Search

Reinforcement Learning from Human Feedback (RLHF) enhances the alignment between LLMs and human preference. The workflow of RLHF typically involves several models and tasks in a series of distinct stages. Existing RLHF training systems view each task as the smallest execution unit thus overlooking the opportunities for subtask-level optimizations. Due to the intrinsic nature of RLHF training, i.e., the data skewness in the generation stage, and the pipeline bubbles in the training stage, existing RLHF systems suffer from low GPU utilization in production deployments. RLHFuse breaks the traditional view of RLHF workflow as a composition of individual tasks, splitting each task into finer-grained subtasks, and performing stage fusion to improve GPU utilization. RLHFuse contains two key ideas. First, for generation and inference tasks, RLHFuse splits them into sample-level subtasks, enabling efficient inter-stage fusion to mitigate the original generation bottleneck dominated by long-tailed samples. Second, for training tasks, RLHFuse breaks them into subtasks of micro-batches. By leveraging the intuition that pipeline execution can be essentially complemented by another pipeline, RLHFuse performs intra-stage fusion to concurrently execute these subtasks in the training stage with a fused pipeline schedule, resulting in fewer pipeline bubbles. In addition, RLHFuse incorporates a series of system optimizations tailored for each stage of RLHF, making it efficient and scalable for our internal product usage. We evaluate RLHFuse on various popular LLMs and the results show that RLHFuse increases the training throughput by up to 3.7x, compared to existing state-of-the-art systems.

Multimodal large language models (LLMs) have demonstrated significant potential in a wide range of AI applications. Yet, training multimodal LLMs suffers from low efficiency and scalability, due to the inherent model heterogeneity and data heterogeneity across different modalities. We present DistTrain, an efficient and adaptive framework to reform the training of multimodal large language models on large-scale clusters. The core of DistTrain is the disaggregated training technique that exploits the characteristics of multimodal LLM training to achieve high efficiency and scalability. Specifically, it leverages disaggregated model orchestration and disaggregated data reordering to address model and data heterogeneity respectively. We also tailor system optimization for multimodal LLM training to overlap GPU communication and computation. We evaluate DistTrain across different sizes of multimodal LLMs on a large-scale production cluster with thousands of GPUs. The experimental results show that DistTrain achieves 54.7% Model FLOPs Utilization (MFU) when training a 72B multimodal LLM on 1172 GPUs and outperforms Megatron-LM by up to 2.2$\times$ on throughput. The ablation study shows the main techniques of DistTrain are both effective and lightweight.

Retrieval-Augmented Generation (RAG) has shown significant improvements in various natural language processing tasks by integrating the strengths of large language models (LLMs) and external knowledge databases. However, RAG introduces long sequence generation and leads to high computation and memory costs. We propose RAGCache, a novel multilevel dynamic caching system tailored for RAG. Our analysis benchmarks current RAG systems, pinpointing the performance bottleneck (i.e., long sequence due to knowledge injection) and optimization opportunities (i.e., caching knowledge's intermediate states). Based on these insights, we design RAGCache, which organizes the intermediate states of retrieved knowledge in a knowledge tree and caches them in the GPU and host memory hierarchy. RAGCache proposes a replacement policy that is aware of LLM inference characteristics and RAG retrieval patterns. It also dynamically overlaps the retrieval and inference steps to minimize the end-to-end latency. We implement RAGCache and evaluate it on vLLM, a state-of-the-art LLM inference system and Faiss, a state-of-the-art vector database. The experimental results show that RAGCache reduces the time to first token (TTFT) by up to 4x and improves the throughput by up to 2.1x compared to vLLM integrated with Faiss.

Large Language Models (LLMs) have shown remarkable progress in automated code generation. Yet, LLM-generated code may contain errors in API usage, class, data structure, or missing project-specific information. As much of this project-specific context cannot fit into the prompts of LLMs, we must find ways to allow the model to explore the project-level code context. We present CoCoGen, a new code generation approach that uses compiler feedback to improve the LLM-generated code. CoCoGen first leverages static analysis to identify mismatches between the generated code and the project's context. It then iteratively aligns and fixes the identified errors using information extracted from the code repository. We integrate CoCoGen with two representative LLMs, i.e., GPT-3.5-Turbo and Code Llama (13B), and apply it to Python code generation. Experimental results show that CoCoGen significantly improves the vanilla LLMs by over 80% in generating code dependent on the project context and consistently outperforms the existing retrieval-based code generation baselines.

We initiate a systematic construction of real analytic Lagrangian fibrations from integer matrices. We prove that when the matrix is of full column rank, the perverse filtration associated with the Lagrangian fibration matches the mixed Hodge-theoretic weight filtration of the isolated cluster variety associated with the matrix., Comment: 16 pages. Comments are welcome!

A three-phase isolated AC-DC-DC power supply is widely used in the industrial field due to its attractive features such as high-power density, modularity for easy expansion and electrical isolation. In high-power application scenarios, it can be realized by multiple AC-DC-DC modules with Input-Parallel Output-Parallel (IPOP) mode. However, it has the problems of slow output voltage response and large ripple in some special applications, such as electrophoresis and electroplating. This paper investigates an improved Adaptive Linear Active Disturbance Rejection Control (A-LADRC) with flexible adjustment capability of the bandwidth parameter value for the high-power DC supply to improve the output voltage response speed. To reduce the DC supply ripple, a control strategy is designed for a single module to adaptively adjust the duty cycle compensation according to the output feedback value. When multiple modules are connected in parallel, a Hierarchical Delay Current Sharing Control (HDCSC) strategy for centralized controllers is proposed to make the peaks and valleys of different modules offset each other. Finally, the proposed method is verified by designing a 42V/12000A high-power DC supply, and the results demonstrate that the proposed method is effective in improving the system output voltage response speed and reducing the voltage ripple, which has significant practical engineering application value., Comment: Accepted by IEEE Transactions on Power Electronics

Large language models (LLMs) power a new generation of interactive AI applications exemplified by ChatGPT. The interactive nature of these applications demands low latency for LLM inference. Existing LLM serving systems use run-to-completion processing for inference jobs, which suffers from head-of-line blocking and long latency. We present FastServe, a distributed inference serving system for LLMs. FastServe exploits the autoregressive pattern of LLM inference to enable preemption at the granularity of each output token. FastServe uses preemptive scheduling to minimize latency with a novel skip-join Multi-Level Feedback Queue scheduler. Based on the new semi-information-agnostic setting of LLM inference, the scheduler leverages the input length information to assign an appropriate initial queue for each arrival job to join. The higher priority queues than the joined queue are skipped to reduce demotions. We design an efficient GPU memory management mechanism that proactively offloads and uploads intermediate state between GPU memory and host memory for LLM inference. We build a system prototype of FastServe and experimental results show that compared to the state-of-the-art solution vLLM, FastServe improves the throughput by up to 31.4x and 17.9x under the same average and tail latency requirements, respectively.