HybridServe: Efficient Serving of Large AI Models with Confidence-Based Cascade Routing

Giant Deep Neural Networks (DNNs), have become indispensable for accurate and robust support of large-scale cloud based AI services. However, serving giant DNNs is prohibitively expensive from an energy consumption viewpoint easily exceeding that of training, due to the enormous scale of GPU clusters needed to hold giant DNN model partitions and replicas. Existing approaches can either optimize energy efficiency or inference accuracy but not both. To overcome this status quo, we propose HybridServe, a novel hybrid DNN model serving system that leverages multiple sized versions (small to giant) of the model to be served in tandem. Through a confidence based hybrid model serving dataflow, HybridServe prefers to serve inference requests with energy-efficient smaller models so long as accuracy is not compromised, thereby reducing the number of replicas needed for giant DNNs. HybridServe also features a dataflow planner for efficient partitioning and replication of candidate models to maximize serving system throughput. Experimental results using a prototype implementation of HybridServe show that it reduces energy footprint by up to 19.8x compared to the state-of-the-art DNN model serving systems while matching the accuracy of serving solely with giant DNNs.

MoE-CAP: Benchmarking Cost, Accuracy and Performance of Sparse Mixture-of-Experts Systems

The sparse Mixture-of-Experts (MoE) architecture is increasingly favored for scaling Large Language Models (LLMs) efficiently, but it depends on heterogeneous compute and memory resources. These factors jointly affect system Cost, Accuracy, and Performance (CAP), making trade-offs inevitable. Existing benchmarks often fail to capture these trade-offs accurately, complicating practical deployment decisions. To address this, we introduce MoE-CAP, a benchmark specifically designed for MoE systems. Our analysis reveals that achieving an optimal balance across CAP is difficult with current hardware; MoE systems typically optimize two of the three dimensions at the expense of the third-a dynamic we term the MoE-CAP trade-off. To visualize this, we propose the CAP Radar Diagram. We further introduce sparsity-aware performance metrics-Sparse Memory Bandwidth Utilization (S-MBU) and Sparse Model FLOPS Utilization (S-MFU)-to enable accurate performance benchmarking of MoE systems across diverse hardware platforms and deployment scenarios.

MoE-CAP: Cost-Accuracy-Performance Benchmarking for Mixture-of-Experts Systems

The sparse Mixture-of-Experts (MoE) architecture is increasingly favored for scaling Large Language Models (LLMs) efficiently; however, MoE systems rely on heterogeneous compute and memory resources. These factors collectively influence the system's Cost, Accuracy, and Performance (CAP), creating a challenging trade-off. Current benchmarks often fail to provide precise estimates of these effects, complicating practical considerations for deploying MoE systems. To bridge this gap, we introduce MoE-CAP, a benchmark specifically designed to evaluate MoE systems. Our findings highlight the difficulty of achieving an optimal balance of cost, accuracy, and performance with existing hardware capabilities. MoE systems often necessitate compromises on one factor to optimize the other two, a dynamic we term the MoE-CAP trade-off. To identify the best trade-off, we propose novel performance evaluation metrics - Sparse Memory Bandwidth Utilization (S-MBU) and Sparse Model FLOPS Utilization (S-MFU) - and develop cost models that account for the heterogeneous compute and memory hardware integral to MoE systems. This benchmark is publicly available on HuggingFace: https://huggingface.co/spaces/sparse-generative-ai/open-moe-llm-leaderboard.

Dynamic Self-Distillation via Previous Mini-batches for Fine-tuning Small Language Models

