AMD released veRL on ROCm on September 8, 2026, saying it provides production-ready reinforcement learning post-training capabilities on AMD Instinct GPUs. It is the company's first software update in the AI/ML category since its previous ROCm release.
AMD reported training accuracy validated on both MI300 and MI355 GPUs, with case studies showing convergence on GSM8K and MATH benchmarks. That compares with prior accuracy benchmarks on other platforms.
veRL is built on ROCm and targets reinforcement learning post-training workflows, including PPO, GRPO, and DAPO. Availability begins with a turnkey container, initially for developers and researchers.
"veRL ships a turnkey ROCm container verlai/verl:rocm7.14_torch2.12_release_0724 based on docker/rocm/Dockerfile.rocm so you can go from zero to training without hand-assembling a single dependency," said Fuwei Yang, lead researcher. The image bundles the entire runtime stack — ROCm, PyTorch, Triton, vLLM, SGLang, AITER, TransformerEngine and Megatron-core — all pinned to a set of versions verified to work together.
The announcement follows AMD's earlier release of ROCm 7.14. AMD said the release expands its software ecosystem for AI and machine learning.
AMD did not say how veRL will perform on future GPU generations, and raised the question of whether the framework will support other hardware platforms. The company said it will continue to validate and expand the framework's capabilities.
Source: amd
