Chaoyi Jiang

Xiangbo Qi, Chaoyi Jiang, Murali Annavaram

Large language models (LLMs) have grown beyond the memory capacity of single GPU devices, necessitating quantization techniques for practical deployment. While NF4 (4-bit NormalFloat) quantization enables 4$\times$ memory reduction, inference on current NVIDIA GPUs (e.g., Ampere A100) requires expensive dequantization back to FP16 format, creating a critical performance bottleneck. This paper presents a lightweight shared memory optimization that addresses this gap through principled memory hierarchy exploitation while maintaining full ecosystem compatibility. We compare our technique against the open-source BitsAndBytes implementation, achieving 2.0--2.2$\times$ kernel speedup across three models (Gemma 27B, Qwen3 32B, and Llama3.3 70B) and up to 1.54$\times$ end-to-end improvement by leveraging the 12--15$\times$ latency advantage of shared memory over global memory access. Our optimization reduces instruction counts through simplified indexing logic while using only 64 bytes of shared memory per thread block, demonstrating that lightweight optimizations can deliver substantial performance gains with minimal engineering effort. This work provides a plug-and-play solution for the HuggingFace ecosystem that democratizes access to advanced models on existing GPU infrastructure.

Hossein Entezari Zarch, Lei Gao, Chaoyi Jiang et al.

Speculative Decoding (SD) is a widely used approach to accelerate the inference of large language models (LLMs) without reducing generation quality. It operates by first using a compact model to draft multiple tokens efficiently, followed by parallel verification using the target LLM. This approach leads to faster inference compared to auto-regressive decoding. While there are multiple approaches to create a draft model, one promising approach is to use early-exit methods. These methods draft candidate tokens by using a subset of layers of the primary model and applying the remaining layers for verification, allowing a single model to handle both drafting and verification. While this technique reduces memory usage and computational cost, its performance relies on the choice of the exit layer for drafting and the number of tokens drafted (speculation length) in each SD round. Prior works use hyperparameter exploration to statically select these values. However, our evaluations show that these hyperparameter values are task-specific, and even within a task they are dependent on the current sequence context. We introduce DEL (Dynamic Exit Layer), a plug-and-play method that adaptively selects the exit layer and speculation length during inference. DEL dynamically tracks the token acceptance rate if the tokens are drafted at each layer of an LLM and uses that knowledge to heuristically select the optimal exit layer and speculation length. Our experiments across a broad range of models and downstream tasks show that DEL achieves overall speedups of $2.16\times$$\sim$$2.62\times$ over vanilla auto-regressive decoding and improves upon state-of-the-art SD methods, which peak at $2.43\times$, by up to $0.19\times$. The code is available at https://github.com/hoenza/DEL.

Chaoyi Jiang, Lei Gao, Hossein Entezari Zarch et al.

Inference for Large Language Models (LLMs) is computationally demanding. To reduce the cost of auto-regressive decoding, Key-Value (KV) cache is used to store intermediate activations, which significantly lowers the computational overhead for token generation. However, the memory required for the KV cache grows rapidly, often exceeding the capacity of GPU memory. A cost-effective alternative is to offload KV cache to CPU memory, which alleviates GPU memory pressure, but shifts the bottleneck to the limited bandwidth of the PCIe connection between the CPU and GPU. Existing methods attempt to address these issues by overlapping GPU computation with I/O or employing CPU-GPU heterogeneous execution, but they are hindered by excessive data movement and dependence on CPU capabilities. Fully overlapping PCIe communication latency gets challenging as the size of the KV cache grows and/or the GPU compute capabilities increase. In this paper, we introduce KVPR, an efficient I/O-aware LLM inference method where the CPU first transfers a partial set of activations, from which the GPU can start recomputing the KV cache values. While the GPU recomputes the partial KV cache, the remaining portion of the KV cache is transferred concurrently from the CPU. This approach overlaps GPU recomputation with KV cache transfer to minimize idle GPU time and maximize inference performance. KVPR is fully automated by integrating a profiler module that utilizes input characteristics and system hardware information, a scheduler module to optimize the distribution of computation and communication workloads, and a runtime module to efficiently execute the derived execution plan. Experimental results show that KVPR achieves up to 35.8% lower latency and 46.2% higher throughput during decoding compared to state-of-the-art approaches. The code is available at https://github.com/chaoyij/KVPR.

Hossein Entezari Zarch, Abdulla Alshabanah, Chaoyi Jiang et al.

Deep learning recommendation models (DLRMs) are at the heart of the current e-commerce industry. However, the amount of training data used to train these large models is growing exponentially, leading to substantial training hurdles. The training dataset contains two primary types of information: content-based information (features of users and items) and collaborative information (interactions between users and items). One approach to reduce the training dataset is to remove user-item interactions. But that significantly diminishes collaborative information, which is crucial for maintaining accuracy due to its inclusion of interaction histories. This loss profoundly impacts DLRM performance. This paper makes an important observation that if one can capture the user-item interaction history to enrich the user and item embeddings, then the interaction history can be compressed without losing model accuracy. Thus, this work, Collaborative Aware Data Compression (CADC), takes a two-step approach to training dataset compression. In the first step, we use matrix factorization of the user-item interaction matrix to create a novel embedding representation for both the users and items. Once the user and item embeddings are enriched by the interaction history information the approach then applies uniform random sampling of the training dataset to drastically reduce the training dataset size while minimizing model accuracy drop. The source code of CADC is available at \href{https://anonymous.4open.science/r/DSS-RM-8C1D/README.md}{https://anonymous.4open.science/r/DSS-RM-8C1D/README.md}.

