{"slug": "deepseek-v4-flash-rl-on-amd-instinct-mi308x-gpus-with-an-aiter-backed-rl-stack", "title": "DeepSeek-V4 Flash RL on AMD Instinct MI308X GPUs with an AITER-Backed RL Stack", "summary": "A three-node test ran end-to-end DeepSeek-V4 Flash reinforcement learning on 24 AMD Instinct MI308X GPUs using an AITER-backed AMD RL stack, covering long-context FP8 rollout, BF16 MoE actor training, online RDMA weight updates, evaluation and checkpoint recovery. Two MI308X OAM nodes host the BF16 actor for gradient and optimizer stability, while a dedicated third node runs an FP8 rollout engine whose smaller per-parameter footprint frees GPU memory for KV cache and longer generated sequences, with Ray placing and coordinating both GPU groups.", "body_md": "We demonstrate end-to-end DeepSeek-V4 Flash reinforcement learning with an\nAITER-backed AMD RL stack on 24 AMD Instinct\nMI308X GPUs. Our test covers the complete workflow on three MI308X OAM\nnodes—not only inference or an\nisolated training step, but long-context FP8 rollout, BF16 MoE actor training,\nonline RDMA weight updates, evaluation, and checkpoint recovery. Training and\ninference run as disaggregated services: two MI308X nodes host a BF16 actor\nthat preserves the numerical stability required by gradient computation and\noptimizer updates, while a dedicated third MI308X node hosts an FP8 rollout\nengine whose smaller per-parameter footprint leaves more GPU memory available\nfor KV cache and longer generated sequences.\nRay places and coordinates both GPU groups. This\nexperiment demonstrates one tested three-node configuration\nfor running both rollout inference and RL training on MI308X GPUs.", "url": "https://wpnews.pro/news/deepseek-v4-flash-rl-on-amd-instinct-mi308x-gpus-with-an-aiter-backed-rl-stack", "canonical_source": "https://rocm.blogs.amd.com/ecosystems-and-partners/deepseek-v4-mi308-aiter-backed-rl/README.html", "published_at": "2026-10-28 00:00:00+00:00", "updated_at": "2026-10-28 16:45:14.152027+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "ai-infrastructure", "ai-chips", "large-language-models"], "entities": ["DeepSeek", "DeepSeek-V4 Flash", "AMD", "AMD Instinct MI308X", "AITER", "Ray"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/deepseek-v4-flash-rl-on-amd-instinct-mi308x-gpus-with-an-aiter-backed-rl-stack", "markdown": "https://wpnews.pro/news/deepseek-v4-flash-rl-on-amd-instinct-mi308x-gpus-with-an-aiter-backed-rl-stack.md", "text": "https://wpnews.pro/news/deepseek-v4-flash-rl-on-amd-instinct-mi308x-gpus-with-an-aiter-backed-rl-stack.txt", "jsonld": "https://wpnews.pro/news/deepseek-v4-flash-rl-on-amd-instinct-mi308x-gpus-with-an-aiter-backed-rl-stack.jsonld"}}