Welcome to MIT HAN Lab! We focus on making AI faster, smarter, and more efficient. Our research covers a broad spectrum, including generative AI (e.g., LLMs and diffusion models), TinyML, system optimization and hardware design. By integrating algorithm and hardware expertise, we strive to push the frontiers of AI efficiency and performance.
Graduated PhD students: Ji Lin (OpenAI), Hanrui Wang (assistant professor @UCLA), Zhijian Liu (assistant professor @UCSD), Han Cai (NVIDIA Research), Haotian Tang (Google DeepMind).
Accelerating LLM and Generative AI [slides]:
🔥⚡ We release TinyChat 2.0, the latest version with significant advancements in prefilling speed of Edge LLMs and VLMs, 1.5-1.7x faster than the previous version of TinyChat. Please refer to our blog for more details.
DistriFusion is integrated in NVIDIA's TensorRT-LLM for distributed inference on high-resolution image generation.
🔥 NVIDIA TensorRT-LLM, AMD, Google Vertex AI, Amazon Sagemaker, Intel Neural Compressor, FastChat, vLLM, HuggingFace TGI, and LMDeploy adopt AWQ to improve LLM serving efficiency. Our AWQ models on HuggingFace has received over 6 million downloads.
Congrats on graduation! Cheers on the next move: Zhijian Liu: assistant professor at UCSD, Hanrui Wang: assistant professor at UCLA, Ji Lin: OpenAI, Han Cai: NVIDIA Research, Wei-Chen Wang (postdoc): Amazon, Wei-Ming Chen (postdoc): NVIDIA.
We show SmoothQuant can enable W8A8 quantization for Llama-1/2, Falcon, Mistral, and Mixtral models with negligible loss.
We supported VILA Vision Languague Models in AWQ & TinyChat! Check our latest demos with multi-image inputs!
StreamingLLM is integrated by HPC-AI Tech SwiftInfer to support infinite input length for LLM inference.
StreamingLLM is integrated by CMU, UW, and OctoAI, enabling endless and efficient LLM generation on iPhone!
Congrats Ji Lin completed and defended his PhD thesis: "Efficient Deep Learning Computing: From TinyML to Large Language Model". Ji joined OpenAI after graduation.
AWQ is integrate by NVIDIA TensorRT-LLM, can fit Falcon-180B on a single H200GPU with INT4 AWQ, and 6.7x faster Llama-70B over A100.
TorchSparse++ has been adopted by One-2-3-45++ from Prof. Hao Su's lab (UCSD) for 3D object generation!
🔥 AWQ is now integrated natively in Hugging Face transformers through from_pretrained. You can either load quantized models from the Hub or your own HF quantized models.
Attention Sinks, an library from community enables StreamingLLM on more Huggingface LLMs. blog.
TorchSparse++ has been adopted by One-2-3-45 from Prof. Hao Su's lab (UCSD) for 3D mesh reconstruction!
Large language models (LLMs) have shown remarkable potential in processing long sequences, yet efficiently serving these long-context models remains challenging due to the quadratic computational complexity of attention in the prefilling stage and the large memory footprint of the KV cache in the decoding stage. To address these issues, we introduce LServe, an efficient system that accelerates long-sequence LLM serving via unified sparse attention. This method unifies different hardware-friendly, structured sparsity patterns for both prefilling and decoding attention into a single framework, where computations on less important tokens are skipped block-wise. LServe demonstrates the compatibility of static and dynamic sparsity in long-context LLM attention. This design enables multiplicative speedups by combining these optimizations. Specifically, we convert half of the attention heads to nearly free streaming heads in both the prefilling and decoding stages. Additionally, we find that only a constant number of KV pages is required to preserve long-context capabilities, irrespective of context length. We then design a hierarchical KV page selection policy that dynamically prunes KV pages based on query-centric similarity. For Llama-3-8B, LServe accelerates LLM prefilling by an average of 2.4x and decoding by up to 3.3x over the state-of-the-art serving systems, maintaining long-context accuracy.
