Reinforcement learning (RL) has shifted from a niche experiment to a critical stage in developing modern AI foundation models. While pretraining builds the initial knowledge base, post-training techniques like RL are what actually teach models how to reason through complex problems, use digital tools, and maintain coherent multi-turn conversations. To facilitate this at scale, the open-source Miles framework has officially launched support for AMD’s ROCm software stack. This integration allows researchers to run intensive RL workflows natively on AMD Instinct GPUs, specifically the high-performance MI300 and MI350 series.

What makes RL training unique compared to standard model training is the rollout phase, where the model generates thousands of trial responses to see which ones work best. In many pipelines, this phase consumes up to 90% of total compute time. Because this process is extremely memory-intensive, AMD’s hardware—known for its massive high-bandwidth memory (HBM) capacity—is particularly well-suited for these tasks. By using a decoupled architecture, Miles separates the data generation (rollouts) from the actual weight updates (training), ensuring that hardware resources are used as efficiently as possible across large clusters.

Early benchmarks on agentic tasks, such as solving math problems using a Python interpreter, show promising results. As training progresses, the models demonstrate an increased ability to handle multiple turns or steps of reasoning before arriving at a solution. This rollout-heavy optimization is essential for the next generation of AI agents that don't just guess an answer but actively verify and correct their work mid-trajectory. This move by AMD and the Miles team strengthens the open-source ecosystem, providing a robust alternative to proprietary hardware stacks for large-scale AI development.

Reinforcement learning (RL) has shifted from a niche experiment to a critical stage in developing modern AI foundation models. While pretraining builds the initial knowledge base, post-training techniques like RL are what actually teach models how to reason through complex problems, use digital tools, and maintain coherent multi-turn conversations. To facilitate this at scale, the open-source Miles framework has officially launched support for AMD’s ROCm software stack. This integration allows researchers to run intensive RL workflows natively on AMD Instinct GPUs, specifically the high-performance MI300 and MI350 series.

What makes RL training unique compared to standard model training is the rollout phase, where the model generates thousands of trial responses to see which ones work best. In many pipelines, this phase consumes up to 90% of total compute time. Because this process is extremely memory-intensive, AMD’s hardware—known for its massive high-bandwidth memory (HBM) capacity—is particularly well-suited for these tasks. By using a decoupled architecture, Miles separates the data generation (rollouts) from the actual weight updates (training), ensuring that hardware resources are used as efficiently as possible across large clusters.

Early benchmarks on agentic tasks, such as solving math problems using a Python interpreter, show promising results. As training progresses, the models demonstrate an increased ability to handle multiple turns or steps of reasoning before arriving at a solution. This rollout-heavy optimization is essential for the next generation of AI agents that don't just guess an answer but actively verify and correct their work mid-trajectory. This move by AMD and the Miles team strengthens the open-source ecosystem, providing a robust alternative to proprietary hardware stacks for large-scale AI development.