Add VideoLLaMA3 interface

app.py

@@ -1,64 +1,176 @@
+import os
+import os.path as osp
+
 import gradio as gr
-from huggingface_hub import InferenceClient
+import spaces
+import torch
+from threading import Thread
+from transformers import AutoModelForCausalLM, AutoProcessor, TextIteratorStreamer
 
+HEADER = """
 """
-For more information on `huggingface_hub` Inference API support, please check the docs: https://huggingface.co/docs/huggingface_hub/v0.22.2/en/guides/inference
-"""
-client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")
 
 
-def respond(
-    message,
-    history: list[tuple[str, str]],
-    system_message,
-    max_tokens,
-    temperature,
-    top_p,
-):
-    messages = [{"role": "system", "content": system_message}]
+class VideoLLaMA3GradioInterface(object):
 
-    for val in history:
-        if val[0]:
-            messages.append({"role": "user", "content": val[0]})
-        if val[1]:
-            messages.append({"role": "assistant", "content": val[1]})
+    def __init__(self, model_name, device="cpu", example_dir=None, **server_kwargs):
+        self.device = device
+        self.model = AutoModelForCausalLM.from_pretrained(
+            model_name,
+            trust_remote_code=True,
+            torch_dtype=torch.bfloat16,
+            attn_implementation="flash_attention_2",
+        )
+        self.model.to(self.device)
+        self.processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)
 
-    messages.append({"role": "user", "content": message})
+        self.server_kwargs = server_kwargs
+        
+        self.image_formats = ("png", "jpg", "jpeg")
+        self.video_formats = ("mp4",)
 
-    response = ""
+        image_examples, video_examples = [], []
+        if example_dir is not None:
+            example_files = [
+                osp.join(example_dir, f) for f in os.listdir(example_dir)
+            ]
+            for example_file in example_files:
+                if example_file.endswith(self.image_formats):
+                    image_examples.append([example_file])
+                elif example_file.endswith(self.video_formats):
+                    video_examples.append([example_file])
 
-    for message in client.chat_completion(
-        messages,
-        max_tokens=max_tokens,
-        stream=True,
-        temperature=temperature,
-        top_p=top_p,
-    ):
-        token = message.choices[0].delta.content
+        with gr.Blocks() as self.interface:
+            gr.Markdown(HEADER)
+            with gr.Row():
+                chatbot = gr.Chatbot(type="messages", elem_id="chatbot", height=710)
 
-        response += token
-        yield response
+                with gr.Column():
+                    with gr.Tab(label="Input"):
 
+                        with gr.Row():
+                            input_video = gr.Video(sources=["upload"], label="Upload Video")
+                            input_image = gr.Image(sources=["upload"], type="filepath", label="Upload Image")
 
