Developer ports Stanford's Memorilla memory framework to train Gemma4 locally on Strix Halo
DataPhreak · reddit · 2026-11-11
Reddit user DataPhreak spent days porting Memorilla (arXiv:2610.21375), a memory framework from Stanford, MIT, UPenn and the Vector Institute, from CUDA to AMD's Strix Halo to train Gemma4 26b a4b. Memorilla compresses huge document collections or chat histories into just 16 dynamic memory tokens — treating long-term memory like a new modality, similar to how vision-language models embed images, instead of massive text copy-pasting.
The port was nontrivial: the original code targeted CUDA and Gemma4's unique architecture required heavy dimension adaptation. By leveraging Stanford's published weights, only 10% of the original dataset is needed for training; the run will take a couple of weeks. Code is open-sourced on GitHub.
More from Infra
- Amazon Prime 360 uses AI to rebuild NBA plays in 3D for live, any-angle replays — import_jmr · 2026-11-11
- Altman says OpenAI aims to replace Nvidia in the medium to long term — zephyr_z9 · 2026-11-11
- Chinese University Bought Nvidia H200 Chips Despite US Blacklist — pstAsiatech · 2026-11-11
- Baidu's ERNIE Evolve tops public bests on NanoGPT Speedrun, brings RSI to real inference optimization — xiaohu · 2026-11-11
- Baidu says RSI auto-optimized ERNIE inference, 6.24x decode speedup on Kunxin P800 — xiaohu · 2026-11-11
- Nvidia's supply commitments jump from $119B to $279B in one quarter, locking up 2027-28 memory — tengyanAI · 2026-11-11