LServe accelerates long-sequence LLM serving with unified sparse attention for both prefilling and decoding, achieving up to 3.3× speedup over state-of-the-art solution without sacrificing accuracy.
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 1.2× on A100, 1.4× on L40S; and Qwen1.5-72B by 2.4× on A100, 3.5× 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 3×.
QServe accelerates large-scale LLM serving on GPUs with QoQ (W4A8KV4) quantization, boosting the generation throughputs by up to 3x over the state-of-the-art solution.
Diffusion models have been proven highly effective at generating high-quality images. However, as these models grow larger, they require significantly more memory and suffer from higher latency, posing substantial challenges for deployment. In this work, we aim to accelerate diffusion models by quantizing their weights and activations to 4 bits. At such an aggressive level, both weights and activations are highly sensitive, where conventional post-training quantization methods for large language models like smoothing become insufficient. To overcome this limitation, we propose SVDQuant, a new 4-bit quantization paradigm. Different from smoothing which redistributes outliers between weights and activations, our approach absorbs these outliers using a low-rank branch. We first consolidate the outliers by shifting them from activations to weights, then employ a high-precision low-rank branch to take in the weight outliers with Singular Value Decomposition (SVD). This process eases the quantization on both sides. However, naïvely running the low-rank branch independently incurs significant overhead due to extra data movement of activations, negating the quantization speedup. To address this, we co-design an inference engine Nunchaku that fuses the kernels of the low-rank branch into those of the low-bit branch to cut off redundant memory access. It can also seamlessly support off-the-shelf low-rank adapters (LoRAs) without the need for re-quantization. Extensive experiments on SDXL, PixArt-∑, and FLUX.1 validate the effectiveness of SVDQuant in preserving image quality. We reduce the memory usage for the 12B FLUX.1 models by 3.5×, achieving 3.0× speedup over the 4-bit weight-only quantized baseline on the 16GB laptop 4090 GPU, paving the way for more interactive applications on PCs. Our quantization library and inference engine are open-sourced.
A new W4A4 quantization paradigm for diffusion models.
FP8 training has emerged as a promising method for improving training efficiency. Existing frameworks accelerate training by applying FP8 computation to linear layers while leaving optimizer states and activations in higher precision, which fails to fully optimize memory usage. This paper introduces COAT (Compressing Optimizer States and Activations for FP8 Training), a novel FP8 training framework designed to significantly reduce memory footprint when training large models. COAT addresses current limitations through two key innovations: (1) Dynamic Range Expansion, which aligns optimizer state distributions more closely with the FP8 representation range, thereby reducing quantization error, and (2) Mixed-Granularity Activation Quantization, which optimizes activation memory using a combination of per-tensor and per-group quantization strategies. Experiments demonstrate that COAT effectively reduces end-to-end training memory footprint by 1.54x compared to BF16 while achieving nearly lossless performance across various tasks, such as Large Language Model pretraining and fine-tuning and Vision Language Model training. COAT also achieves a 1.43x end-to-end training speedup compared to BF16, performing on par with or surpassing TransformerEngine's speedup. COAT enables efficient full-parameter training of large models on fewer GPUs, and facilitates doubling the batch size in distributed training settings, providing a practical solution for scaling large-scale model training.
We propose COAT, a memory efficient FP8 training method for large language models.
We actively collaborate with industry partners on efficient AI, model compression and acceleration. Our research has influenced and landed in many industrial products: Intel OpenVino, Intel Neural Network Distiller, Intel Neural Compressor, Apple Neural Engine, NVIDIA Sparse Tensor Core, NVIDIA TensorRT LLM, AMD-Xilinx Vitis AI, Qualcomm AI Model Efficiency Toolkit (AIMET), Amazon AutoGluon, Facebook PyTorch, Microsoft NNI, SONY Neural Architecture Search Library, SONY Model Compression Toolkit, ADI MAX78000/MAX78002 Model Training and Synthesis Tool.