Arun Ramachandran, Ramaswamy Govindarajan, Murali Annavaram et al.

With the skyrocketing costs of GPUs and their virtual instances in the cloud, there is a significant desire to use CPUs for large language model (LLM) inference. KV cache update, often implemented as allocation, copying, and in-place strided update for each generated token, incurs significant overhead. As the sequence length increases, the allocation and copy overheads dominate the performance. Alternate approaches may allocate large KV tensors upfront to enable in-place updates, but these matrices (with zero-padded rows) cause redundant computations. In this work, we propose a new KV cache allocation mechanism called Balancing Memory and Compute (BMC). BMC allocates, once every r iterations, KV tensors with r redundant rows, allowing in-place update without copy overhead for those iterations, but at the expense of a small amount of redundant computation. Second, we make an interesting observation that the extra rows allocated in the KV tensors and the resulting redundant computation can be repurposed for Speculative Decoding (SD) that improves token generation efficiency. Last, BMC represents a spectrum of design points with different values of r. To identify the best-performing design point(s), we derive a simple analytical model for BMC. The proposed BMC method achieves an average throughput acceleration of up to 3.2x over baseline HuggingFace (without SD). Importantly when we apply BMC with SD, it results in an additional speedup of up to 1.39x, over and above the speedup offered by SD. Further, BMC achieves a throughput acceleration of up to 1.36x and 2.29x over state-of-the-art inference servers vLLM and DeepSpeed, respectively. Although the BMC technique is evaluated extensively across different classes of CPUs (desktop and server class), we also evaluate the scheme with GPUs and demonstrate that it works well for GPUs.

Lei Gao, Chaoyi Jiang, Hossein Entezari Zarch et al.

Modern LLM serving systems must sustain high throughput while meeting strict latency SLOs across two distinct inference phases: compute-intensive prefill and memory-bound decode phases. Existing approaches either (1) aggregate both phases on shared GPUs, leading to interference between prefill and decode phases, which degrades time-between-tokens (TBT); or (2) disaggregate the two phases across GPUs, improving latency but wasting resources through duplicated models and KV cache transfers. We present DuetServe, a unified LLM serving framework that achieves disaggregation-level isolation within a single GPU. DuetServe operates in aggregated mode by default and dynamically activates SM-level GPU spatial multiplexing when TBT degradation is predicted. Its key idea is to decouple prefill and decode execution only when needed through fine-grained, adaptive SM partitioning that provides phase isolation only when contention threatens latency service level objectives (SLOs). DuetServe integrates (1) an attention-aware roofline model to forecast iteration latency, (2) a partitioning optimizer that selects the optimal SM split to maximize throughput under TBT constraints, and (3) an interruption-free execution engine that eliminates CPU-GPU synchronization overhead. Evaluations show that DuetServe improves total throughput by up to 1.3x while maintaining low generation latency compared to state-of-the-art frameworks.

Hossein Entezari Zarch, Lei Gao, Chaoyi Jiang et al.

Large reasoning models (LRMs) achieve state-of-the-art performance on challenging benchmarks by generating long chains of intermediate steps, but their inference cost is dominated by decoding, where each new token must attend to the entire growing sequence. Existing sparse attention methods reduce computation by pruning the key-value (KV) cache, yet they suffer from severe accuracy degradation on reasoning tasks due to cumulative selection errors and the dynamic importance of tokens over long derivations. We present \textbf{DELTA}, a training-free sparse attention mechanism that achieves computational efficiency without sacrificing model accuracy. DELTA partitions transformer layers into three groups: initial layers that use full attention, a small set of \emph{selection layers} that identify salient tokens via aggregated head-level attention scores, and subsequent \emph{sparse-attention layers} that attend only to the selected subset. This design preserves the full KV cache in GPU memory for accuracy, while avoiding expensive full-attention computation over many layers. On reasoning benchmarks such as AIME and GPQA-Diamond, DELTA matches or surpasses full attention in accuracy, while reducing the number of attended tokens by up to $5\times$ and delivering $1.5\times$ end-to-end speedup. Our results show that selective reuse of intermediate attention maps offers a robust path toward efficient long-context reasoning.

Chaoyi Jiang, Sungwoo Kim, Lei Gao et al.

Masked autoregressive (MAR) models unify the strengths of masked and autoregressive generation by predicting tokens in a fixed order using bidirectional attention for image generation. While effective, MAR models suffer from significant computational overhead, as they recompute attention and feed-forward representations for all tokens at every decoding step, despite most tokens remaining semantically stable across steps. We propose a training-free generation framework MARché to address this inefficiency through two key components: cache-aware attention and selective KV refresh. Cache-aware attention partitions tokens into active and cached sets, enabling separate computation paths that allow efficient reuse of previously computed key/value projections without compromising full-context modeling. But a cached token cannot be used indefinitely without recomputation due to the changing contextual information over multiple steps. MARché recognizes this challenge and applies a technique called selective KV refresh. Selective KV refresh identifies contextually relevant tokens based on attention scores from newly generated tokens and updates only those tokens that require recomputation, while preserving image generation quality. MARché significantly reduces redundant computation in MAR without modifying the underlying architecture. Empirically, MARché achieves up to 1.7x speedup with negligible impact on image quality, offering a scalable and broadly applicable solution for efficient masked transformer generation.