QServe: W4A8KV4 Quantization and System Co-design for Efficient LLM Serving
read the original abstract
Quantization can accelerate large language model (LLM) inference. Going beyond INT8 quantization, the research community is actively exploring even lower precision, such as INT4. Nonetheless, state-of-the-art INT4 quantization techniques only accelerate low-batch, edge LLM inference, failing to deliver performance gains in large-batch, cloud-based LLM serving. We uncover a critical issue: existing INT4 quantization methods suffer from significant runtime overhead (20-90%) when dequantizing either weights or partial sums on GPUs. To address this challenge, we introduce QoQ, a W4A8KV4 quantization algorithm with 4-bit weight, 8-bit activation, and 4-bit KV cache. QoQ stands for quattuor-octo-quattuor, which represents 4-8-4 in Latin. QoQ is implemented by the QServe inference library that achieves measured speedup. The key insight driving QServe is that the efficiency of LLM serving on GPUs is critically influenced by operations on low-throughput CUDA cores. Building upon this insight, in QoQ algorithm, we introduce progressive quantization that can allow low dequantization overhead in W4A8 GEMM. Additionally, we develop SmoothAttention to effectively mitigate the accuracy degradation incurred by 4-bit KV quantization. In the QServe system, we perform compute-aware weight reordering and take advantage of register-level parallelism to reduce dequantization latency. We also make fused attention memory-bound, harnessing the performance gain brought by KV4 quantization. As a result, QServe improves the maximum achievable serving throughput of Llama-3-8B by 1.2x on A100, 1.4x on L40S; and Qwen1.5-72B by 2.4x on A100, 3.5x on L40S, compared to TensorRT-LLM. Remarkably, QServe on L40S GPU can achieve even higher throughput than TensorRT-LLM on A100. Thus, QServe effectively reduces the dollar cost of LLM serving by 3x. Code is available at https://github.com/mit-han-lab/omniserve.
This paper has not been read by Pith yet.
Forward citations
Cited by 25 Pith papers
-
QCFuse: Query-Aware Cache Fusion via Compressed View for Efficient RAG Serving
QCFuse achieves full-prefill quality in RAG with 1.7x average prefill speedup over full prefill and 1.5x over ProphetKV via compressed query-aware cache fusion.
-
Multi-Scale Dequant: Eliminating Dequantization Bottleneck via Activation Decomposition for Efficient LLM Inference
MSD eliminates dequantization from the GEMM path by decomposing BF16 activations into multiple low-precision parts that multiply directly with INT8 or MXFP4 weights, achieving near-16 effective bits for INT8 and 6.6 f...
-
DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Heads
DuoAttention identifies retrieval heads requiring full KV cache and streaming heads using constant-length cache to reduce memory and latency in long-context LLM inference.
-
MxGLUT: A Reconfigurable LUT-Centric Broadcast Dataflow Accelerator for Mixed-Precision GEMM
MxGLUT introduces a reconfigurable LUT-centric broadcast dataflow accelerator with mixed-precision LUT-based PEs that unifies FP8-INT4 and FP8-FP8 GEMM without separate FP datapaths, reporting up to 2.16x prefill spee...
-
TileFuse: A Fused Mixed-Precision Kernel Library for Efficient Quantized LLM Inference on AMD NPUs
TileFuse introduces fused kernels and data layouts for W4A16/W8A16 on AMD XDNA2 NPUs, reporting up to 2.0x lower LLM prefilling latency and 64.6% lower energy versus baselines.
-
SPARQLe: Sub-Precision Activation Representation for Quantized LLM Inference
SPARQLe is a hardware-software co-design that splits quantized activations into dense low bits and sparse high bits to run inference on narrower datapaths while claiming to preserve full-precision accuracy.
-
Runtime-Certified Bounded-Error Quantized Attention
A tiered KV cache architecture computes per-head per-step error bounds on quantized attention and uses adaptive fallback to guarantee bounded or exact outputs relative to FP16 reference.
-
Search Your Block Floating Point Scales!
ScaleSearch optimizes block floating point scales via fine-grained search to cut quantization error by 27% for NVFP4, improving PTQ by up to 15 points on MATH500 for Qwen3-8B and attention PPL by 0.77 on Llama 3.1 70B.
-
WindowQuant: Mixed-Precision KV Cache Quantization based on Window-Level Similarity for VLMs Inference Optimization
WindowQuant performs window-adaptive mixed-precision KV cache quantization guided by similarity to the text prompt, with reordering to enable efficient inference in VLMs.
