Consistency at the core.
Deterministic reduction paths and explicit operator contracts help control numerical divergence between training and rollout.
Explore the approachOpen infrastructure / Reinforcement learning
Train. Infer.
Stay aligned.
Building cross-hardware and multi-model RL post-training infrastructure for kernel-level train–inference consistency.
OUR PARTNERS
01 / OUR FOUNDATION
Every rollout explores
the unknown.
Build on certainty.
We work to eliminate numerical mismatch between training and inference, with bitwise alignment as the goal and explicit operator contracts as the foundation.
More room to scale.
Fused logprob computation, chunking, and shared-prefix attention target the memory bottlenecks in large RL workloads.
Explore the operatorsYour stack. Your hardware.
Custom operator hooks connect to existing engines, with CUDA and ROCm backends plus partial Ascend adaptation. Availability is operator-specific.
Check compatibility02 / START BUILDING
Build your
next breakthrough.
Build from source and choose the operator path that fits your workload. Start with the installation guide for hardware prerequisites.
Read the documentationNative GPU kernels require Python 3.10+, a matching PyTorch build, and a CUDA or ROCm toolchain. Install backend dependencies first.
git clone https://github.com/RL-Align/RL-Kernel.git
cd RL-Kernel
# NVIDIA SM90: H100, H200, GH200
MAX_JOBS=8 \
RL_KERNEL_REQUIRE_EXT=1 \
TORCH_CUDA_ARCH_LIST="9.0+PTX" \
python3 -m pip install --no-build-isolation --no-deps -e .
# Verify the GPU, SM capability, and native extension.
python3 -c "import torch, rl_engine._C as C; \
print('GPU:', torch.cuda.get_device_name(0)); \
print('SM:', torch.cuda.get_device_capability(0)); \
print('Extension:', C.__file__); assert hasattr(C, 'fused_logp')"
git clone https://github.com/RL-Align/RL-Kernel.git
cd RL-Kernel
# AMD gfx942: MI300A, MI300X, MI325X
PYTORCH_ROCM_ARCH=gfx942 python3 setup.py develop
# Verify the ROCm environment and native extension.
python3 scripts/check_rocm_env.py
python3 -c "import torch, rl_engine._C as C; \
print('GPU:', torch.cuda.get_device_name(0)); \
print('HIP:', torch.version.hip); \
print('Extension:', C.__file__); assert hasattr(C, 'fused_logp')"
git clone https://github.com/RL-Align/RL-Kernel.git
cd RL-Kernel
# CPU-only / pure-Python fallback.
python3 -m pip install -e .
03 / THE COMMUNITY
The next frontier.
Built together.
We’re a community of kernel developers, infrastructure engineers, and researchers. Bring a benchmark, a question, or your next pull request.