Skip to content

What is LLM Compressor?

LLM Compressor is an easy-to-use library for optimizing large language models for deployment with vLLM. It provides a comprehensive toolkit for applying state-of-the-art compression algorithms to reduce model size, lower hardware requirements, and improve inference performance.

LLM Compressor Flow

What challenges does LLM Compressor address?

Model optimization through quantization and pruning addresses the key challenges of deploying AI at scale:

Challenge How LLM Compressor helps
GPU and infrastructure costs Reduces memory requirements by 50-75%, enabling deployment on fewer GPUs
Response latency Reduces data movement overhead because quantized weights load faster
Request throughput Utilizes lower-precision tensor cores for faster computation
Energy consumption Smaller models consume less power during inference

For more information, see Why use LLM Compressor?

New in this release

Review the LLM Compressor v0.8.0 release notes for details about new features. Highlights include:

Support for multiple modifiers in oneshot compression runs

LLM Compressor now supports using multiple modifiers in oneshot compression runs such as applying both AWQ and GPTQ in a single model.

Using multiple modifiers is an advanced usage of LLM Compressor and an active area of research. See Non-uniform Quantization for more detail and example usage.

Quantization and calibration support for Qwen3 models

Quantization and calibration support for Qwen3 Next models has been added to LLM Compressor.

LLM Compressor now supports quantization for Qwen3 Next and Qwen3 VL MoE models. You can now use data-free pathways such as FP8 channel-wise and block-wise quantization. Pathways requiring data such W4A16 and NVFP4 are planned for a future release.

Examples for NVFP4 and FP8 quantization have been added for the Qwen3-Next-80B-A3B-Instruct model.

For the Qwen3 VL MoE model, support has been added for the data-free pathway. The data-free pathway applies FP8 quantization, for example, channel-wise and block-wise quantization.

NOTE: These models are not supported in tranformers<=4.56.2. You may need to install transformers from source.

Transforms support for non-full-size rotation sizes

You can now set a transform_block_size field in the Transform-based modifier classes SpinQuantModifier and QuIPModifier. You can configure transforms of variable size with this field, and you don't need to restrict hadamards to match the size of the weight.

Recent updates

QuIP and SpinQuant-style Transforms

The newly added QuIPModifier and SpinQuantModifier transforms allow you to quantize models after injecting hadamard weights into the computation graph, reducing quantization error and greatly improving accuracy recovery for low bit-weight and activation quantization.

DeepSeekV3-style Block Quantization Support

Allows for more efficient compression of large language models without needing a calibration dataset. Quantize a Qwen3 model to W8A8.

FP4 Quantization - now with MoE and non-uniform support

Quantize weights and activations to FP4 and seamlessly run the compressed model in vLLM. Model weights and activations are quantized following the NVFP4 configuration. See examples of FP4 activation support, MoE support, and Non-uniform quantization support where some layers are selectively quantized to FP8 for better recovery. You can also mix other quantization schemes, such as INT8 and INT4.

Llama4 Quantization Support

Quantize a Llama4 model to W4A16 or NVFP4. The checkpoint produced can seamlessly run in vLLM.

For more information, check out the latest release on GitHub.

Supported algorithms

Algorithm Description Use Case
RTN (Round-to-Nearest) Fast baseline quantization Quick compression with minimal setup
GPTQ Weighted quantization with calibration High-accuracy 4-bit weight quantization
AWQ Activation-aware weight quantization Preserves accuracy for important weights
SmoothQuant Outlier handling for W8A8 Weight and activation quantization
SparseGPT Pruning with quantization 2:4 sparsity patterns
SpinQuant Rotation-based transforms Improved low-bit accuracy
QuIP Incoherence processing Advanced quantization preprocessing
FP8 KV Cache KV cache quantization Long context inference on Hopper-class and newer GPUs

Supported quantization formats

Format Targets Compute Capability Use Case
W4A16 Weights only 8.0 (Ampere and up) Optimize for latency on older hardware
W8A8-INT8 Weights + activations 7.5 (Turing and up) Balanced performance and compatibility
W8A8-FP8 Weights + activations 8.9 (Hopper and up) High throughput on modern GPUs
NVFP4/MXFP4 Weights + activations 10.0 (Blackwell) Maximum compression on latest hardware
W4AFP8 Mixed precision 8.9 (Hopper and up) Low-bit weights with FP8 activations
W4AINT8 Mixed precision 7.5 (Turing and up) Low-bit weights with INT8 activations
2:4 Sparse Weights 8.0 (Ampere and up) Sparsity-accelerated inference

Note

Listed compute capability indicates the minimum architecture required for hardware acceleration.