RBA gives LLaVA-OneVision-2 a 3x lossless training speed-up, merged upstream
liuziwei7 · x · 2026-11-10
The team used RBA to fully automatically optimize LLaVA-OneVision-2's training, achieving a 3x lossless speed-up: one full epoch dropped from 9.64h to 3.19h on 8×Blackwell with matching loss curves. On 4×Hopper 80GB it went from 21.94h to 9.84h (2.2x). MFU jumped from 11.6% to 45.5% on Blackwell and 22.1% to 45.5% on Hopper, with eval loss essentially unchanged. Verified across 2 nodes × 8 GPUs with multi-node launch and load balancing, under both TP 2 / PP 4 presets. Merged into the main repo via PR #205.
More from Infra
- "Biblically accurate" low-loss waveguide crossing looks majestic, performs terribly — jwt0625 · 2026-11-10
- Comet 60B (A12B) gets GGUF quants with MTP head for speculative decoding in llama.cpp — Felladrin · 2026-11-10
- Nvidia lobbied hard against rules on remote access to restricted chips, officials say — rohanpaul_ai · 2026-11-10
- AI daily: Fable 5.2 tops evals, orchestrates 1000 agents for 8 days to build a prototype OS — jiqizhixin · 2026-11-10
- Engineer calls megakernel hype a psyop: CUDA graphs already express a strict superset — mike64_t · 2026-11-10
- Senate advances energy permitting bill 73-25, requiring data centers to pay more grid costs — firstadopter · 2026-11-10