Spaces:
Sleeping
Requesting GPU Grant for Academic Research - Harmonic SSM
I am Petr Nyoma, an independent researcher at Harmonic Labs (Nizhny Novgorod). I am building two complementary open sequence models:
Harmonic — a hierarchical state-space model where each level receives the prediction error of the level below. On enwiki8, the quality advantage over a matched Transformer grows with context: +1.4% at 1K tokens to +11.4% at 32K. The paper is on arXiv, endorsed by Prof. Martin Jaggi (EPFL) and Dr. Sebastian Ruder.
What we need: compute to run a parameter ladder (30M → 200M) for both models and a joules-per-token benchmark vs matched Transformers at long context. We have no institutional funding. All output is released openly (arXiv + code).
We applied to Google TRC. Any GPU grant would directly enable the next result.