High-performance RL post-training infra. Achieves bitwise train-inference consistency across heterogeneous engines & extreme memory efficiency for GRPO, PPO.
Singapore
Joined August 2026
- Tweets14
- Following10
- Followers57
- Likes12
RL-Align retweeted
大模型 RL 后训练中，Rollout 与 Training 往往由不同引擎执行。即使模型、权重和 token 完全相同，仍可能因浮点归约顺序、并行分片、通信合并树及精度转换边界不同，产生 LogP 偏差，进而影响 PPO/GRPO 的 importance ratio、clipping 与训练稳定性。
北京时间 10月20日（周二）晚8点，#青稞社区 邀请到 RL-Kernel 核心 Member、西安交通大学硕士研究生李佳杰，以 @RLKernel RL-Kernel v0.1.0 为例，介绍训推差异的定位方法、关键算子与数值契约的对齐方式
以及固定 replay、逐层比对和端到端 200-step 验证流程，并展示 Qwen3-8B 在 CUDA 与 ROCm 环境下的验证结果。最后讨论 PD 跨机分离与异步 RL 场景的一致性边界和后续接入计划。
RL-Kernel v0.1.0 is now available. 🎉
Bitwise train–inference consistency across parallel configurations and heterogeneous execution engines for RL post-training.
With Qwen3-8B on a full vime pipeline, RL-Kernel recorded zero train–rollout LogP mismatches across all 200 training steps.
1,129 commits from 29 contributors.
Highlights:
🎯 Strict numerical alignment across rollout and training engines
⚡ 68.7% higher rollout throughput and 8.2% lower end-to-end step time on NVIDIA H100
🔶 AMD Instinct MI300X: zero mismatches across 9,400,614 active-token LogP elements from 1,600 samples
🔁 End-to-end integration with vime, vLLM rollout, and Megatron-LM training
🧩 Deterministic Attention, dense FFN, LogP, reduction, and collective operators
🌐 NVIDIA CUDA and AMD ROCm support, with initial Ascend integration and MUSA adaptation underway
GitHub 👇
github.com/RL-Align/RL-Kerne…
RL-Align retweeted
RL-Kernel v0.1.0 is now available. 🎉
Bitwise train–inference consistency across parallel configurations and heterogeneous execution engines for RL post-training.
With Qwen3-8B on a full vime pipeline, RL-Kernel recorded zero train–rollout LogP mismatches across all 200 training steps.
1,129 commits from 29 contributors.
Highlights:
🎯 Strict numerical alignment across rollout and training engines
⚡ 68.7% higher rollout throughput and 8.2% lower end-to-end step time on NVIDIA H100
🔶 AMD Instinct MI300X: zero mismatches across 9,400,614 active-token LogP elements from 1,600 samples
🔁 End-to-end integration with vime, vLLM rollout, and Megatron-LM training
🧩 Deterministic Attention, dense FFN, LogP, reduction, and collective operators
🌐 NVIDIA CUDA and AMD ROCm support, with initial Ascend integration and MUSA adaptation underway
GitHub 👇
github.com/RL-Align/RL-Kerne…