dowload_models

English | 简体中文

Model Download Script Guide

This script is for demonstration purposes only. It is for personal use only and must not be used for commercial purposes.

What it does

Downloads models from ModelScope, then moves and renames them under models\. It now supports interactive selection instead of always downloading everything. Downloads run concurrently (default 3 at a time), and each download is retried automatically on timeout or failure.

Available models (number, size, destination folder and description are listed at runtime):

Group	Model	Destination
LLM	DeviLeo/Qwen3-4B-int4-ov	models\llm
Embedding	DeviLeo/bge-m3-int4-sym-ov	models\emb
Rerank	DeviLeo/bge-reranker-v2-m3-int4-sym-ov	models\rerank
MMR	DeviLeo/gme-Qwen2-VL-2B-Instruct-int4-sym-ov	models\mmr\gme
ASR	DeviLeo/FunASR-ov	models\asr
OCR	DeviLeo/PaddleOCR-ov	models\ocr
Action	DeviLeo/BossActionRecognition	models\bar
VLM	DeviLeo/Qwen3-VL-4B-Instruct-int4-ov	models\vlm
Splitter	DeviLeo/zh_core_web_sm-3.8.0	models\splitter
Splitter	DeviLeo/en_core_web_sm-3.8.0	models\splitter

Only one splitter model can be used — both map to models\splitter. If both are selected, the Chinese splitter is kept and the English one is skipped.

Prerequisites

1. Python 3.9+
2. Install ModelScope SDK:

pip install modelscope

Usage

Run in your command line:

python download_modelscope_models.py

The script will:

1. Ask you to choose the script language (Chinese / English);
2. If --root is not given, ask for the GameAssistantToolServer directory (press Enter to use the current directory, so models land under ./models);
3. List all models with their size (based on the installed size under ..\GameAssistantToolServer\models), destination folder and a short description;
4. Choose which models to download:
   • On an interactive terminal: use ↑/↓ to move, Space to toggle, a to select all, Enter to confirm, q/0 to cancel;
   • In non-interactive environments (piped / scripted): falls back to numbered input — comma-separated numbers or ranges are supported (e.g. 1,3-5), all selects everything, 0 cancels;
5. Check the destination folders first — models that are already present are skipped;
6. After confirmation, download the remaining models (concurrently by default) and move them to their destinations. A failed or timed-out download is retried automatically.

Pre-select the splitter model as the default choice:

python download_modelscope_models.py --splitter zh
python download_modelscope_models.py --splitter en

Specify the GameAssistantToolServer directory (where models\ will be created; skips the interactive prompt):

python download_modelscope_models.py --root "C:\path\to\GameAssistantToolServer" --splitter zh

Download settings

Downloads run concurrently (default: 3 models at a time), each with a timeout and automatic retry on timeout or failure.

Option	Default	Description
--workers N	3	Number of models to download at the same time. Set 1 to download one at a time.
--retries N	3	Max retries per model after a failed / timed-out attempt. Set 0 to never retry.
--timeout N	1800	Per-attempt download timeout in seconds. Set 0 to disable the timeout.

python download_modelscope_models.py --workers 4 --retries 5 --timeout 3600

While concurrent downloads are running, the progress bars printed by the ModelScope SDK may interleave on the terminal — this is cosmetic only.

Notes

• Before downloading, the script checks each destination folder:
  • If the same model is already present (detected via the .modelscope_model_id marker, or by comparing the folder size) it is skipped;
  • If the folder exists but its content does not match (it may be a different model), you are asked whether to overwrite it;
  • If the folder is empty or missing, the model is downloaded normally.
• If all selected models are already present, the script finishes without downloading anything.
• After installing a model, the script writes a .modelscope_model_id marker into the folder for later detection.
• The script creates a temporary cache folder _modelscope_download_cache. It is cleaned up automatically when all selected models download successfully; if some models still fail after retries, the cache is kept so a re-run can resume the partial downloads.