update

app.py

@@ -1,38 +1,105 @@
+import time
+from threading import Thread
+
 import gradio as gr
-from transformers import pipeline, BitsAndBytesConfig
 import torch
+from PIL import Image
+from transformers import AutoProcessor, LlavaForConditionalGeneration
+from transformers import TextIteratorStreamer
+
+import spaces
+
+
+PLACEHOLDER = """
+<div style="padding: 30px; text-align: center; display: flex; flex-direction: column; align-items: center;">
+   <img src="https://cdn-uploads.huggingface.co/production/uploads/64ccdc322e592905f922a06e/DDIW0kbWmdOQWwy4XMhwX.png" style="width: 80%; max-width: 550px; height: auto; opacity: 0.55;  "> 
+   <h1 style="font-size: 28px; margin-bottom: 2px; opacity: 0.55;">LLaVA-Llama-3-8B</h1>
+   <p style="font-size: 18px; margin-bottom: 2px; opacity: 0.65;">Llava-Llama-3-8b is a LLaVA model fine-tuned from Meta-Llama-3-8B-Instruct and CLIP-ViT-Large-patch14-336 with ShareGPT4V-PT and InternVL-SFT by XTuner</p>
+</div>
+"""
 
-quantization_config = BitsAndBytesConfig(
-    load_in_4bit=True,
-    bnb_4bit_compute_dtype=torch.bfloat16
-)
 
 model_id = "rogerxi/llava-finetune-test"
-pipe = pipeline("image-text-to-text", model=model_id, model_kwargs={"quantization_config": quantization_config})
 
-def update_conversation(new_message, history, image):
-    if image is None:
-        return "Please upload an image first using the widget on the left"
+processor = AutoProcessor.from_pretrained(model_id)
+
+model = LlavaForConditionalGeneration.from_pretrained(
+    model_id,
+    torch_dtype=torch.float16,
+    low_cpu_mem_usage=True,
+)
+
+model.to("cuda:0")
+model.generation_config.eos_token_id = 128009
+
+
+@spaces.GPU
+def bot_streaming(message, history):
+    print(message)
+    if message["files"]:
+        # message["files"][-1] is a Dict or just a string
+        if type(message["files"][-1]) == dict:
+            image = message["files"][-1]["path"]
+        else:
+            image = message["files"][-1]
+    else:
+        # if there's no image uploaded for this turn, look for images in the past turns
+        # kept inside tuples, take the last one
+        for hist in history:
+            if type(hist[0]) == tuple:
+                image = hist[0][0]
+    try:
+        if image is None:
+            # Handle the case where image is None
+            gr.Error("You need to upload an image for LLaVA to work.")
+    except NameError:
+        # Handle the case where 'image' is not defined at all
+        gr.Error("You need to upload an image for LLaVA to work.")
+
+    prompt = f"<|start_header_id|>user<|end_header_id|>\n\n<image>\n{message['text']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+    # print(f"prompt: {prompt}")
+    image = Image.open(image)
+    inputs = processor(prompt, image, return_tensors='pt').to(0, torch.float16)
 
-    messages = [{"role": "user", "content": "<image>"}]
+    streamer = TextIteratorStreamer(processor, **{"skip_special_tokens": False, "skip_prompt": True})
+    generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=1024, do_sample=False)
 
-    for user_text, assistant_text in history:
-        messages.append({"role": "user", "content": user_text})
-        messages.append({"role": "assistant", "content": assistant_text})
-    
-    messages.append({"role": "user", "content": new_message})
+    thread = Thread(target=model.generate, kwargs=generation_kwargs)
+    thread.start()
 
-    outputs = pipe(messages=messages, generate_kwargs={"max_new_tokens": 200})[0]['generated_text']
-    return outputs
+    text_prompt = f"<|start_header_id|>user<|end_header_id|>\n\n{message['text']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+    # print(f"text_prompt: {text_prompt}")
 
+    buffer = ""
+    time.sleep(0.5)
+    for new_text in streamer:
+        # find <|eot_id|> and remove it from the new_text
+        if "<|eot_id|>" in new_text:
+            new_text = new_text.split("<|eot_id|>")[0]
+        buffer += new_text
 
-with gr.Blocks() as demo:
+        # generated_text_without_prompt = buffer[len(text_prompt):]
+        generated_text_without_prompt = buffer
+        # print(generated_text_without_prompt)
+        time.sleep(0.06)
+        # print(f"new_text: {generated_text_without_prompt}")
+        yield generated_text_without_prompt
 
-    with gr.Row():
-      image = gr.Image(type='pil', interactive=True)
 
-      gr.ChatInterface(
-          update_conversation, additional_inputs=[image]
-      )
+chatbot=gr.Chatbot(placeholder=PLACEHOLDER,scale=1)
+chat_input = gr.MultimodalTextbox(interactive=True, file_types=["image"], placeholder="Enter message or upload file...", show_label=False)
+with gr.Blocks(fill_height=True, ) as demo:
+    gr.ChatInterface(
+    fn=bot_streaming,
+    title="LLaVA Llama-3-8B",
+    examples=[{"text": "What is on the flower?", "files": ["./bee.jpg"]},
+              {"text": "How to make this pastry?", "files": ["./baklava.png"]}],
+    description="Try [LLaVA Llama-3-8B](https://huggingface.co/xtuner/llava-llama-3-8b-v1_1-transformers). Upload an image and start chatting about it, or simply try one of the examples below. If you don't upload an image, you will receive an error.",
+    stop_btn="Stop Generation",
+    multimodal=True,
+    textbox=chat_input,
+    chatbot=chatbot,
+    )
 
-demo.launch()
+# demo.queue(api_open=False)
+demo.launch(debug=True)

requirements.txt

@@ -1,4 +1,5 @@
 transformers
 torch
 accelerate
-bitsandbytes
+bitsandbytes
+spaces