Field report: your day-0 NVFP4 quant as the one-piece baseline in a 1.11M-token KV pool campaign on RTX 6000 Pro 96G
#11
by AntigravityAI - opened
We used your day-0 NVFP4 quant as the one-piece control group before moving to the SSD-Stream split shell — and it holds up: boots clean at TP1 on the RTX PRO 6000, which matches what others reported. Our campaign ended up pushing the KV pool from 374,208 to 1,111,168 tokens (1M ctx YaRN x4, native MTP, FP8 KV) on a 62GB-RAM box, with your quant serving as the reference point for stability and layout the whole way. Write-up with logs: https://github.com/AntigravityAI/Qwen3.8-Flash-Next-on-RTX-PRO-6000 — thanks for the solid day-0 baseline.