AMD给vLLM生态的RL训练框架vime加了ROCm支持，现在你的Qwen3-8B在MI355X上能跑到4100 tokens/gpu/s，训练和rollout的logprob很稳，还直接提供预构建容器，省去编译麻烦。
AMD团队将ROCm支持引入vime，vime是vLLM生态的RL后训练框架，端到端RL后训练现在原生运行在AMD Instinct MI355X GPU上。vime使用vLLM作为rollout后端，在ROCm上继承完整vLLM栈，无需独立代码路径。AMD验证了pipeline并上游了ROCm修复，提供了预构建容器。支持GRPO训练、colocated和非colocated训练/rollout、Megatron-LM训练+vLLM rollout、Qwen3密集和MoE模型。在MI355X上，Qwen3-8B达到约4100 tokens/gpu/s，训练-rollout logprob差异低且稳定。
🎉 Great to see the @AMD team for bringing ROCm sup…
🎉 Great to see the @AMD team for bringing ROCm support to vime, the vLLM ecosystem's RL post-training framework. End-to-end RL post-training now runs natively on AMD Instinct MI355X GPUs.
vime uses vLLM as its rollout backend, so on ROCm it inherits the full vLLM rollout stack with no separate code path. The @AIatAMD team validated the pipeline end-to-end, upstreamed the ROCm-specific fixes, and shipped a prebuilt container so you can skip building from source.
What works today: - GRPO training - Colocated and async (non-colocated) train/rollout - Megatron-LM training + vLLM rollout backends - Qwen3 dense and MoE models
On MI355X, Qwen3-8B sustains ~4,100 tokens/gpu/s, and the train-rollout logprob diff holds low and stable.
🔗 https://t.co/wYTc83APXH