Qwen2.5-VL-7B-Instruct-int8-convrot-comfyui

Qwen2.5-VL-7B-Instruct INT8-ConvRot (ComfyUI Native)

ComfyUI-native INT8 quantization with ConvRot (group-wise Hadamard rotation) of Qwen/Qwen2.5-VL-7B-Instruct — the text encoder used by the Qwen-Image family: Qwen-Image (base), Qwen-Image-2512 (text-to-image generation), and Qwen-Image-Edit-2511 (image editing).

This is the first public INT8-ConvRot conversion of the vanilla (non-abliterated) instruct model. The weights load directly in ComfyUI's native Load CLIP node — no custom nodes required.

Recipe / how it was made: github.com/AnsteinHuynh/qwen-2.5-vl-int8-convrot-comfyui — reproducible tooling + docs (shard merge, convert scripts for Windows/Linux, option guide).

Key facts

Usage (ComfyUI)

1. Put qwen_2.5_vl_7b_instruct_int8_convrot.safetensors in ComfyUI/models/clip/
2. Load CLIP → select the file → type qwen_image
3. Works with Qwen-Image, Qwen-Image-2512 (text-to-image), and Qwen-Image-Edit-2511 (editing) workflows — all use the same qwen_image clip type (e.g. TextEncodeQwenImageEdit, TextEncodeQwenImageEditPlus)

How it was made

ctq -i Qwen2.5-VL-7B-Instruct-bf16.safetensors \
    -o Qwen2.5-VL-7B-Instruct_int8-convrot-tensor-simple.safetensors \
    --int8 --scaling_mode tensor --convrot --qwen35 \
    --comfy_quant --save-quant-metadata --low-memory --simple

• Tool: silveroxides/convert_to_quant (ctq v1.3.1)
• ConvRot: group-wise Hadamard rotation (group size 256) — spreads activation/weight outliers before quantization (QuaRot/ConvRot line of work), a small but free quality gain
• Scaling: per-tensor (int8_tensorwise), matching the recipe used for the popular Gemma-3-12B INT8-ConvRot text encoder
• --simple: fast path, no learned-rounding (SVD) optimization — output verified in ComfyUI before release
• --qwen35 filter: keeps first/last layers, embeddings, lm_head, and vision-tower components in high precision

Notes & limitations

• INT8 trades a small amount of fidelity for ~45% smaller footprint and faster loads. Test on your own prompts.
• This is a quantization of Qwen's model — no fine-tuning was performed. Redistribution under Apache 2.0 with attribution.
• SHA256SUMS included for integrity verification.

Credits

• Base model: Qwen/Qwen2.5-VL-7B-Instruct (Apache 2.0)
• Quantization tooling: silveroxides/convert_to_quant
• ComfyUI native INT8 support: Comfy-Org/ComfyUI