ResearchBenchmarksRL Training on LongReasonFollow80Peak Memory (GB)DualKV78.9685.9893100.02May 14, 2026Evaluation ResultsMethodMethodLinksPeak Memory (GB)Policy Update Time (s)Old Log Prob Time (s)Reference Log Prob Time (s)Rollout Time (s)Step Time (s)Validation AccuracyMFU (%)Reward (Step 25)DualKVmb/GPU=4mb/GPU=42026.058031363—231606—59.385.3FA2mb/GPU=8, SP=4mb/GPU=8, SP=42026.05924,1154475072145,28480.2——DualKVmb/GPU=8mb/GPU=82026.059324263—237542—77.486.8DualKVmb/GPU=8, SP=1mb/GPU=8, SP=12026.051031,0781291422151,56477.4——FA2mb/GPU=4mb/GPU=42026.05106597153—236987—30.984.8FA3mb/GPU=4mb/GPU=42026.05106543142—234919—3485.9