Update app.py

app.py

@@ -1,34 +1,40 @@
 import gradio as gr
 import spaces
-from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
+from transformers import Qwen2VLForConditionalGeneration, AutoProcessor, TextIteratorStreamer
 from qwen_vl_utils import process_vision_info
 import torch
 from PIL import Image
 import subprocess
 import numpy as np
 import os
-
-# Install flash-attn
-subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
+from threading import Thread
 
 # Model and Processor Loading (Done once at startup)
 MODEL_ID = "Qwen/Qwen2-VL-2B-Instruct"
-model = Qwen2VLForConditionalGeneration.from_pretrained(MODEL_ID, attn_implementation="flash_attention_2", trust_remote_code=True, torch_dtype=torch.float16).to("cuda").eval()
+model = Qwen2VLForConditionalGeneration.from_pretrained(
+    MODEL_ID,
+    trust_remote_code=True,
+    torch_dtype=torch.float16
+).to("cuda").eval()
 processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
 
 DESCRIPTION = "[Qwen2-VL-2B Demo](https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct)"
 
+image_extensions = Image.registered_extensions()
+video_extensions = ("avi", "mp4", "mov", "mkv", "flv", "wmv", "mjpeg", "wav", "gif", "webm", "m4v", "3gp")
+
+
 @spaces.GPU
 def qwen_inference(media_path, text_input=None):
-
-    image_extensions = Image.registered_extensions()
     if media_path.endswith(tuple([i for i, f in image_extensions.items()])):
         media_type = "image"
-    elif media_path.endswith(("avi", "mp4", "mov", "mkv", "flv", "wmv", "mjpeg", "wav", "gif", "webm", "m4v", "3gp")):  # Check if it's a video path
+    elif media_path.endswith(video_extensions):  # Check if it's a video path
         media_type = "video"
     else:
-        raise ValueError("Unsupported media type. Please upload an image or video.")
-    
+        raise ValueError(
+            "Unsupported media type. Please upload an image or video."
+        )
+
     messages = [
         {
             "role": "user",
@@ -36,15 +42,17 @@ def qwen_inference(media_path, text_input=None):
                 {
                     "type": media_type,
                     media_type: media_path,
-                    **({"fps": 8.0} if media_type == "video" else {}), 
+                    **({"fps": 8.0} if media_type == "video" else {}),
                 },
                 {"type": "text", "text": text_input},
             ],
         }
     ]
 
-    text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
-    image_inputs, video_inputs = process_vision_info(messages) 
+    text = processor.apply_chat_template(
+        messages, tokenize=False, add_generation_prompt=True
+    )
+    image_inputs, video_inputs = process_vision_info(messages)
     inputs = processor(
         text=[text],
         images=image_inputs,
@@ -53,11 +61,19 @@ def qwen_inference(media_path, text_input=None):
         return_tensors="pt",
     ).to("cuda")
 
-    generated_ids = model.generate(**inputs, max_new_tokens=1024)
-    generated_ids_trimmed = [out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
-    output_text = processor.batch_decode(generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
-    
-    return output_text
+    streamer = TextIteratorStreamer(
+        processor, skip_prompt=True, **{"skip_special_tokens": True}
+    )
+    generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=1024)
+
+    thread = Thread(target=model.generate, kwargs=generation_kwargs)
+    thread.start()
+
+    buffer = ""
+    for new_text in streamer:
+        buffer += new_text
+        yield buffer
+
 
 css = """
   #output {
@@ -73,12 +89,16 @@ with gr.Blocks(css=css) as demo:
     with gr.Tab(label="Image/Video Input"):
         with gr.Row():
             with gr.Column():
-                input_media = gr.File(label="Upload Image or Video", type="filepath") 
+                input_media = gr.File(
+                    label="Upload Image or Video", type="filepath"
+                )
                 text_input = gr.Textbox(label="Question")
                 submit_btn = gr.Button(value="Submit")
             with gr.Column():
                 output_text = gr.Textbox(label="Output Text")
 
-        submit_btn.click(qwen_inference, [input_media, text_input], [output_text])
+        submit_btn.click(
+            qwen_inference, [input_media, text_input], [output_text]
+        )
 
 demo.launch(debug=True)