Rob1234567/axiom-pilot-l3full-step300
Updated
Axiom
Axiom JAX/MXFP4 pilot — now with the MFU-55 research card and its trace dataset: 452 → 1,203.7 tok/s on one DGX Spark in 8 days.
Note Pretrain pilot: l3-full endurance leg, step 300/300 — base for full SFT
The Kernel War for MFU
⚡28 agent cycles, one JAX LLM, one DGX Spark
Note Research card: 28 agent-driven optimisation cycles took a 1.01B JAX MoE from 452 to 1,203.7 tok/s on one DGX Spark. Includes the seven falsified diagnoses.
Rob1234567/axiom-mfu55-traces
Updated • 59Note Trace-derived aggregates behind the MFU-55 card: per-cycle measurements, gates, phase profiles, corrections, two-layer agent traces (12–19 Oct 2026).