BioEmu2 fork speeds up inference with context projection reuse, torch.compile, and CUDA Graphs
FrankNoeBerlin · x · 2026-11-14
Developer YumizSui released an accelerated fork of Microsoft's BioEmu2 protein conformation model, adding an opt-in --accelerate flag for equilibrium sampling, structure prediction, and time propagation.
The speedup combines three techniques:
- Context projection reuse: static context projections are precomputed instead of recomputed each step
- torch.compile: the denoiser is compiled
- CUDA Graphs: cuts kernel launch overhead
The fork is MIT-licensed on GitHub and targets users who want faster BioEmu2 inference out of the box.
More from Infra
- Analyst warns of hyperscaler monopoly: NVIDIA loses pricing power, startup funding dries up — BenBajarin · 2026-11-15
- Anthropic CEO Dario Amodei explains why the company buys compute conservatively — _AustinCalvert_ · 2026-11-15
- Anthropic IPO deck chart sparks debate: Mythos could be retrained with a fraction of the compute — Witty_County5128 · 2026-11-15
- Google open-sources AX, a Kubernetes-style orchestrator for running fleets of autonomous agents — rakyll · 2026-11-15
- RustingStrata runs Flint 40B on a 12 GB GPU at 64 tok/s, 33% faster than llama.cpp — Serious_Scarcity7818 · 2026-11-14
- Citizen Distributed LLM Training PoC: Port 443 Only, Safetensors, MultiKrum — OrganizationTop9026 · 2026-11-14