Gemma-4 2B IT - INT8 ConvRot Quantization

This is an INT8 quantization of the gemma4_e2b_it_bf16.safetensors text encoder from Comfy-Org/gemma-4, optimized for use in ComfyUI.

This quantized version significantly reduces memory usage while maintaining high quality, making it ideal for running Gemma-4 2B text encoders in memory-constrained environments.

Quantization Details

• Tool: silveroxides/convert_to_quant (CTQ)
• Format: INT8 row-wise quantization with embedded ConvRot metadata
• Scaling Mode: Row-wise
• ConvRot Group Size: 256
• Method: Learned rounding (AdaRound) with low-memory streaming conversion

Conversion Command

ctq -i gemma4_e2b_it_bf16.safetensors -o gemma4_e2b_it_int8_convrot.safetensors \
  --int8 --scaling_mode row \
  --convrot --convrot-group-size 256 \
  --comfy_quant --save-quant-metadata \
  --low-memory --device cuda \
  --exclude-layers "(^model\.embed_tokens(_per_layer)?\.weight$|^vision_model\.|^model\.layers\.(0|47)\.)" \
  --verbose NORMAL