PhotoGuider explainer base, NF4
The base model of the PhotoGuider explainer:
Qwen/Qwen2.5-VL-7B-Instruct
(revision cc594898137f460bfe9f0759e9844b3ce807cfb5) quantized once with bitsandbytes to 4-bit NF4
(bnb_4bit_compute_dtype=bfloat16, llm_int8_skip_modules=["visual", "lm_head"]), exactly the
configuration PhotoGuider applies when it loads the bf16 checkpoint. Loading these weights reads
6 GB instead of quantizing the 16 GB checkpoint on every load; the generated explanations are
identical token for token.
PhotoGuider adds its released LoRA adapter on top (tools/download_explainer.sh in the repository)
and decodes greedily. This repository contains no adapter and no training data. The model card,
license and acceptable-use terms of the base model apply unchanged.
# inside a PhotoGuider checkoutenvs/explainer/venv/bin/python tools/download_explainer_nf4.py