📜 The Story of FastFlowLM (FLM)
August 11, 2026 · View on GitHub
From a 2025 university research project to the NPU-first runtime for AMD Ryzen™ AI — now part of AMD.
🎓 2025 — A University Project
Modern AMD Ryzen™ AI laptops all shipped with a capable, power-efficient NPU. Almost nobody could run an LLM on it with good performance and long context length.
FastFlowLM started as an NSF-funded university research project aimed squarely at that gap. The bet: closing it required co-designing kernels around the NPU's dataflow architecture from the ground up — not porting an existing runtime onto it.
What began as federally funded lab work became FastFlowLM, Inc. in 2025, and the runtime was released publicly as open source that June.
Founders
- Tao Wei — Professor of ECE, Clemson University; directs the NEXT Lab (domain-specific accelerators, reconfigurable computing, applied ML). Leads FLM kernel strategy.
- Ken Qing Yang — Distinguished Engineering Professor, University of Rhode Island. 30+ years in computer architecture; serial entrepreneur behind four deep-tech startups, including VeloBit (acquired by Western Digital) and DapuStor.
- Zhenyu (Alfred) Xu — Research Assistant Professor, Clemson University. Accelerator design and on-device AI inference, with hardware–software co-optimization across FPGA, CGRA, and AI accelerators.
🚀 2025–2026 — Building in the Open
First public commit landed June 16, 2025; v0.1.0 nine days later. The design principle was borrowed from tools developers already liked — "Think Ollama, but deeply optimized for NPUs." Install in seconds, flm run <model>, done.
FLM was MIT-licensed and open source from the start, built on IRON, the open-source NPU compiler technology developed and released by AMD's Research and Advanced Development (RAD) Group. That open foundation, plus close alignment with Lemonade (AMD's open-source inference initiative), drove adoption across developers and ISVs.
| Date | Milestone |
|---|---|
| Jun 16, 2025 | First public commit |
| Jun 25, 2025 | v0.1.0 — first tagged release |
| Oct 1, 2025 | Integrated into AMD's Lemonade Server 🍋 |
| Mar 11, 2026 | Linux support launches 🐧 |
| Jul 17, 2026 | FastFlowLM joins AMD 🎉 |
| Aug 7, 2026 | Repo moves to the ROCm organization |
| Aug 10, 2026 | v1.0.0 🎊 |
Over that span: 60+ releases, 1,500+ commits, ~38 contributors. The runtime grew past text to Vision, Audio, Embedding, and MoE models with context up to 256k tokens — in a 17 MB package.
🤝 July 17, 2026 — Joining AMD
The FastFlowLM team joined AMD, joining the AMD Artificial Intelligence Group to accelerate the client and workstation AI software stack and Day-0 enablement of new models. AMD called it:
"another key step in our strategy to advance AI performance and efficiency across the stack."
And on the open ecosystem:
"We remain committed to investing in this open ecosystem, and we're thrilled to build the future of on-device AI together."
Joining AMD did not close the project. FLM stays open source under the MIT license, and its NPU kernels remain free for any use, including commercial use.
🏠 August 2026 — Home in ROCm
On August 7, 2026 the repo moved from FastFlowLM/FastFlowLM to ROCm/FastFlowLM, alongside AMD's open-source ROCm organization. v1.0.0 shipped three days later.
The mission is unchanged: turn idle NPU silicon into instant, private, power-efficient AI.