Update app.py

app.py

@@ -1,14 +1,7 @@
-import os
-import subprocess
-import time
-import requests
-import json
-import zipfile
-import tarfile
-import stat
 import gradio as gr
 import spaces
 from huggingface_hub import hf_hub_download
+from llama_cpp import Llama
 
 # 1. DUMMY GPU FUNCTION: satisfies the Hugging Face startup checker 
 @spaces.GPU(duration=5)
@@ -19,168 +12,53 @@ def dummy_gpu():
 # Automatically execute it once right away
 dummy_gpu()
 
-# 2. DOWNLOAD MODEL
-def setup_model():
-    print("Downloading model...")
-    repo = "Abiray/MiniCPM5-1B-GGUF"
-    filename = "minicpm5-1b-Q6_K.gguf"
-    model_path = hf_hub_download(repo_id=repo, filename=filename)
-    print(f"Model downloaded to: {model_path}")
-    return model_path
+# 2. LOAD MODEL DIRECTLY IN PYTHON
+print("Downloading model...")
+model_path = hf_hub_download(
+    repo_id="Abiray/MiniCPM5-1B-GGUF", 
+    filename="minicpm5-1b-Q6_K.gguf"
+)
 
-# 3. DOWNLOAD PRE-BUILT LLAMA.CPP BINARY
-def get_prebuilt_llama():
-    server_bin = "./llama-server"
-    if os.path.exists(server_bin):
-        print("Pre-built llama-server already exists.")
-        return server_bin
-        
-    print("Fetching latest pre-built llama.cpp release...")
-    api_url = "https://api.github.com/repos/ggml-org/llama.cpp/releases/latest"
-    
-    try:
-        resp = requests.get(api_url).json()
-        target_asset = None
-        # Support both .zip and .tar.gz (Linux recently migrated to tar.gz)
-        for asset in resp.get("assets", []):
-            name = asset["name"].lower()
-            if "ubuntu-x64" in name and (name.endswith(".zip") or name.endswith(".tar.gz")) and not any(x in name for x in ["vulkan", "rocm", "sycl", "openvino", "cuda"]):
-                target_asset = asset
-                break
-                
-        if not target_asset:
-            raise ValueError("No matching CPU asset found in latest release.")
-            
-        download_url = target_asset["browser_download_url"]
-        archive_name = target_asset["name"]
-        
-    except Exception as e:
-        print(f"GitHub API check failed ({e}), using fallback direct URL...")
-        # Updated to a highly recent July 2026 build that supports MiniCPM architecture natively
-        download_url = "https://github.com/ggml-org/llama.cpp/releases/download/b9940/llama-b9940-bin-ubuntu-x64.tar.gz"
-        archive_name = "llama-fallback.tar.gz"
-
-    print(f"Downloading {archive_name} from {download_url}...")
-    with requests.get(download_url, stream=True) as r:
-        r.raise_for_status()
-        with open(archive_name, "wb") as f:
-            for chunk in r.iter_content(chunk_size=8192):
-                f.write(chunk)
-                
-    print("Extracting archive...")
-    if archive_name.endswith(".zip"):
-        with zipfile.ZipFile(archive_name, 'r') as zip_ref:
-            zip_ref.extractall("./llama_extracted")
-    elif archive_name.endswith(".tar.gz"):
-        with tarfile.open(archive_name, "r:gz") as tar_ref:
-            tar_ref.extractall("./llama_extracted")
-        
-    # Locate the binary inside the extracted folder
-    found_bin = None
-    for root, dirs, files in os.walk("./llama_extracted"):
-        if "llama-server" in files:
-            found_bin = os.path.join(root, "llama-server")
-            break
-            
-    if not found_bin:
-        raise RuntimeError("llama-server binary not found in the extracted files!")
-        
-    # Move it to root and clean up
-    os.rename(found_bin, server_bin)
-    
-    # Make it executable (chmod +x)
-    st = os.stat(server_bin)
-    os.chmod(server_bin, st.st_mode | stat.S_IEXEC)
-    
-    print("Pre-built llama-server is ready!")
-    return server_bin
-
-# 4. START BACKGROUND SERVER
-def start_server(model_path, server_bin_path):
-    cmd = [
-        server_bin_path,
-        "-m", model_path,
-        "--host", "127.0.0.1",
-        "--port", "8080",
-        "-t", "2",
-        "--cache-type-k", "q8_0",
-        "--cache-type-v", "q8_0",
-        "-c", "8192", 
-        "-n", "4096"
-    ]
-    
-    print("Booting local llama-server with command:")
-    print(" ".join(cmd))
-    
-    server_process = subprocess.Popen(cmd)
-    
-    print("Waiting for llama-server to initialize...")
-    for _ in range(120):
-        # Fail-fast check
-        if server_process.poll() is not None:
-            raise RuntimeError(f"llama-server crashed instantly with return code {server_process.returncode}! Check the C++ logs above for the reason.")
-            
-        try:
-            response = requests.get("http://127.0.0.1:8080/health")
-            if response.status_code == 200:
-                print("llama-server is up and running!")
-                return server_process
-        except requests.exceptions.ConnectionError:
-            time.sleep(2)
-            
-    raise RuntimeError("Server failed to start within the timeout period.")
+print("Loading model into memory via llama-cpp-python...")
+llm = Llama(
+    model_path=model_path,
+    n_ctx=8192,       # Safe 8k context limit to avoid RAM crashes
+    n_threads=2,      # Perfectly matches Hugging Face standard CPU cores
+    verbose=False     # Keeps the terminal clean
+)
+print("Model loaded successfully!")
 
