DeepSeek-V4 Flash Reinforcement Learning Training Lands on AMD Instinct MI355X GPUs with Miles Framework Support
Original title:News Bringing DeepSeek-V4 Flash RL Training to AMD Instinct MI355X GPUs with Miles DeepSeek-V4 RL is now supported in Miles on AMD Instinct™ MI355X GPUs with ROCm™! RL requires SGLang rollout and Megatron training to implement the same policy closely enough that token probabilities ... AMD & Miles Team
AI summary
Reinforcement learning training for DeepSeek-V4 Flash is now supported on AMD Instinct MI355X GPUs through the Miles framework, running on the ROCm software stack. The 284-billion-parameter MoE model (13 billion activated parameters per token) requires SGLang for rollout generation and Megatron for policy updates, with Miles handling the asynchronous loop and weight synchronization. The team addressed three major challenges: model alignment between SGLang and Megatron, online updates of quantized states, and multi-node parallel stability, ultimately completing end-to-end validation on four eight-GPU nodes: over 100 optimizer steps, the training-rollout log-probability difference remained controllable, online rewards continued to improve, and offline AIME-2024 benchmark scores rose in tandem.
Key Takeaways
- Reinforcement learning training for DeepSeek-V4 Flash is now supported on AMD Instinct MI355X GPUs…
- The 284-billion-parameter MoE model (13 billion activated parameters per token) requires SGLang…
- The team addressed three major challenges: model alignment between SGLang and Megatron,online…
Quick Facts
- Published
- Updated
- Source
- LMSYS Blog
- Category
- Papers & Research
- AI score
- 61
把 DeepSeek-V4 Flash 的 RL 训练完整搬到 AMD MI355X 上,不是画饼,而是给了实测数据和可跑的配置。想用非 NVIDIA 卡做 RL 训练的,可以照着踩坑。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: DeepSeek-V4 Flash Reinforcement Learning Training Lands on AMD Instinct MI355X GPUs with Miles Framework Support. July 10, 2026. https://new.nat6.net/en/news/cmrf7a5kg000wihpr80vvnaxn/