Temporary fix qwenvl2 queue error (#688)

In order to get the best performance from Qwenvl2, we need to temporary
add this piece of code that is going to unqueue qwenvl queue.

This is directly linked to: https://github.com/dora-rs/dora/pull/652
which is that we currently having a queue issue within nodes for
`queue_size: 1` where the queue is actually pending older data and not
the newest one.

examples/vlm/dataflow.yml

@@ -18,7 +18,7 @@ nodes:
       image:
         source: camera/image
         queue_size: 1
-      tick: dora/timer/millis/300
+      tick: dora/timer/millis/100
     outputs:
       - text
       - tick

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):
@@ -101,6 +110,16 @@ def main():
         event_type = event["type"]
 
         if event_type == "INPUT":
+
+            # pylint: disable=fixme
+            # TODO: Remove this after https://github.com/dora-rs/dora/pull/652
+            while True:
+                next_event = node.next(timeout=0.001)
+                if next_event is not None and next_event["type"] == "INPUT":
+                    event = next_event
+                else:
+                    break
+
             event_id = event["id"]
 
             if "image" in event_id:

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