ThunderAgent (ICML 2026): Boosts Multi-Turn Agent RL Rollout Throughput by 1.94x
simran_s_arora · x · 2026-07-30
ThunderAgent, accepted as a Spotlight paper at ICML 2026, proposes an optimization for the inference bottleneck in multi-turn agent reinforcement learning. The work has been integrated into the Dynamo rollout backend of verl-recipe.
Core Problem & Solution
Multi-turn agent inference often wastes GPU compute due to KV cache thrashing. ThunderAgent addresses this at the scheduler level by introducing program-aware routing.
Performance Metrics
- Achieves a 2.5x increase in single-node inference throughput
- Reduces P50 latency by approximately 10x under high concurrency
- Delivers 1.94x rollout throughput and 1.39x full-step throughput in a matched synchronous GRPO run
This system provides significant value for improving the infrastructure efficiency of large-scale agent training and deployment.
Related event: Together AI Unveils ThunderAgent, Doubling Agent Inference Throughput(7 posts)→
More from Infra
- NVIDIA launches RTX Spark for local AI on Windows PCs — ostrisai · 2026-10-16
- SkyRL ships Day-0 RL support for Beam with +62% LoRA decode throughput in vLLM — vllm_project · 2026-10-16
- DFlash2 delivers up to 3.03x decode-only speedup for Beam across five benchmarks — vllm_project · 2026-10-16
- vLLM's Beam serving kernels cut GB200 forward time from 8.15 to 6.47 ms at TP8 batch 128 — vllm_project · 2026-10-16
- Reflection's 501B MoE Beam hits 463 tok/s on 8x GB200 with NVFP4 at day 0 — zhyncs42 · 2026-10-16
- Gemma 4 31B QAT on one 24GB card: 46 tok/s plain, ~80 with draft model on RTX 4090 — Subject_Poet_3861 · 2026-10-16