About LLM Compressor

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.

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.