Modal Shows Mixed-Precision Samplers Across Six Accelerator Types Cut RL Cost ~15%
lmoroney · x · 2026-10-15
Modal's Nan Jiang and team trained Qwen3.6-35B-A3B on agentic coding with fully asynchronous RL, running samplers on six accelerator types (B200, B300, H100, H200, A100, RTX PRO 6000) at mixed precisions including NVFP4, FP8 and BF16, while a BF16 trainer ran on B200s. Off-policy corrections (score centering, masked importance sampling, top-p with mask replay) kept the mismatch bounded. On SWE-bench Pro V2 the mixed setup tracked the homogeneous B200 run with 15% lower sampler cost per update — significant since sampling already consumes 80–85% of GPUs in large RL runs.
More from Infra
- 1GW of Compute Matches the Gross Profit of 1M Cybercabs — SpaceX to Add 3-5GW Next Year — JOBhakdi · 2026-10-15
- Cloudflare data: Yandex outage didn't kill malicious bots, they just moved to other Russian hosts — Cloudflare Blog · 2026-10-15
- AI cost falling 54x faster than electricity ever did, five years in — LinusEkenstam · 2026-10-15
- Inference has no yardstick: 4.5-quadrillion token run rate, but no proof of what buyers get — AccBalanced · 2026-10-15
- Booming business, falling stocks: scarcity caps ASML and Applied Materials upside — TiernanRayTech · 2026-10-15
- Ben Bajarin: semicap theme is accelerating the cadence of capacity builds — BenBajarin · 2026-10-15