --recycling_steps 3 --sampling_steps 50 --diffusion_samples 1 --num_workers 0

Verified

inf2 v8, on a freshly launched inf2.xlarge installed only by setup_inf2.sh: 0 NEFFs compiled at ds 1, 8 and 25; output bit-identical to the build host's; predict wall time 72 s (ds=1), 72 s (ds=8) and 93 s (ds=25) per command, after a first command of about 4 minutes on a new instance.

trn2 v8, restored into empty cache directories on a trn2.3xlarge: 0 NEFFs compiled at ds 1, 8 and 25; seeded output bit-identical to an uncached run; predict wall time 56 / 63 / 73 s.

License

MIT. Boltz-2 is MIT-licensed by its authors.