RL Training and Inference Engineer for Multimodal Generation

Skills

About the Role

You will build the reinforcement learning training and inference infrastructure for video generation models. You will optimize rollout and inference performance, improve distributed training, coordinate resources across training and reward models, and establish validation, monitoring, and recovery mechanisms.

Requirements

• Proficiency in Python and familiarity with deep learning frameworks such as PyTorch
• Practical expertise in distributed training, model inference optimization, or RL systems
• Ability to analyze performance bottlenecks and implement optimizations independently
• Knowledge of GPU computing and memory management
• Understanding of mixed precision, gradient accumulation, activation recomputation, data parallelism, and model parallelism
• Experience with multi-node, multi-GPU training or inference
• Understanding of reinforcement learning training workflows, rollout, reward calculation, policy updates, and weight synchronization
• Knowledge of Diffusion, Flow Matching, or DiT generative models
• Understanding of long-sequence computation, temporal modeling, and multi-step sampling for video generation
• Experience with image or video generation model training, inference acceleration, RL post-training, or preference optimization
• Experience developing CUDA or Triton operators, Attention optimization, sequence parallelism, or communication-computation overlap optimization
• Experience building large-scale RL systems, asynchronous rollout, training-inference resource coordination, reward-model services, or fault recovery
• Experience with model distillation, few-step sampling, quantization, sparse computation, or core contributions to relevant open-source projects

Responsibilities

• Build reinforcement learning training and inference infrastructure for video generation models
• Connect sample generation, reward calculation, policy updates, and evaluation workflows
• Optimize video-generation inference and rollout performance
• Improve distributed RL training frameworks for long videos, high resolution, and multi-sample workloads
• Optimize scheduling, resource allocation, weight synchronization, data transfer, and pipeline parallelism
• Collaborate with algorithm specialists to implement reinforcement learning, preference optimization, and model acceleration solutions
• Establish performance analysis, numerical correctness validation, monitoring, and fault-recovery mechanisms