Knowledge distillation (KD) has become a widely adopted approach for compressing large language models (LLMs) to reduce computational costs and memory footprints. However, the availability of complex teacher models is a prerequisite for running most KD pipelines. Thus, the traditional KD procedure can be unachievable or budget-unfriendly, particularly when relying on commercial LLMs like GPT4. In this regard, Self-distillation (SelfD) emerges as an advisable alternative, enabling student models to learn without teachers' guidance. Nonetheless, existing SelfD approaches for LMs often involve architectural modifications, assuming the models are open-source, which may not always be practical. In this work, we introduce a model-agnostic and task-agnostic method named dynamic SelfD from the previous minibatch (DynSDPB), which realizes current iterations' distillation from the last ones' generated logits. Additionally, to address prediction inaccuracies during the early iterations, we dynamically adjust the distillation influence and temperature values to enhance the adaptability of fine-tuning. Furthermore, DynSDPB is a novel fine-tuning policy that facilitates the seamless integration of existing self-correction and self-training techniques for small language models (SLMs) because they all require updating SLMs' parameters. We demonstrate the superior performance of DynSDPB on both encoder-only LMs (e.g., BERT model families) and decoder-only LMs (e.g., LLaMA model families), validating its effectiveness across natural language understanding (NLU) and natural language generation (NLG) benchmarks.

Interactive and Expressive Code-Augmented Planning with Large Language Models

Large Language Models (LLMs) demonstrate strong abilities in common-sense reasoning and interactive decision-making, but often struggle with complex, long-horizon planning tasks. Recent techniques have sought to structure LLM outputs using control flow and other code-adjacent techniques to improve planning performance. These techniques include using variables (to track important information) and functions (to divide complex tasks into smaller re-usable sub-tasks). However, purely code-based approaches can be error-prone and insufficient for handling ambiguous or unstructured data. To address these challenges, we propose REPL-Plan, an LLM planning approach that is fully code-expressive (it can utilize all the benefits of code) while also being dynamic (it can flexibly adapt from errors and use the LLM for fuzzy situations). In REPL-Plan, an LLM solves tasks by interacting with a Read-Eval-Print Loop (REPL), which iteratively executes and evaluates code, similar to language shells or interactive code notebooks, allowing the model to flexibly correct errors and handle tasks dynamically. We demonstrate that REPL-Plan achieves strong results across various planning domains compared to previous methods.

DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Heads

Deploying long-context large language models (LLMs) is essential but poses significant computational and memory challenges. Caching all Key and Value (KV) states across all attention heads consumes substantial memory. Existing KV cache pruning methods either damage the long-context capabilities of LLMs or offer only limited efficiency improvements. In this paper, we identify that only a fraction of attention heads, a.k.a, Retrieval Heads, are critical for processing long contexts and require full attention across all tokens. In contrast, all other heads, which primarily focus on recent tokens and attention sinks--referred to as Streaming Heads--do not require full attention. Based on this insight, we introduce DuoAttention, a framework that only applies a full KV cache to retrieval heads while using a light-weight, constant-length KV cache for streaming heads, which reduces both LLM's decoding and pre-filling memory and latency without compromising its long-context abilities. DuoAttention uses a lightweight, optimization-based algorithm with synthetic data to identify retrieval heads accurately. Our method significantly reduces long-context inference memory by up to 2.55x for MHA and 1.67x for GQA models while speeding up decoding by up to 2.18x and 1.50x and accelerating pre-filling by up to 1.73x and 1.63x for MHA and GQA models, respectively, with minimal accuracy loss compared to full attention. Notably, combined with quantization, DuoAttention enables Llama-3-8B decoding with 3.3 million context length on a single A100 GPU. Code is provided in https://github.com/mit-han-lab/duo-attention.

ProTrain: Efficient LLM Training via Memory-Aware Techniques

It is extremely memory-hungry to train Large Language Models (LLM). To solve this problem, existing work exploits the combination of CPU and GPU for the training process, such as ZeRO-Offload. Such a technique largely democratizes billion-scale model training, making it possible to train with few consumer graphics cards. However, based on our observation, existing frameworks often provide coarse-grained memory management and require experienced experts in configuration tuning, leading to suboptimal hardware utilization and performance. This paper proposes ProTrain, a novel training system that intelligently balances memory usage and performance by coordinating memory, computation, and IO. ProTrain achieves adaptive memory management through Chunk-Based Model State Management and Block-Wise Activation Management, guided by a Memory-Aware Runtime Profiler without user intervention. ProTrain does not change the training algorithm and thus does not compromise accuracy. Experiments show that ProTrain improves training throughput by 1.43$\times$ to 2.71$\times$ compared to the SOTA training systems.

