# vime x RL-Kernel x AMD: Bitwise Train-Rollout Consistency on ROCm

DevFeed: [vime x RL-Kernel x AMD: Bitwise Train-Rollout Consistency on ROCm](<https://devfeed.tech/articles/vime-rl-kernel-amd-bitwise-train-rollout-consistency-on-rocm-79486.md>)

Original publisher: [Read original article](<https://vllm.ai/blog/2026-09-14-rl-kernel-v0-1-0-train-rollout-bitwise-consistency>)

Author: RL-Kernel Team, vime Team, and AMD Team

Published: 2026-09-14T00:00:00Z

Content type: article

Language: en

Sources: [vLLM Blog](<https://devfeed.tech/sources/vllm-blog.md>)

Topics: [rocm](<https://devfeed.tech/topics/rocm.md>), [grpo](<https://devfeed.tech/topics/grpo.md>), [vllm](<https://devfeed.tech/topics/vllm.md>), [Kernel](<https://devfeed.tech/topics/kernel.md>), [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>), [Linux Kernel](<https://devfeed.tech/topics/linux-kernel.md>)

Tags: [ai-engineering](<https://devfeed.tech/tags/ai-engineering.md>), [amd-gpu](<https://devfeed.tech/tags/amd-gpu.md>), [attention](<https://devfeed.tech/tags/attention.md>), [ecosystem](<https://devfeed.tech/tags/ecosystem.md>), [hardware](<https://devfeed.tech/tags/hardware.md>), [instinct](<https://devfeed.tech/tags/instinct.md>), [kernel](<https://devfeed.tech/tags/kernel.md>), [llm-inference](<https://devfeed.tech/tags/llm-inference.md>), [model-serving](<https://devfeed.tech/tags/model-serving.md>), [open-source-ai](<https://devfeed.tech/tags/open-source-ai.md>), [performance](<https://devfeed.tech/tags/performance.md>), [post-training](<https://devfeed.tech/tags/post-training.md>), [reinforcement-learning](<https://devfeed.tech/tags/reinforcement-learning.md>), [rl](<https://devfeed.tech/tags/rl.md>), [rocm](<https://devfeed.tech/tags/rocm.md>), [token](<https://devfeed.tech/tags/token.md>), [train](<https://devfeed.tech/tags/train.md>), [training](<https://devfeed.tech/tags/training.md>), [vllm](<https://devfeed.tech/tags/vllm.md>), [vllm-blog](<https://devfeed.tech/tags/vllm-blog.md>)

## AI overview

vime coordinates Megatron training, vLLM rollout and weight synchronization, while RL-Kernel aligns numerical execution so selected-token logprobs match bit for bit on AMD Instinct MI300X. The article reports zero mismatches and zero maximum absolute difference across a 200-step Qwen3-8B GRPO run, and describes ROCm-specific kernel, attention, communication and execution changes used to achieve that result.

## Source excerpt

vime and RL-Kernel align selected-token logprobs bit for bit across Megatron training and vLLM rollout on AMD Instinct MI300X, with zero mismatches across 200 GRPO steps.