Senior Software Engineer, RL Post-Training Frameworks

NVIDIA•Santa Clara, CA

About The Position

Reinforcement learning post-training is driving some of the most significant capability gains in AI today. It is the process that teaches a model to reason through hard problems, follow complex instructions, and act as an autonomous agent. It is also one of the hardest infrastructure challenges in the field. RL requires inference, rollout generation, and training running in a continuous loop. The rollout step is what makes it hard: the model must interact with environments, tools, and other models to produce the signal that drives learning. Coordinating actor, critic, and reward models across heterogeneous hardware at scale pushes the limits of what distributed systems can do. NVIDIA is building an RL Frameworks engineering team to develop the open-source tools and infrastructure that AI researchers and post-training teams depend on. The team spans the full software stack, from collaborating closely with the researchers and labs pushing the frontier, to contributing to RL frameworks like VeRL, Miles, and TorchTitan, to improving the distributed runtimes they depend on, including Ray and Monarch. Whether your strength is working with researchers to understand and address their need optimizing deep learning frameworks, or building distributed infrastructure, we want to hear from you. Come join us to build the systems that enable the next generation of AI.

Requirements

MS or PhD in Computer Science, Computer Engineering, or a related field (or equivalent experience)
5+ years of professional experience in distributed systems, high-performance computing, deep learning infrastructure, or ML systems engineering
Strong proficiency in Python and C/C++
Demonstrated experience building or contributing to large-scale distributed systems or runtime frameworks in production at a frontier AI lab, hyperscaler, or major technology company
Strong verbal and written communication skills and the ability to collaborate across organizational and geographic boundaries
Depth in one or more of the following technical areas: Reinforcement learning for LLM post-training (RLHF, PPO, GRPO, DPO, reward modeling), including how algorithms map to distributed execution and the systems challenges they create (heterogeneous placement, rollouts, environment execution, resharding between training and generation)
PyTorch internals, including distributed training primitives (FSDP, tensor parallelism, pipeline parallelism) and their composition
Kubernetes runtime internals (container lifecycle, pod scheduling, resource quotas, GPU allocation)
End-to-end distributed systems design (service boundaries, data flows, consistency models, failure modes, recovery approaches)

Nice To Haves

Deep expertise in networking (NCCL, NVLink, InfiniBand), advanced multi-dimensional parallelisms (Megatron-LM, FSDP2, TP/DP/PP, MoE), or memory optimizations (quantization-aware training, mixed precision)
Experience integrating high-performance inference engines (vLLM, SGLang, TensorRT-LLM) into RL training loops for GPU-accelerated rollout
Strong background in actor- and task-based distributed programming (Ray, Monarch, or comparable systems)
Familiarity with multi-turn training, multi-agent co-evolution, or VLM post-training
Open-source contributions to RL post-training or distributed training projects (e.g., VeRL, Miles, TorchTitan, OpenRLHF, NeMo-Aligner, DeepSpeed-Chat), including significant work on framework internals where applicable
Kubernetes work beyond routine operations (custom operators, GPU device plugins, or scheduling contributions)
Direct experience operating frontier-scale training (RL post-training at thousands of GPUs and/or large-scale LLM or multimodal pre-training)
Hands-on experience with production distributed failures at scale (stragglers, resource contention, hardware faults)

Responsibilities

Architect and build RL post-training infrastructure that scales efficiently from experimentation on a single GPU to production across thousands of nodes.
Tune RL training-inference-rollout loops on GPUs, CPUs, and LPUs for performance where it matters.
Contribute to and improve the performance and usability of open-source RL frameworks.
Partner with the teams who own open-source RL frameworks.
Ensure fault tolerance, elastic scaling, and fast restarts for long-running distributed training jobs to survive failures, stragglers, and resource contention.
Partner with teams building CPU-driven rollout workloads, including tool-use, code execution, and agentic environments, supplying the systems and framework engineering needed to run them efficiently alongside GPU- or LPU-accelerated generation and GPU-accelerated training.
Advocate for researcher and partner needs with NVIDIA's networking, math library, and compiler teams so the capabilities RL workloads require get prioritized and delivered.
Work with hardware teams to take advantage of next-generation hardware capabilities in post-training workloads.