-# Execute initialization
-model_filepath = setup_model()
-llama_executable = get_prebuilt_llama()
-start_server(model_filepath, llama_executable)
-
-# 5. GRADIO UI TO INTERACT WITH LOCAL SERVER
+# 3. CHAT FUNCTION
 def chat_with_llama(message, history):
+    # Format the conversation history
     messages = [{"role": "system", "content": "You are a helpful AI assistant."}]
+    
     for user_msg, assistant_msg in history:
         messages.append({"role": "user", "content": user_msg})
         messages.append({"role": "assistant", "content": assistant_msg})
-    
+        
     messages.append({"role": "user", "content": message})
     
-    payload = {
-        "messages": messages,
-        "stream": True,
-        "temperature": 0.7
-    }
+    # Generate the response in a stream
+    stream = llm.create_chat_completion(
+        messages=messages,
+        stream=True,
+        temperature=0.7,
+        max_tokens=1024
+    )
     
-    try:
-        response = requests.post(
-            "http://127.0.0.1:8080/v1/chat/completions",
-            json=payload,
-            stream=True
-        )
-        
-        partial_response = ""
-        for line in response.iter_lines():
-            if line:
-                decoded_line = line.decode('utf-8')
-                if decoded_line.startswith("data: "):
-                    data_str = decoded_line[6:]
-                    if data_str == "[DONE]":
-                        break
-                    try:
-                        data = json.loads(data_str)
-                        if "choices" in data and len(data["choices"]) > 0:
-                            delta = data["choices"][0].get("delta", {})
-                            if "content" in delta:
-                                partial_response += delta["content"]
-                                yield partial_response
-                    except json.JSONDecodeError:
-                        continue
-    except Exception as e:
-        yield f"Error communicating with local server: {str(e)}"
+    partial_response = ""
+    for chunk in stream:
+        if "choices" in chunk and len(chunk["choices"]) > 0:
+            delta = chunk["choices"][0].get("delta", {})
+            if "content" in delta:
+                partial_response += delta["content"]
+                yield partial_response
 
-# 6. LAUNCH GRADIO APP
+# 4. LAUNCH GRADIO APP
 with gr.Blocks() as demo:
-    gr.Markdown("# Native CPU Pre-Built `llama.cpp` on ZeroGPU Space")
-    gr.Markdown(f"Running **MiniCPM5-1B-GGUF** natively using the pre-compiled Ubuntu `llama-server`! Bypassing compilation and GPU logic entirely.")
+    gr.Markdown("# Native CPU `llama-cpp-python` on ZeroGPU Space")
+    gr.Markdown("Running **MiniCPM5-1B-GGUF** natively using Python bindings! No background servers, no zip downloads, and no compiling.")
     
     gr.ChatInterface(
         fn=chat_with_llama,