vLLM RL Training Now Supports AMD ROCm
vllm_project · x · 2026-07-14
The vLLM team announces that **vime**, the RL post-training framework in the vLLM ecosystem, now natively supports end-to-end RL post-training on **AMD Instinct MI355X** GPUs. Key highlights: - Using vLLM as the rollout backend, vime inherits the full vLLM rollout stack on ROCm without needing a separate code path. - The AMD team has validated the end-to-end pipeline and upstreamed ROCm-related fixes. - Pre-built containers are provided to reduce the overhead of building from source. - Currently supported features include GRPO training, co-located/async (non-co-located) train-rollout, Megatron-LM training + vLLM rollout backend, and Qwen3 dense and MoE models. Performance-wise, Qwen3-8B on MI355X achieves around **4,100 tokens/gpu/s**, with train-rollout logprob differences remaining low and stable.
More from coding & agent
- Codex turns out 123 screensavers in one playful batch — intellectronica · 2026-07-21
- Grok Build adds `grok doctor`, resumable sessions and remote image paste — mark_k · 2026-07-21
- Autoresearch proposes packaging ML runs as studies with questions, analysis, and code diffs — morgymcg · 2026-07-21
- CHAP defines approvals, handoffs, and audit logs for human-agent workflows — DeliveryTechnical199 · 2026-07-21
- The author says Codex reached 20x and is now debugging spec decoding on a hybrid parallel setup — TheZachMueller · 2026-07-21
- Axcess adds an MCP connector for WCAG accessibility checks that scanners miss — modelcontextprotocol · 2026-07-21