Training SWE agents with Miles on MI355X: taming long-tail trajectories and complex sandbox coordination

In 30 Sekunden

Was passiert ist

Researchers trained a 30B coding model using reinforcement learning on AMD MI355X GPUs, improving task solve rate from 7.3% to 27.5% on SWE-bench.

Warum es zählt

Engineers building agentic AI systems need to understand how to scale RL training for code generation across multiple GPUs and manage sandbox infrastructure efficiently.

Achtung

Long-tail trajectories and CPU sandbox bottlenecks dominated training time, not GPU compute. Infrastructure limits like Docker network caps and connection pools broke before network bandwidth became limiting.