-
AQPIM: Breaking the PIM Capacity Wall for LLMs with In-Memory Activation Quantization
AQPIM performs in-memory product quantization of activations for LLMs on PIM hardware, reducing GPU-CPU communication by 90-98.5% and delivering 3.4x speedup over prior PIM methods.
-
FlexiCache: Leveraging Temporal Stability of Attention Heads for Efficient KV Cache Management
FlexiCache reduces GPU memory for long-context LLM requests by up to 70% and boosts throughput 1.38-1.55x and latency 1.6-2.1x by exploiting per-head differences in temporal stability of critical tokens.
-
LogQuant: Log-Distributed 2-Bit Quantization of KV Cache with Superior Accuracy Preservation
LogQuant applies log-based filtering for 2-bit KV cache quantization in LLMs, claiming 25% higher throughput, 60% larger batches, and 40-200% accuracy gains on math/code tasks versus existing compression approaches.
-
Analytical FFN-to-MoE Restructuring via Activation Pattern Analysis
An analytical post-training method restructures FFNs into MoE by partitioning neurons based on activation patterns and building a router from statistics, achieving 1.17x speedup with minimal resources.
-
Demystifying the Design Space and Best Practices for Heterogeneous LLM Inference and Serving
Organizes the heterogeneous LLM prefill-decode design space along four axes and extracts three boundary decisions with guidance on precision, KV representation, and ownership.
-
Demystifying the Design Space and Best Practices for Heterogeneous LLM Inference and Serving
The paper organizes heterogeneous prefill-decode LLM serving into a four-axis design space and identifies three recurring boundary decisions that require joint choices.
-
PALUTE: Processing-In-Memory Acceleration via Lookup Table for Edge LLM Inference
PALUTE is a new PIM accelerator using in-DRAM LUTs on M3D DRAM that reports 1264 TPS at 0.16 W with 12.8x energy efficiency gains over CHIME for quantized edge LLM inference.
-
Cassandra: Enabling Reasoning LLMs at Edge via Self-Speculative Decoding
Cassandra is a self-speculative decoding system that builds a draft model via fine-grained data selection and optimized pruning/mantissa truncation, achieving up to 2.41x speedup over BF16 and 1.81x more tokens than E...
-
EVA: Accelerating LLM Decoding via an Efficient Vector Quantization Architecture
EVA is a vector-quantization hardware architecture that transforms LLM decoding from GEMV to GEMM via direct codebook dot products and conflict-free output buffering, claiming up to 11.17x speedup over prior lookup designs.
-
IO-SVD: Input-Output Whitened SVD for Adaptive-Rank LLM Compression
IO-SVD performs SVD-based LLM compression by constructing a KL-aware double-sided whitening space and using first-order loss estimates for heterogeneous rank allocation.
-
LMDeploy Accelerates Mixed-Precision LLM Inference with TurboMind
TurboMind delivers up to 61% lower latency and 156% higher throughput for mixed-precision LLM inference across 16 models and 4 GPU architectures via optimized weight packing, adaptive alignment, instruction parallelis...
-
MixLLM: LLM Quantization with Global Mixed-precision between Output-features and Highly-efficient System Design
MixLLM uses global output-feature importance to set mixed bit-widths for LLM quantization and adds two-step dequantization plus software pipelining for system efficiency.
-
SEADA: An efficient methodology for optimizing mixed-precision DNNs on multi-precision spatial architectures
SEADA introduces an analytical framework combining cost models, mapping tools, and entropy-based precision selection to optimize mixed-precision DNNs on multi-precision spatial architectures.
-
TileFuse: A Fused Mixed-Precision Kernel Library for Efficient Quantized LLM Inference on AMD NPUs
TileFuse introduces a fused kernel library enabling AWQ W4A16/W8A16 quantized LLM inference on AMD NPUs, reporting up to 2.0x lower prefilling latency and 64.6% lower energy on Ryzen AI laptops.
-
DuQuant++: Fine-grained Rotation Enhances Microscaling FP4 Quantization
DuQuant++ adapts outlier-aware fine-grained rotation to MXFP4 by matching block size to the 32-element microscaling group, enabling a single rotation that smooths distributions and achieves SOTA performance on LLaMA-3...
-
A Comparative Study of CNN Optimization Methods for Edge AI: Exploring the Role of Early Exits
Combining pruning, quantization, and early exits in CNNs reduces inference latency and memory on real edge devices with minimal accuracy loss.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.