Update app.py

app.py

@@ -1,21 +1,23 @@
 import gradio as gr
-import cv2
-import numpy as np
-import time
-import torch
-import spaces
-from threading import Thread
-from PIL import Image
 from transformers import (
     AutoProcessor,
     Qwen2_5_VLForConditionalGeneration,
     TextIteratorStreamer,
-    AutoTokenizer,
     AutoModelForCausalLM,
+    AutoTokenizer,
 )
 from transformers.image_utils import load_image
+from threading import Thread
+import time
+import torch
+import spaces
+import cv2
+import numpy as np
+from PIL import Image
 
+# -----------------------
 # Progress Bar Helper
+# -----------------------
 def progress_bar_html(label: str) -> str:
     """
     Returns an HTML snippet for a thin progress bar with a label.
@@ -36,7 +38,9 @@ def progress_bar_html(label: str) -> str:
 </style>
     '''
 
-# Video Downsampling Helper
+# -----------------------
+# Video Processing Helper
+# -----------------------
 def downsample_video(video_path):
     """
     Downsamples the video to 10 evenly spaced frames.
@@ -62,7 +66,9 @@ def downsample_video(video_path):
     vidcap.release()
     return frames
 
-# Qwen2.5-VL Setup (for image and video understanding)
+# -----------------------
+# Qwen2.5-VL Model (Multimodal)
+# -----------------------
 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(
@@ -71,8 +77,10 @@ vl_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
     torch_dtype=torch.bfloat16
 ).to("cuda").eval()
 
-# Text Generation Setup (Ganymede)
-TG_MODEL_ID = "prithivMLmods/Ganymede-Llama-3.3-3B-Preview"
+# -----------------------
+# Text Generation Setup (DeepHermes)
+# -----------------------
+TG_MODEL_ID = "prithivMLmods/DeepHermes-3-Llama-3-3B-Preview-abliterated"
 tg_tokenizer = AutoTokenizer.from_pretrained(TG_MODEL_ID)
 tg_model = AutoModelForCausalLM.from_pretrained(
     TG_MODEL_ID,
@@ -81,38 +89,37 @@ tg_model = AutoModelForCausalLM.from_pretrained(
 )
 tg_model.eval()
 
+# -----------------------
+# Main Inference Function
+# -----------------------
 @spaces.GPU
 def model_inference(input_dict, history):
     text = input_dict["text"]
-    files = input_dict.get("files", [])
+    files = input_dict["files"]
 
-    # Video inference branch using a tag @video-infer
+    # Video inference branch
     if text.strip().lower().startswith("@video-infer"):
-        # Remove the tag from the query.
         text = text[len("@video-infer"):].strip()
         if not files:
-            gr.Error("Please upload a video file along with your @video-infer query.")
+            yield gr.Error("Please upload a video file along with your @video-infer query.")
             return
-        # Assume the first file is a video.
         video_path = files[0]
         frames = downsample_video(video_path)
         if not frames:
-            gr.Error("Could not process video.")
+            yield gr.Error("Could not process video.")
             return
-        # Build messages: start with the text prompt.
+        # Build messages starting with the text prompt and then add each frame with its timestamp.
         messages = [
             {
                 "role": "user",
                 "content": [{"type": "text", "text": text}]
             }
         ]
-        # Append each frame with a timestamp label.
         for image, timestamp in frames:
             messages[0]["content"].append({"type": "text", "text": f"Frame {timestamp}:"})
             messages[0]["content"].append({"type": "image", "image": image})
-        # Collect only the images from the frames.
+        # Collect images from the frames.
         video_images = [image for image, _ in frames]
-        # Prepare the prompt.
         prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
         inputs = processor(
             text=[prompt],
@@ -120,7 +127,6 @@ def model_inference(input_dict, history):
             return_tensors="pt",
             padding=True,
         ).to("cuda")
-        # 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=vl_model.generate, kwargs=generation_kwargs)
@@ -133,12 +139,20 @@ def model_inference(input_dict, history):
             yield buffer
         return
 
-    # If files are provided (e.g. images), use the VL model.
+    # Multimodal branch if images are provided (non-video)
     if files:
+        # If more than one file is provided, load them as images.
         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 == "":
+            yield gr.Error("Please input a text query along with the image(s).")
+            return
+
         messages = [
             {
                 "role": "user",
@@ -167,34 +181,37 @@ def model_inference(input_dict, history):
             yield buffer
         return
 
-    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,
-        }
-        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
+    # Text-only branch using DeepHermes text generation.
+    if text.strip() == "":
+        yield gr.Error("Please input a query.")
         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).")
+    input_ids = tg_tokenizer(text, return_tensors="pt").to(tg_model.device)
+    streamer = TextIteratorStreamer(tg_tokenizer, skip_prompt=True, skip_special_tokens=True)
+    generation_kwargs = {
+        "input_ids": input_ids,
+        "streamer": streamer,
+        "max_new_tokens": 2048,
+        "do_sample": True,
+        "top_p": 0.9,
+        "top_k": 50,
+        "temperature": 0.6,
+        "repetition_penalty": 1.2,
+    }
+    thread = Thread(target=tg_model.generate, kwargs=generation_kwargs)
+    thread.start()
+    buffer = ""
+    yield progress_bar_html("Processing text with DeepHermes Model")
+    for new_text in streamer:
+        buffer += new_text
+        time.sleep(0.01)
+        yield buffer
 
-# Gradio Chat Interface Setup
+# -----------------------
+# Gradio Chat Interface
+# -----------------------
 examples = [
-    [{"text": "Explain the image and highlight the key points.", "files": ["example_images/campeones.jpg"]}],
+    [{"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"]}],