Running π0.5 on an RTX 4070 Ti SUPER: full walkthrough of local embodied AI inference with open-source APXInf
lxfater · x · 2026-10-27
A detailed first-hand walkthrough of locally deploying the π0.5 embodied model on an unofficially supported RTX 4070 Ti SUPER 16GB using the open-source APXInf inference engine, driving a LIBERO-simulated robot arm.
Setup: two machines split the work — an Ubuntu box with the 4070 Ti SUPER runs APXInf inference, a Mac runs the LIBERO simulation and client, connected over LAN. π0.5 acts as the decision-making brain; every 5 sim steps the frame and arm state are sent to the model, which returns the next action chunk. The sim handles execution, collisions, and object positions, producing full pick-and-place sequences (turning knobs, grabbing and placing a moka pot).
Debugging notes:
- After compiling, the Python extension failed with "function not found." A coding agent traced it to APXinf-robo's pinned older version: the Ada build referenced operators compiled only for Blackwell-class GPUs. Fix: adjust build conditions so non-Blackwell builds skip that operator, then recompile and force-reinstall.
- BF16 inference ran out of VRAM; switching to the project's built-in INT8 mode made minimal inference pass.
Bonus: APXInf also ships an Agentic Engineering workflow for porting new models — it specifies the reference model, target hardware, and acceptance criteria, then guides an agent to implement the model, add operators, and verify against reference outputs. The project is open source on GitHub.
Related event: π0.5 Runs on a Consumer RTX 4070 Ti SUPER(3 posts)→
More from coding & agent
- Void v1.3 ships JSX routes, email aliases, and a wave of Cloudflare fixes — cnakazawa · 2026-10-27
- Dev says building game worlds on Spawn with Opus enables endless shooter maps — TAbrodi · 2026-10-27
- RL environments essay: knowing how to operate software isn't knowing how to achieve goals — ShenRaphael · 2026-10-27
- Rolldown ships Inline Common Chunks to cut small common chunk files — irvinebroque · 2026-10-27
- NUS's DynaFlow Learns Free-Form Multi-Agent Orchestration, Beats AFLOW — NationalUniversityofSingapore · 2026-10-27
- ProAS Fixes Uniform Credit Assignment in Critic-Free Agentic RL — LINs-lab · 2026-10-27