AMD released Verl 0.9.0 on ROCm 10.0, saying it enables next-generation reinforcement learning on AMD Instinct GPUs with improved performance and platform integration. It is the company's first software update in the AI/ML category since the previous blog post in August 2026.
AMD reported the key figure of a production-ready platform, measured on the ROCm 10.0 framework. That compares with the earlier version, which was validated on 8× MI355X GPUs.
Verl 0.9.0 is built on the ROCm 10.0 platform and targets reinforcement learning applications with support for async policy training and improved throughput. Availability begins with the release/0.9.0.amd0 branch, initially for developers and researchers.
"PlatformROCm is a first-class platform, not a CUDA lookalike," said Tiffany Mintz, AI/ML lead at AMD. The platform layer ensures compatibility with Ray, vLLM, and SGLang through updated defaults and environment variables.
The announcement follows the release of the ROCm 10.0 framework. AMD framed the significance as a step toward full platform integration for production reinforcement learning.
AMD did not say the platform is compatible with all CUDA-based models, and it raised the open question of how well the AITER kernels perform on different GPU architectures. The source says the platform is covered by an AMD Instinct regression workflow.
Source: amd
