Update app.py

app.py

@@ -1,14 +1,21 @@
 import gradio as gr
-from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration, TextIteratorStreamer
-from transformers.image_utils import load_image
-from threading import Thread
+import cv2
+import numpy as np
 import time
 import torch
 import spaces
-import cv2
-import numpy as np
+from threading import Thread
 from PIL import Image
+from transformers import (
+    AutoProcessor,
+    Qwen2_5_VLForConditionalGeneration,
+    TextIteratorStreamer,
+    AutoTokenizer,
+    AutoModelForCausalLM,
+)
+from transformers.image_utils import load_image
 
+# Progress Bar Helper
 def progress_bar_html(label: str) -> str:
     """
     Returns an HTML snippet for a thin progress bar with a label.
@@ -29,6 +36,7 @@ def progress_bar_html(label: str) -> str:
 </style>
     '''
 
+# Video Downsampling Helper
 def downsample_video(video_path):
     """
     Downsamples the video to 10 evenly spaced frames.
@@ -54,19 +62,31 @@ def downsample_video(video_path):
     vidcap.release()
     return frames
 
-MODEL_ID = "Qwen/Qwen2.5-VL-7B-Instruct"  # Alternatively: "Qwen/Qwen2.5-VL-3B-Instruct" 
-processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
-model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
-    MODEL_ID,
+# Qwen2.5-VL Setup (for image and video understanding)
+MODEL_ID_VL = "Qwen/Qwen2.5-VL-7B-Instruct"  # Alternatively: "Qwen/Qwen2.5-VL-3B-Instruct"
+processor = AutoProcessor.from_pretrained(MODEL_ID_VL, trust_remote_code=True)
+vl_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+    MODEL_ID_VL,
     trust_remote_code=True,
     torch_dtype=torch.bfloat16
 ).to("cuda").eval()
 
+# Text Generation Setup (Ganymede)
+TG_MODEL_ID = "prithivMLmods/Ganymede-Llama-3.3-3B-Preview"
+tg_tokenizer = AutoTokenizer.from_pretrained(TG_MODEL_ID)
+tg_model = AutoModelForCausalLM.from_pretrained(
+    TG_MODEL_ID,
+    device_map="auto",
+    torch_dtype=torch.bfloat16,
+)
+tg_model.eval()
+
 @spaces.GPU
 def model_inference(input_dict, history):
     text = input_dict["text"]
-    files = input_dict["files"]
+    files = input_dict.get("files", [])
 
+    # Video inference branch using a tag @video-infer
     if text.strip().lower().startswith("@video-infer"):
         # Remove the tag from the query.
         text = text[len("@video-infer"):].strip()
@@ -103,7 +123,7 @@ def model_inference(input_dict, history):
         # Set up streaming generation.
         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 = Thread(target=vl_model.generate, kwargs=generation_kwargs)
         thread.start()
         buffer = ""
         yield progress_bar_html("Processing video with Qwen2.5VL Model")
@@ -113,49 +133,69 @@ def model_inference(input_dict, history):
             yield buffer
         return
 
-    if len(files) > 1:
-        images = [load_image(image) for image in files]
-    elif len(files) == 1:
-        images = [load_image(files[0])]
-    else:
-        images = []
-
-    if text == "" and not images:
-        gr.Error("Please input a query and optionally image(s).")
-        return
-    if text == "" and images:
-        gr.Error("Please input a text query along with the image(s).")
+    # If files are provided (e.g. images), use the VL model.
+    if files:
+        if len(files) > 1:
+            images = [load_image(image) for image in files]
+        elif len(files) == 1:
+            images = [load_image(files[0])]
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    *[{"type": "image", "image": image} for image in images],
+                    {"type": "text", "text": text},
+                ],
+            }
+        ]
+        prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+        inputs = processor(
+            text=[prompt],
+            images=images,
+            return_tensors="pt",
+            padding=True,
+        ).to("cuda")
+        streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
+        generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=1024)
+        thread = Thread(target=vl_model.generate, kwargs=generation_kwargs)
+        thread.start()
+        buffer = ""
+        yield progress_bar_html("Processing with Qwen2.5VL Model")
+        for new_text in streamer:
+            buffer += new_text
+            time.sleep(0.01)
+            yield buffer
         return
 
-    messages = [
-        {
-            "role": "user",
-            "content": [
-                *[{"type": "image", "image": image} for image in images],
-                {"type": "text", "text": text},
-            ],
+    if text and not files:
+        # Prepare input for text generation.
+        input_ids = tg_tokenizer.encode(text, return_tensors="pt").to("cuda")
+        streamer = TextIteratorStreamer(tg_tokenizer, skip_prompt=True, skip_special_tokens=True)
+        generation_kwargs = {
+            "input_ids": input_ids,
+            "streamer": streamer,
+            "max_new_tokens": 1024,
+            "do_sample": True,
+            "temperature": 0.7,
+            "top_p": 0.9,
         }
-    ]
-    prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
-    inputs = processor(
-        text=[prompt],
-        images=images if images else None,
-        return_tensors="pt",
-        padding=True,
-    ).to("cuda")
-    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 = ""
-    yield progress_bar_html("Processing with Qwen2.5VL Model")
-    for new_text in streamer:
-        buffer += new_text
-        time.sleep(0.01)
-        yield buffer
+        thread = Thread(target=tg_model.generate, kwargs=generation_kwargs)
+        thread.start()
+        buffer = ""
+        yield progress_bar_html("Processing text with Ganymede Model")
+        for new_text in streamer:
+            buffer += new_text
+            time.sleep(0.01)
+            yield buffer
+        return
+
+    # Fallback error in case neither text nor proper file input is provided.
+    gr.Error("Please input a query (and optionally images or video for multimodal processing).")
 
+# Gradio Chat Interface Setup
 examples = [
     [{"text": "Describe the Image?", "files": ["example_images/document.jpg"]}],
+    [{"text": "Tell me a story about a brave knight."}],
     [{"text": "@video-infer Explain the content of the Advertisement", "files": ["example_images/videoplayback.mp4"]}],
     [{"text": "@video-infer Explain the content of the video in detail", "files": ["example_images/breakfast.mp4"]}],
     [{"text": "@video-infer Explain the content of the video.", "files": ["example_images/sky.mp4"]}],
@@ -172,4 +212,5 @@ demo = gr.ChatInterface(
     cache_examples=False,
 )
 
-demo.launch(debug=True)
+if __name__ == "__main__":
+    demo.launch(debug=True)