From 309b40aaf097e840f55e8c975bbbaa5f30a62646 Mon Sep 17 00:00:00 2001
From: haixuanTao <tao.xavier@outlook.com>
Date: Sun, 13 Oct 2024 06:02:07 +0200
Subject: [PATCH] Add `modelscope` to qwenvl

---
 node-hub/dora-qwenvl/dora_qwenvl/main.py | 17 +++++++++++++----
 node-hub/dora-qwenvl/pyproject.toml      |  1 +
 2 files changed, 14 insertions(+), 4 deletions(-)

diff --git a/node-hub/dora-qwenvl/dora_qwenvl/main.py b/node-hub/dora-qwenvl/dora_qwenvl/main.py
index 4538c63e..7a2b6119 100644
--- a/node-hub/dora-qwenvl/dora_qwenvl/main.py
+++ b/node-hub/dora-qwenvl/dora_qwenvl/main.py
@@ -5,10 +5,19 @@ from qwen_vl_utils import process_vision_info
 import numpy as np
 import pyarrow as pa
 from PIL import Image
+from pathlib import Path
 import cv2
 
 DEFAULT_PATH = "Qwen/Qwen2-VL-2B-Instruct"
-CUSTOM_MODEL_PATH = os.getenv("CUSTOM_MODEL_PATH", DEFAULT_PATH)
+
+MODEL_NAME_OR_PATH = os.getenv("MODEL_NAME_OR_PATH", DEFAULT_PATH)
+
+if bool(os.getenv("MODELSCOPE")) is True:
+    from modelscope import snapshot_download
+
+    if not Path(MODEL_NAME_OR_PATH).exists():
+        MODEL_NAME_OR_PATH = snapshot_download(MODEL_NAME_OR_PATH)
+
 DEFAULT_QUESTION = os.getenv(
     "DEFAULT_QUESTION",
     "Describe this image",
@@ -20,14 +29,14 @@ try:
     import flash_attn as _
 
     model = Qwen2VLForConditionalGeneration.from_pretrained(
-        CUSTOM_MODEL_PATH,
+        MODEL_NAME_OR_PATH,
         torch_dtype="auto",
         device_map="auto",
         attn_implementation="flash_attention_2",
     )
 except (ImportError, ModuleNotFoundError):
     model = Qwen2VLForConditionalGeneration.from_pretrained(
-        CUSTOM_MODEL_PATH,
+        MODEL_NAME_OR_PATH,
         torch_dtype="auto",
         device_map="auto",
     )
@@ -38,7 +47,7 @@ if ADAPTER_PATH != "":
 
 
 # default processor
-processor = AutoProcessor.from_pretrained(CUSTOM_MODEL_PATH)
+processor = AutoProcessor.from_pretrained(MODEL_NAME_OR_PATH)
 
 
 def generate(frames: dict, question):
diff --git a/node-hub/dora-qwenvl/pyproject.toml b/node-hub/dora-qwenvl/pyproject.toml
index eac55946..f4f9f1bd 100644
--- a/node-hub/dora-qwenvl/pyproject.toml
+++ b/node-hub/dora-qwenvl/pyproject.toml
@@ -20,6 +20,7 @@ transformers = "^4.45"
 qwen-vl-utils = "^0.0.2"
 accelerate = "^0.33"
 opencv-python = ">= 4.1.1"
+modelscope = "^1.18.1"
 # flash_attn = "^2.6.1" # Install using: pip install -U flash-attn --no-build-isolation