Breaking Heterogeneity Barriers: Unified Cloud-to-Robot AI System SW Stack for...
summary
A PyTorch-native stack called FlagOS stitches cloud VLMs to edge VLAs using a Triton-based operator library FlagGEM (200+ Triton kernels) plus compiler optimizations in FlagTree and orchestration in FlagScale. Claims: >2x training throughput vs common data loaders, 10–20% better parallel performance vs alternatives, Triton kernels that match or beat CUDA on Media hardware. Demonstrated running 3.3B VLA models aimed at ~100 TOPS and 20–30 GB edge memory with cloud‑edge latency targets approaching human 100 ms limits.
AI-written summary · about
chapters (9)
- 0:00Embodied AI introduction
- 3:40Brain and VLA models overview
- 8:45Cross-chip architecture support
- 12:00FL OS for AI chipsets
- 16:42FRA compiler and Frescale
- 19:10Training and inference workflows
- 24:11Cloud-edge collaboration
- 26:01New hardware support
- 27:03Q&A
related talks
14mLightning Talk: Full-Stack PyTorch Robotics VLA: From Data To...
29mTeam Teardown: MLOps in the Cloud for IoT
16mLightning Talk: Efficient Inference at the Edge: Performance You Need at the Lowest...
23mBringing ExecuTorch To the Next Frontiers of Edge AI
40mCloud-Native Automotive Edge: AGL for Software-Defined Vehicles
28mML on the Edge for Industry 4.0 with Arm Ethos N-78 Neural Processing Unit
50mAI acceleration at the edge
04mMultiple embedded and Cloud AI solutions at ST