Kaiqi Chen

Kaiqi Chen, Eugene Lim, Kelvin Lin et al.

Imitation learning empowers artificial agents to mimic behavior by learning from demonstrations. Recently, diffusion models, which have the ability to model high-dimensional and multimodal distributions, have shown impressive performance on imitation learning tasks. These models learn to shape a policy by diffusing actions (or states) from standard Gaussian noise. However, the target policy to be learned is often significantly different from Gaussian and this mismatch can result in poor performance when using a small number of diffusion steps (to improve inference speed) and under limited data. The key idea in this work is that initiating from a more informative source than Gaussian enables diffusion methods to mitigate the above limitations. We contribute both theoretical results, a new method, and empirical findings that show the benefits of using an informative source policy. Our method, which we call BRIDGER, leverages the stochastic interpolants framework to bridge arbitrary policies, thus enabling a flexible approach towards imitation learning. It generalizes prior work in that standard Gaussians can still be applied, but other source policies can be used if available. In experiments on challenging simulation benchmarks and on real robots, BRIDGER outperforms state-of-the-art diffusion policies. We provide further analysis on design considerations when applying BRIDGER. Code for BRIDGER is available at https://github.com/clear-nus/bridger.

Jinwei Yao, Kaiqi Chen, Kexun Zhang et al.

Large language models (LLMs) are increasingly employed for complex tasks that process multiple generation calls in a tree structure with shared prefixes of tokens, including few-shot prompting, multi-step reasoning, speculative decoding, etc. However, existing inference systems for tree-based applications are inefficient due to improper partitioning of queries and KV cache during attention calculation. This leads to two main issues: (1) a lack of memory access (IO) reuse for KV cache of shared prefixes, and (2) poor load balancing.As a result, there is redundant KV cache IO between GPU global memory and shared memory, along with low GPU utilization. To address these challenges, we propose DeFT(Decoding with Flash Tree-Attention), a hardware-efficient attention algorithm with prefix-aware and load-balanced KV cache partitions. DeFT reduces the number of read/write operations of KV cache during attention calculation through KV-Guided Grouping, a method that avoids repeatedly loading KV cache of shared prefixes in attention computation. Additionally, we propose Flattened Tree KV Splitting, a mechanism that ensures even distribution of the KV cache across partitions with little computation redundancy, enhancing GPU utilization during attention computations. By reducing 73-99% KV cache IO and nearly 100% IO for partial results during attention calculation, DeFT achieves up to 2.23/3.59x speedup in the end-to-end/attention latency across three practical tree-based workloads compared to state-of-the-art attention algorithms. Our code is available at https://github.com/LINs-lab/DeFT.