-"""
-For information on how to customize the ChatInterface, peruse the gradio docs: https://www.gradio.app/docs/chatinterface
-"""
-demo = gr.ChatInterface(
-    respond,
-    additional_inputs=[
-        gr.Textbox(value="You are a friendly Chatbot.", label="System message"),
-        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
-        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
-        gr.Slider(
-            minimum=0.1,
-            maximum=1.0,
-            value=0.95,
-            step=0.05,
-            label="Top-p (nucleus sampling)",
-        ),
-    ],
-)
+                        if len(image_examples):
+                            gr.Examples(image_examples, inputs=[input_image], label="Example Images")
+                        if len(video_examples):
+                            gr.Examples(video_examples, inputs=[input_video], label="Example Videos")
+
+                        input_text = gr.Textbox(label="Input Text", placeholder="Type your message here and press enter to submit")
+
+                        submit_button = gr.Button("Generate")
+
+                    with gr.Tab(label="Configure"):
+                        with gr.Accordion("Generation Config", open=True):
+                            do_sample = gr.Checkbox(value=True, label="Do Sample")
+                            temperature = gr.Slider(minimum=0.0, maximum=1.0, value=0.2, label="Temperature")
+                            top_p = gr.Slider(minimum=0.0, maximum=1.0, value=0.9, label="Top P")
+                            max_new_tokens = gr.Slider(minimum=0, maximum=4096, value=2048, step=1, label="Max New Tokens")
+
+                        with gr.Accordion("Video Config", open=True):
+                            fps = gr.Slider(minimum=0.0, maximum=10.0, value=1, label="FPS")
+                            max_frames = gr.Slider(minimum=0, maximum=256, value=180, step=1, label="Max Frames")
+
+            input_video.change(self._on_video_upload, [chatbot, input_video], [chatbot, input_video])
+            input_image.change(self._on_image_upload, [chatbot, input_image], [chatbot, input_image])
+            input_text.submit(self._on_text_submit, [chatbot, input_text], [chatbot, input_text])
+            submit_button.click(
+                self._predict,
+                [
+                    chatbot, input_text, do_sample, temperature, top_p, max_new_tokens,
+                    fps, max_frames
+                ],
+                [chatbot],
+            )
+
+    def _on_video_upload(self, messages, video):
+        if video is not None:
+            # messages.append({"role": "user", "content": gr.Video(video)})
+            messages.append({"role": "user", "content": {"path": video}})
+        return messages, None
+
+    def _on_image_upload(self, messages, image):
+        if image is not None:
+            # messages.append({"role": "user", "content": gr.Image(image)})
+            messages.append({"role": "user", "content": {"path": image}})
+        return messages, None
+
+    def _on_text_submit(self, messages, text):
+        messages.append({"role": "user", "content": text})
+        return messages, ""
+
+    @spaces.GPU(duration=120)
+    def _predict(self, messages, input_text, do_sample, temperature, top_p, max_new_tokens,
+                 fps, max_frames):
+        if len(input_text) > 0:
+            messages.append({"role": "user", "content": input_text})
+        new_messages = []
+        contents = []
+        for message in messages:
+            if message["role"] == "assistant":
+                if len(contents):
+                    new_messages.append({"role": "user", "content": contents})
+                    contents = []
+                new_messages.append(message)
+            elif message["role"] == "user":
+                if isinstance(message["content"], str):
+                    contents.append(message["content"])
+                else:
+                    media_path = message["content"][0]
+                    if media_path.endswith(self.video_formats):
+                        contents.append({"type": "video", "video": {"video_path": media_path, "fps": fps, "max_frames": max_frames}})
+                    elif media_path.endswith(self.image_formats):
+                        contents.append({"type": "image", "image": {"image_path": media_path}})
+                    else:
+                        raise ValueError(f"Unsupported media type: {media_path}")
+
+        if len(contents):
+            new_messages.append({"role": "user", "content": contents})
+
+        if len(new_messages) == 0 or new_messages[-1]["role"] != "user":
+            return messages
+
+        generation_config = {
+            "do_sample": do_sample,
+            "temperature": temperature,
+            "top_p": top_p,
+            "max_new_tokens": max_new_tokens
+        }
+
+        inputs = self.processor(
+            conversation=new_messages,
+            add_system_prompt=True,
+            add_generation_prompt=True,
+            return_tensors="pt"
+        )
+        inputs = {k: v.to(self.device) if isinstance(v, torch.Tensor) else v for k, v in inputs.items()}
+        if "pixel_values" in inputs:
+            inputs["pixel_values"] = inputs["pixel_values"].to(torch.bfloat16)
+
+        streamer = TextIteratorStreamer(self.processor.tokenizer, skip_prompt=True, skip_special_tokens=True)
+        generation_kwargs = {
+            **inputs,
+            **generation_config,
+            "streamer": streamer,
+        }
+
+        thread = Thread(target=self.model.generate, kwargs=generation_kwargs)
+        thread.start()
+
+        messages.append({"role": "assistant", "content": ""})
+        for token in streamer:
+            messages[-1]['content'] += token
+            yield messages
+
+    def launch(self):
+        self.interface.launch(**self.server_kwargs)
 
 
 if __name__ == "__main__":
-    demo.launch()
+    interface = VideoLLaMA3GradioInterface(
+        model_name="DAMO-NLP-SG/VideoLLaMA3-7B",
+        device="cuda",
+        example_dir="./examples",
+    )
+    interface.launch()