Search for: All records

High-demand LLM inference services (e.g., ChatGPT and BARD) support a wide range of requests from short chat conversations to long document reading. To ensure that all client requests are processed fairly, most major LLM inference services have request rate limits, to ensure that no client can dominate the request queue. However, this rudimentary notion of fairness also results in under-utilization of the resources and poor client experience when there is spare capacity. While there is a rich literature on fair scheduling, serving LLMs presents new challenges due to their unpredictable request lengths and their unique batching characteristics on parallel accelerators. This paper introduces the definition of LLM serving fairness based on a cost function that accounts for the number of input and output tokens processed. To achieve fairness in serving, we propose a novel scheduling algorithm, the Virtual Token Counter (VTC), a fair scheduler based on the continuous batching mechanism. We prove a 2× tight upper bound on the service difference between two backlogged clients, adhering to the requirement of work-conserving. Through extensive experiments, we demonstrate the superior performance of VTC in ensuring fairness, especially in contrast to other baseline methods, which exhibit shortcomings under various conditions. The reproducible code is available at https://github.com/Ying1123/VTC-artifact. 

The "pretrain-then-finetune" paradigm is commonly adopted in the deployment of large language models. Low-Rank Adaptation (LoRA), a parameter-efficient fine-tuning method, is often employed to adapt a base model to a multitude of tasks, resulting in a substantial collection of LoRA adapters derived from one base model. We observe that this paradigm presents significant opportunities for batched inference during serving. To capitalize on these opportunities, we present S-LoRA, a system designed for the scalable serving of many LoRA adapters. S-LoRA stores all adapters in the main memory and fetches the adapters used by the currently running queries to the GPU memory. To efficiently use the GPU memory and reduce fragmentation, S-LoRA proposes Unified Paging. Unified Paging uses a unified memory pool to manage dynamic adapter weights with different ranks and KV cache tensors with varying sequence lengths. Additionally, S-LoRA employs a novel tensor parallelism strategy and highly optimized custom CUDA kernels for heterogeneous batching of LoRA computation. Collectively, these features enable S-LoRA to serve thousands of LoRA adapters on a single GPU or across multiple GPUs with a small overhead. Compared to state-of-the-art libraries such as HuggingFace PEFT and vLLM (with naive support of LoRA serving), S-LoRA can improve the throughput by up to 4 times and increase the number of served adapters by several orders of magnitude. As a result, S-LoRA enables scalable serving of many task-specific fine-tuned models and offers the potential for large-scale customized fine-tuning services. The code is available at this https URL 

We study the problem of online learning in a two-player decentralized cooperative Stackelberg game. In each round, the leader first takes an action, followed by the follower who takes their action after observing the leader’s move. The goal of the leader is to learn to minimize the cumulative regret based on the history of interactions. Differing from the traditional formulation of repeated Stackelberg games, we assume the follower is omniscient, with full knowledge of the true reward, and that they always best-respond to the leader’s actions. We analyze the sample complexity of regret minimization in this repeated Stackelberg game. We show that depending on the reward structure, the existence of the omniscient follower may change the sample complexity drastically, from constant to exponential, even for linear cooperative Stackelberg games. 

We provide a theoretical framework for Reinforcement Learning with Human Feedback (RLHF). We show that when the underlying true reward is linear, under both Bradley-Terry-Luce (BTL) model (pairwise comparison) and Plackett-Luce (PL) model ($$K$$-wise comparison), MLE converges under certain semi-norm for the family of linear reward. On the other hand, when training a policy based on the learned reward model, we show that MLE fails while a pessimistic MLE provides policies with good performance under certain coverage assumption. We also show that under the PL model, both the true MLE and a different MLE which splits the $$K$$-wise comparison into pairwise comparisons converge, while the true MLE is asymptotically more efficient. Our results validate the empirical success of the existing RLHF algorithms, and provide new insights for algorithm design. Our analysis can also be applied for the problem of online RLHF and inverse reinforcement learning. 

Abstract We explore why many recently proposed robust estimation problems are efficiently solvable, even though the underlying optimization problems are non-convex. We study the loss landscape of these robust estimation problems, and identify the existence of ’generalized quasi-gradients’. Whenever these quasi-gradients exist, a large family of no-regret algorithms are guaranteed to approximate the global minimum; this includes the commonly used filtering algorithm. For robust mean estimation of distributions under bounded covariance, we show that any first-order stationary point of the associated optimization problem is an approximate global minimum if and only if the corruption level $$\epsilon < 1/3$$. Consequently, any optimization algorithm that approaches a stationary point yields an efficient robust estimator with breakdown point $1/3$. With carefully designed initialization and step size, we improve this to $1/2$, which is optimal. For other tasks, including linear regression and joint mean and covariance estimation, the loss landscape is more rugged: there are stationary points arbitrarily far from the global minimum. Nevertheless, we show that generalized quasi-gradients exist and construct efficient algorithms. These algorithms are simpler than previous ones in the literature, and for linear regression we improve the estimation error from $$O(\sqrt{\epsilon })$$ to the optimal rate of $$O(\epsilon )$$ for small $$\epsilon $$ assuming certified hypercontractivity. For mean estimation with near-identity covariance, we show that a simple gradient descent algorithm achieves breakdown point $1/3$ and iteration complexity $$\tilde{O}(d/\epsilon ^2)$$.