Day-0 RL for Mistral Large 4: LoRA Training on 8 GB200 Nodes Lifts AIME 2024 From 64% to 75%
vllm_project · x · 2026-10-30
The vLLM project announced Day-0 RL support for Mistral's Mistral Large 4, built with NovaSky's SkyRL. On 8 NVIDIA GB200 nodes, rank-32 LoRA RL on the NVFP4 checkpoint raised AIME 2024 avg@8 from 64% to 75% in 75 training steps.
Key recipe details:
- vLLM generates rollouts; FSDP2 or Megatron handles training
- NVFP4 quantization-aware training with FP8 KV-cache simulation
- Routing replay (R3) and token-level truncated importance sampling
- The FSDP2 run kept the mean absolute trainer–vLLM logprob gap at 0.02–0.026; Megatron reached similar accuracy with a comparable gap
More from Infra
- 797GB of CPU-side n-gram lookups match 57% more compute on 1B LM training — PangWeiKoh · 2026-10-30
- Redditor jokes about finding racks of idle GPUs at a local data center — Beneficial-Ad-8127 · 2026-10-30
- Intel shares plunge 20% as AI-driven memory shortage hurts its PC chip business — pstAsiatech · 2026-10-30
- Wafer hits 2 trillion daily tokens on OpenRouter, doubling from 1T in just 4 days — ycombinator · 2026-10-30
- Databricks VP: Claude Code is just the start, AI software factories need a new agent-native infrastructure — FinanceYF5 · 2026-10-30
- IBM/Cornell's LPT shares KV-cache across looped transformer passes, cutting memory 79% while improving quality — burny_tech · 2026-10-30