Challenges in Deploying Long-Context Transformers: A Theoretical Peak Performance Analysis

Transformer-based long context generative models power emerging AI applications like hour-long video understanding and project-level coding agent. Deploying long context transformers (e.g., 100K to 10M tokens) is prohibitively expensive compared to short context (e.g., 4K tokens) model variants. Reducing the cost of long-context transformers is becoming a pressing research and engineering challenge starting from the year of 2024. This work describes a concurrent programming framework for quantitatively analyzing the efficiency challenges in serving multiple long-context requests under limited size of GPU high-bandwidth memory (HBM) regime. We give a detailed analysis of how all additional computational costs, compared to 4K context, trace back to \textit{one single source: the large size of the KV cache}. We use a 34B GPT-3.5 level model of 50K context on A100 NVLink as a running example, and describe how its large KV cache causes four types of deployment challenges: (1) prefilling long inputs takes much longer compute time and GPU memory than short inputs; (2) after prefilling, the large KV cache residing on the GPU HBM substantially restricts the number of concurrent users being served; (3) during decoding, repeatedly reading the KV cache from HBM to SM largely increases latency; (4) when KV cache memory overflows, swapping it from HBM to DDR causes significant context switching latency. We use this framework to analyze existing works and identify possibilities of combining them to build end-to-end systems. Overall, this work offers a foundational framework for analyzing long context transformer deployment and identifies directions towards reducing the inference cost of 1M context to be as cheap as 4K.

Long Context Alignment with Short Instructions and Synthesized Positions

Effectively handling instructions with extremely long context remains a challenge for Large Language Models (LLMs), typically necessitating high-quality long data and substantial computational resources. This paper introduces Step-Skipping Alignment (SkipAlign), a new technique designed to enhance the long-context capabilities of LLMs in the phase of alignment without the need for additional efforts beyond training with original data length. SkipAlign is developed on the premise that long-range dependencies are fundamental to enhancing an LLM's capacity of long context. Departing from merely expanding the length of input samples, SkipAlign synthesizes long-range dependencies from the aspect of positions indices. This is achieved by the strategic insertion of skipped positions within instruction-following samples, which utilizes the semantic structure of the data to effectively expand the context. Through extensive experiments on base models with a variety of context window sizes, SkipAlign demonstrates its effectiveness across a spectrum of long-context tasks. Particularly noteworthy is that with a careful selection of the base model and alignment datasets, SkipAlign with only 6B parameters achieves it's best performance and comparable with strong baselines like GPT-3.5-Turbo-16K on LongBench.

Retrieval Head Mechanistically Explains Long-Context Factuality

Despite the recent progress in long-context language models, it remains elusive how transformer-based models exhibit the capability to retrieve relevant information from arbitrary locations within the long context. This paper aims to address this question. Our systematic investigation across a wide spectrum of models reveals that a special type of attention heads are largely responsible for retrieving information, which we dub retrieval heads. We identify intriguing properties of retrieval heads:(1) universal: all the explored models with long-context capability have a set of retrieval heads; (2) sparse: only a small portion (less than 5\%) of the attention heads are retrieval. (3) intrinsic: retrieval heads already exist in models pretrained with short context. When extending the context length by continual pretraining, it is still the same set of heads that perform information retrieval. (4) dynamically activated: take Llama-2 7B for example, 12 retrieval heads always attend to the required information no matter how the context is changed. The rest of the retrieval heads are activated in different contexts. (5) causal: completely pruning retrieval heads leads to failure in retrieving relevant information and results in hallucination, while pruning random non-retrieval heads does not affect the model's retrieval ability. We further show that retrieval heads strongly influence chain-of-thought (CoT) reasoning, where the model needs to frequently refer back the question and previously-generated context. Conversely, tasks where the model directly generates the answer using its intrinsic knowledge are less impacted by masking out retrieval heads. These observations collectively explain which internal part of the model seeks information from the input tokens. We believe our insights will foster future research on reducing hallucination, improving reasoning, and compressing the KV cache.