code for i2t inference

[pkulium]

great work! I am wondering can you provide code for i2t inference? There is only code for t2i inference now(scripts/run_t2i.sh)

[QuLiao1117]

great work! I am wondering can you provide code for i2t inference? There is only code for t2i inference now(scripts/run_t2i.sh)

The i2t of TokenFlow is built on LLaVA, so you can directly use lmms-eval for inference and evaluation.

Specifically,

1. First install the i2t environment by following the installation guide. This will install a "llava" pip package.
2. Then refer to the LLaVA testing instructions in the lmms-eval repository. Example command:

python3 -m accelerate.commands.launch \
    --num_processes=8 \
    -m lmms_eval \
    --model llava \
    --model_args pretrained="ByteFlow-AI/Tokenflow-llava-qwen2.5-14B-finetuning" \
    --tasks mme \
    --batch_size 1 \
    --log_samples \
    --log_samples_suffix llava_v1.5_mme \
    --output_path ./logs/
    # you can also change the `--model_args pretrained=` to local path of your model.

Things might need to be modified:

• In case lmm-eval can't find the llava package, add a system path in lmms_eval/models/llava.py

from pathlib import Path
sys.path.insert(0, str(Path("/path/to/your/TokenFlow/i2t").resolve()))

[pkulium]

Thank you for your reply. Yes, I have used lmms-eval before. I am wondering if there is code available to run inference on a single image of my own?

[QuLiao1117]

Thank you for your reply. Yes, I have used lmms-eval before. I am wondering if there is code available to run inference on a single image of my own?

This repo does not support this feature currently. PR is welcomed to enable this.