verl runs DeepSeek-V4-Flash GRPO on 32 AMD Instinct MI355X GPUs: training aligned with FP8 rollout quantization, and BF16, BF16+QAT and FP8 recipes each completed 300 stable steps.
VeRL-Omni v0.3.0 expands async RL, makes the V1 trainer the default, and adds reproducible recipes for audio-video generation and multimodal understanding.
Use RL Insight to locate Trainer and Rollouter bubbles, properly balance training and inference resources, maximize single-card throughput, and enable more efficient training.
MXFP8 on both sides of RL in verl on B200: a sampler on the learner's weight grid cuts train-rollout KL by 16%, training pays off above about 4.5k tokens per micro-batch, and seven failures show what a sync must preserve.
Uni-Agent is an open framework for scalable long-horizon agent RL. We share our insights on harness adaptation, RL training, and practical recipes for agentic RL.
verl-vla is an open-source framework built on verl that unifies human-in-the-loop data collection, fine-tuning, evaluation, and reinforcement learning for vision-language-action policies across cloud and edge resources.
A complete GRPO workflow using verl, Qwen3-30B-A3B, SGLang, and PyTorch FSDP ran for 100 consecutive steps on four AMD Instinct MI455X GPUs, establishing an end-to-end RL baseline for CDNA 5.
A release focused on higher-throughput diffusion rollout, reusable omni adapters, and broader recipe coverage.
RL post-training on AMD Instinct GPUs is here: a turnkey ROCm container, AITER-accelerated vLLM and SGLang rollout, and accuracy validated on both MI300 and MI350 series.
