Yirui Zhou

Haoyang Zhang, Yirui Eric Zhou, Yuqi Xue et al.

To break the GPU memory wall for scaling deep learning workloads, a variety of architecture and system techniques have been proposed recently. Their typical approaches include memory extension with flash memory and direct storage access. However, these techniques still suffer from suboptimal performance and introduce complexity to the GPU memory management, making them hard to meet the scalability requirement of deep learning workloads today. In this paper, we present a unified GPU memory and storage architecture named G10 driven by the fact that the tensor behaviors of deep learning workloads are highly predictable. G10 integrates the host memory, GPU memory, and flash memory into a unified memory space, to scale the GPU memory capacity while enabling transparent data migrations. Based on this unified GPU memory and storage architecture, G10 utilizes compiler techniques to characterize the tensor behaviors in deep learning workloads. Therefore, it can schedule data migrations in advance by considering the available bandwidth of flash memory and host memory. The cooperative mechanism between deep learning compilers and the unified memory architecture enables G10 to hide data transfer overheads in a transparent manner. We implement G10 based on an open-source GPU simulator. Our experiments demonstrate that G10 outperforms state-of-the-art GPU memory solutions by up to 1.75$\times$, without code modifications to deep learning workloads. With the smart data migration mechanism, G10 can reach 90.3\% of the performance of the ideal case assuming unlimited GPU memory.

Yuan Meng, Xiangtong Yao, Haihui Ye et al.

Embodied long-horizon manipulation requires robotic systems to process multimodal inputs-such as vision and natural language-and translate them into executable actions. However, existing learning-based approaches often depend on large, task-specific datasets and struggle to generalize to unseen scenarios. Recent methods have explored using large language models (LLMs) as high-level planners that decompose tasks into subtasks using natural language and guide pretrained low-level controllers. Yet, these approaches assume perfect execution from low-level policies, which is unrealistic in real-world environments with noise or suboptimal behaviors. To overcome this, we fully discard the pretrained low-level policy and instead use the LLM to directly generate executable code plans within a closed-loop framework. Our planner employs chain-of-thought (CoT)-guided few-shot learning with incrementally structured examples to produce robust and generalizable task plans. Complementing this, a reporter evaluates outcomes using RGB-D and delivers structured feedback, enabling recovery from misalignment and replanning under partial observability. This design eliminates per-step inference, reduces computational overhead, and limits error accumulation that was observed in previous methods. Our framework achieves state-of-the-art performance on 30+ diverse seen and unseen long-horizon tasks across LoHoRavens, CALVIN, Franka Kitchen, and cluttered real-world settings.