Topic

Breaking Through the Bottlenecks of LLM Post-Training: AReno’s Unified Training-Inference Architecture and Embodied AI Practice

As post-training tasks such as RLHF and DPO become increasingly complex, many bottlenecks are shifting from algorithms to systems: low GPU utilization, costly weight synchronization, limited GPU memory, and fragmented training and inference pipelines. AReno addresses these challenges through a unified training-inference architecture that brings the two workflows into a single system. This talk will walk through AReno’s core architecture, covering algorithm interfaces, model adaptation, training and inference engines, weight synchronization, fused operators, and memory management. Rather than focusing on concepts, the talk will explore practical engineering trade-offs: which optimizations actually reduce resource consumption and improve performance, and which may introduce stability challenges. The talk will also cover AReno’s applications in embodied AI, including VLA model integration, simulation environment interaction, trajectory collection, and policy post-training. A demo of ARenoflow will show how an Agent can turn natural-language requirements into training configurations and connect the entire workflow from training and diagnostics to deployment. Outline Why LLM Post-Training Is Becoming a Systems Problem Resource inefficiency, communication overhead, and the debugging and iteration costs caused by separate training and inference pipelines. AReno’s Unified Training-Inference Architecture How AReno is structured across algorithm interfaces, model adaptation, and execution engines, and how the architecture manages system complexity. Trade-offs Between Performance, Memory, and Stability Practical experience with weight synchronization, low-bit optimizers, fused operators, bucket-based prefetching, and asynchronous offloading. System Support for Embodied AI How VLA integration, simulation interaction, trajectory collection, and policy post-training are connected within a unified pipeline. Closing the Loop with ARenoflow Using Agents to generate training configurations and automate training orchestration, runtime diagnostics, and model deployment. Key Takeaways Understand the major system-level bottlenecks in LLM post-training, beyond the algorithms themselves. Learn how a unified training-inference architecture can be structured and implemented in practice. Gain practical insights into GPU memory optimization, weight synchronization, fused operators, and other system-level techniques. Understand the key integration paths for VLA models and simulation environments in post-training systems. See how AI-native platforms can reduce repetitive work and streamline the experimentation lifecycle.