NVFP4 forward, MXFP8 backward with stochastic rounding: how Inkling 2 Small was trained
snowclipsed · x · 2026-11-14
The author calls the recipe a small surprise versus other NVFP4 approaches but an agreeable, great decision. Quoting the blog: Inkling 2 Small uses NVFP4 precision in expert layers for forward passes with a dedicated kernel, MXFP8 for backward passes with re-quantization and stochastic rounding to reduce gradient bias, plus quantization techniques to accelerate GPU communications. It is the largest known model trained end-to-end with a native 4-bit forward pass, configured identically from pretraining through final RL rollouts, eliminating checkpoint calibration.
More from Infra
- Atreides sends letter to Nvidia board outlining minimum governance standards — GavinSBaker · 2026-11-14
- OSINT deep-dive: FAA nationwide grounding caused by long-haul ATC comms, radar and GPS fine — eigenrobot · 2026-11-14
- Airtel chairman: Starlink V3 delivers 1 Tbps per launch, matching all of OneWeb — elonmusk · 2026-11-14
- Dev goes all-in on Cloudflare, citing its excellent MCP as the deciding factor — ritakozlov · 2026-11-14
- Brave Leo adds verifiable encrypted LLM inference via TEEs, calls ZDR clauses a pinky promise — gnukeith · 2026-11-14
- 97% of coding-agent tokens are cache reads, but cache writes eat 31% of the cost — Nitjsefnie · 2026-11-14