Update app.py

app.py

@@ -1,88 +1,168 @@
-import os
-import gradio as gr
-import copy
+import spaces
+import json
+import subprocess
 from llama_cpp import Llama
-from huggingface_hub import hf_hub_download  
+from llama_cpp_agent import LlamaCppAgent, MessagesFormatterType
+from llama_cpp_agent.providers import LlamaCppPythonProvider
+from llama_cpp_agent.chat_history import BasicChatHistory
+from llama_cpp_agent.chat_history.messages import Roles
+import gradio as gr
+from huggingface_hub import hf_hub_download
 
+llm = None
+llm_model = None
 
-llm = Llama(
-    model_path=hf_hub_download(
-        repo_id=os.environ.get("REPO_ID", "microsoft/Phi-3-mini-4k-instruct-gguf"),
-        filename=os.environ.get("MODEL_FILE", "Phi-3-mini-4k-instruct-q4.gguf"),
-    ),
-    n_ctx=2048,
-    n_gpu_layers=50, # change n_gpu_layers if you have more or less VRAM 
-) 
+# Download the new model
+hf_hub_download(
+    repo_id="hugging-quants/Llama-3.2-1B-Instruct-Q4_K_M-GGUF",
+    filename="llama-3.2-1b-instruct-q4_k_m.gguf",
+    local_dir="./models"
+)
 
+def get_messages_formatter_type(model_name):
+    if "Llama" in model_name:
+        return MessagesFormatterType.LLAMA_3
+    else:
+        raise ValueError(f"Unsupported model: {model_name}")
 
-def generate_text(
+@spaces.GPU
+def respond(
     message,
     history: list[tuple[str, str]],
+    model,
     system_message,
     max_tokens,
     temperature,
     top_p,
+    top_k,
+    repeat_penalty,
 ):
-    temp = ""
-    input_prompt = f"[INST] <<SYS>>\n{system_message}\n<</SYS>>\n\n "
-    for interaction in history:
-        input_prompt = input_prompt + str(interaction[0]) + " [/INST] " + str(interaction[1]) + " </s><s> [INST] "
+    global llm
+    global llm_model
+    
+    chat_template = get_messages_formatter_type(model)
+    
+    if llm is None or llm_model != model:
+        llm = Llama(
+            model_path=f"models/{model}",
+            flash_attn=True,
+            n_gpu_layers=81,
+            n_batch=1024,
+            n_ctx=8192,
+        )
+        llm_model = model
+    
+    provider = LlamaCppPythonProvider(llm)
+
+    agent = LlamaCppAgent(
+        provider,
+        system_prompt=f"{system_message}",
+        predefined_messages_formatter_type=chat_template,
+        debug_output=True
+    )
+    
+    settings = provider.get_provider_default_settings()
+    settings.temperature = temperature
+    settings.top_k = top_k
+    settings.top_p = top_p
+    settings.max_tokens = max_tokens
+    settings.repeat_penalty = repeat_penalty
+    settings.stream = True
 
-    input_prompt = input_prompt + str(message) + " [/INST] "
+    messages = BasicChatHistory()
 
-    output = llm(
-        input_prompt,
-        temperature=temperature,
-        top_p=top_p,
-        top_k=40,
-        repeat_penalty=1.1,
-        max_tokens=max_tokens,
-        stop=[
-            "<|prompter|>",
-            "<|endoftext|>",
-            "<|endoftext|> \n",
-            "ASSISTANT:",
-            "USER:",
-            "SYSTEM:",
-        ],
-        stream=True,
+    for msn in history:
+        user = {
+            'role': Roles.user,
+            'content': msn[0]
+        }
+        assistant = {
+            'role': Roles.assistant,
+            'content': msn[1]
+        }
+        messages.add_message(user)
+        messages.add_message(assistant)
+    
+    stream = agent.get_chat_response(
+        message,
+        llm_sampling_settings=settings,
+        chat_history=messages,
+        returns_streaming_generator=True,
+        print_output=False
     )
-    for out in output:
-        stream = copy.deepcopy(out)
-        temp += stream["choices"][0]["text"]
-        yield temp
+    
+    outputs = ""
+    for output in stream:
+        outputs += output
+        yield outputs
+
+description = """<p><center>
+<a href="https://huggingface.co/hugging-quants/Llama-3.2-1B-Instruct-Q4_K_M-GGUF" target="_blank">[Meta Llama 3.2 (1B)]</a>
+
+Meta Llama 3.2 (1B) is a multilingual large language model (LLM) optimized for conversational dialogue use cases, including agentic retrieval and summarization tasks. It outperforms many open-source and closed chat models on industry benchmarks, and is intended for commercial and research use in multiple languages.
 
+</center></p>
+"""
 
 demo = gr.ChatInterface(
-    generate_text,
-    title="llama-cpp-python on GPU",
-    description="Running LLM with https://github.com/abetlen/llama-cpp-python",
-    examples=[
-        ['How to setup a human base on Mars? Give short answer.'],
-        ['Explain theory of relativity to me like I’m 8 years old.'],
-        ['What is 9,000 * 9,000?'],
-        ['Write a pun-filled happy birthday message to my friend Alex.'],
-        ['Justify why a penguin might make a good king of the jungle.']
-    ],
-    cache_examples=False,
-    retry_btn=None,
-    undo_btn="Delete Previous",
-    clear_btn="Clear",
+    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.Dropdown([
+                "llama-3.2-1b-instruct-q4_k_m.gguf"
+            ],
+            value="llama-3.2-1b-instruct-q4_k_m.gguf",
+            label="Model"
+        ),
+        gr.Textbox(value="You are a world-class AI system, capable of complex reasoning and reflection. Reason through the query inside <thinking> tags, and then provide your final response inside <output> tags. If you detect that you made a mistake in your reasoning at any point, correct yourself inside <reflection> tags.", label="System message"),
+        gr.Slider(minimum=1, maximum=8192, value=2048, step=1, label="Max 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)",
+            label="Top-p",
+        ),
+        gr.Slider(
+            minimum=0,
+            maximum=100,
+            value=40,
+            step=1,
+            label="Top-k",
+        ),
+        gr.Slider(
+            minimum=0.0,
+            maximum=2.0,
+            value=1.1,
+            step=0.1,
+            label="Repetition penalty",
         ),
     ],
+    theme=gr.themes.Soft(primary_hue="violet", secondary_hue="violet", neutral_hue="gray",font=[gr.themes.GoogleFont("Exo"), "ui-sans-serif", "system-ui", "sans-serif"]).set(
+        body_background_fill_dark="#16141c",
+        block_background_fill_dark="#16141c",
+        block_border_width="1px",
+        block_title_background_fill_dark="#1e1c26",
+        input_background_fill_dark="#292733",
+        button_secondary_background_fill_dark="#24212b",
+        border_color_accent_dark="#343140",
+        border_color_primary_dark="#343140",
+        background_fill_secondary_dark="#16141c",
+        color_accent_soft_dark="transparent",
+        code_background_fill_dark="#292733",
+    ),
+    retry_btn="Retry",
+    undo_btn="Undo",
+    clear_btn="Clear",
+    submit_btn="Send",
+    title="Meta Llama 3.2 (1B)",
+    description=description,
+    chatbot=gr.Chatbot(
+        scale=1, 
+        likeable=False,
+        show_copy_button=True
+    )
 )
 
-
 if __name__ == "__main__":
-    demo.launch()
-
+    demo.launch()