gemma-4-26B-A4B-it-W4A16

This repository contains a llm-compressor GPTQ export of google/gemma-4-26B-A4B-it using 4-bit weight-only quantization with bf16 activations.

What This Is

• Base model: google/gemma-4-26B-A4B-it
• Export format: compressed-tensors / pack-quantized
• Quantization style: W4A16 GPTQ
• Weight quantization: 4-bit signed integer, symmetric, grouped
• Weight group size: 64
• Activation dtype at runtime: bfloat16
• Modalities: text and image

This checkpoint uses group_size=64 intentionally. Gemma 4 26B A4B contains MoE down_proj widths such as 704 and 2112, which are not divisible by 128, so a default W4A16 G128 export is not valid for these layers.

Calibration Setup

• Text calibration baseline: HuggingFaceH4/ultrachat_200k
• Image calibration baseline: lmms-lab/flickr30k
• Calibration mode: mixed text/image

vLLM Serving

Serve Command (Works after "[Gemma4] Support quantized MoE" commit (3aecdf08b4a896a92e2cbd11c3d5a83d3c09abc1))

vllm serve dhruvil237/gemma-4-26B-A4B-it-W4A16 \
  --gpu-memory-utilization 0.8 \
  --reasoning-parser gemma4 \
  --dtype float16