Q1 2026 Roadmap

[dsikka]

Overview

The near-term focus for LLM Compressor will centre on performance improvements across core workflows, targeted enhancements to NVFP4, stabilizing and hardening MXFP4 support, and broad improvements to modifier functionality. These efforts aim to improve efficiency, robustness, and overall usability while enabling more reliable and scalable model compression workflows.

In addition, we will continue to expand quantization support for the latest model releases to ensure timely compatibility with newly introduced architectures and checkpoints, including adopting transformers v5.0

We will also be focusing on improving the quality of our documentation, examples, CI/CD for easier access and understanding.

Q1 Roadmap

Performance Refactor - Enable Distributed Quantization Support

Status: In Progress

RFC: #2180

Issues:

• [Performance Refactor] Replace accelerate’s offloading with distributed-friendly implementations #2215
• [Performance Refactor] Implement a user interface for loading distributed/offloaded models #2216
  • [Offload] Convert accelerate for loading/saving compressed-tensors#572
• [Performance Refactor] Add data parallelism support to LLM Compressor #2217

Enable Modifier Specific Support:

• [Performance Refactor] Extend modifiers to support weight-parallel optimization - GPTQModifier #2218
• [Performance Refactor] Extend modifiers to support weight-parallel optimization - AWQModifier #2219
• [Performance Refactor] Extend modifiers to support weight-parallel optimization - QuantizationModifier #2220
• [AutoRound] Add DDP Support and Example #2411

MXFP4 vLLM Integration / Validation

Status: In Progress

• MXFP4A16 Support :
  • [MXFP4A16] Add Qwen3 MoE Example; reorganize examples #2227
  • [Quant] Support MXFP4 W4A16 for compressed-tensors MoE models  vllm#32285
  • [Quant] Support MXFP4 W4A16 for compressed-tensors dense models vllm#31926
• MXFP4 Support - Activation Quantization Validation (move examples out of the experimental folder)

MXFP8 Support

Status: In Progress

• Add MXFP8 Support compressed-tensors#630
• Add MXFP8 Example #2487

AWQ, GPTQ Improvements and Benchmarking

Status: In Progress

• [AWQ] Modifier Speedups #2265
• [AWQ] Option to disable quantization #2206
• [AWQ] Add Token Masking Support for Calibration #2250
• [NVFP4][GPTQ] Fix GPTQModifier + NVFP4 Support #2294
• [MXFP4][GPTQ] Extend rounding to support FP32 compressed-tensors#551
• [MXFP4][GPTQ] Add GPTQ + MXFP4A16 Example #2304

NVFP4 Improvements

Status: Not Yet Started

• vLLM Support for NVFP4 + micro-rotations: [RFC]: MR-GPTQ (GPTQ+NVFP4) #2006
• General benchmarking with AutoRound, QuantizationModifier, and AWQ

Transformers v5 Support

Status: Not Yet Started

• Support for updated MoE Calibration: [upstream] Expecting future huggingface/transformer incompatibility #2036

Quantized Model Support

Status: In Progress

• FP8 Block + NVFp4 DSR1 Support
• GLM Support: Added GLM Modeling #2170
• MinMax-M2 Support: Minimax-M2 / M2.1 calibration #2171
• Qwen3.5 Support:
  - Calibration Support: [Qwen3.5] Calibration support and NVFP4 Example #2482
  - VL Examples: [Examples] Add Qwen3.5-27B NVFP4A16 and MXFP4A16 quantization examples #2467

CI/CD Buildkite Migration

Status: In Progress

• Migrate CI/CD to Buildkite