# app.py (ZeroGPU Version) # A production-quality, local, and uncensored text-editing agent # designed specifically for the Hugging Face ZeroGPU platform. import gradio as gr import pathlib import re import textwrap from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline from transformers.agents import Agent, Tool from fastmcp import FastMCP from huggingface_hub import snapshot_download # --- Hugging Face Spaces GPU Decorator --- # This is the key to making the app work on ZeroGPU from spaces import GPU as spaces_GPU # --- Configuration --- MODEL_ID = "NousResearch/Meta-Llama-3-8B-Instruct-GPTQ" ROOT = pathlib.Path("workspace") ROOT.mkdir(exist_ok=True) # --- 1. MCP Text-Editing Server (The "Tools" Backend) --- # This part remains the same. server = FastMCP("DocTools") # (All your @server.tool() functions: list_files, search_in_file, read_lines, patch_file go here) # ... [Paste your tool functions here to keep the script self-contained] ... @server.tool() def list_files(relative_path: str = ".") -> list[str]: """Lists all files and directories within a given subdirectory of the workspace.""" try: safe_path = (ROOT / relative_path).resolve() if not safe_path.is_relative_to(ROOT.resolve()): return ["Error: Access denied."] if not safe_path.exists(): return [f"Error: Directory '{relative_path}' not found."] return [p.name for p in safe_path.iterdir()] except Exception as e: return [f"An error occurred: {str(e)}"] @server.tool() def search_in_file(file_path: str, pattern: str, max_hits: int = 40) -> list[str]: """Searches for a regex pattern within a specified file in the workspace.""" try: safe_path = (ROOT / file_path).resolve() if not safe_path.is_relative_to(ROOT.resolve()): return ["Error: Access denied."] if not safe_path.is_file(): return [f"Error: File '{file_path}' not found."] output, regex = [], re.compile(pattern, re.IGNORECASE) with open(safe_path, 'r', encoding='utf-8') as f: for i, line in enumerate(f): if regex.search(line): output.append(f"{i+1}: {line.rstrip()}") if len(output) >= max_hits: break return output if output else ["No matches found."] except Exception as e: return [f"An error occurred: {str(e)}"] @server.tool() def read_lines(file_path: str, start_line: int, end_line: int) -> str: """Reads and returns a specific range of lines from a file.""" try: safe_path = (ROOT / file_path).resolve() if not safe_path.is_relative_to(ROOT.resolve()): return "Error: Access denied." if not safe_path.is_file(): return f"Error: File '{file_path}' not found." with open(safe_path, 'r', encoding='utf-8') as f: lines = f.readlines() return "".join(lines[max(0, start_line - 1):min(len(lines), end_line)]) except Exception as e: return f"An error occurred: {str(e)}"] @server.tool() def patch_file(file_path: str, start_line: int, end_line: int, new_content: str) -> str: """Replaces a range of lines in a file with new content.""" try: safe_path = (ROOT / file_path).resolve() if not safe_path.is_relative_to(ROOT.resolve()): return "Error: Access denied." if not safe_path.is_file(): return f"Error: File '{file_path}' not found." with open(safe_path, 'r', encoding='utf-8') as f: lines = f.readlines() new_lines = (lines[:max(0, start_line - 1)] + [line + '\n' for line in new_content.splitlines()] + lines[end_line:]) with open(safe_path, 'w', encoding='utf-8') as f: f.writelines(new_lines) return f"Success: Patched lines {start_line}-{end_line} in '{file_path}'." except Exception as e: return f"An error occurred: {str(e)}"] # --- 2. Agent and Model Loading (ZeroGPU compatible) --- # We initialize the agent as None. It will be created on the first user request. agent = None # This is our GPU-accelerated function. # It will load the model on the first run and cache it for subsequent calls. @spaces_GPU(duration=120) # Request GPU for 120 seconds per call def get_agent(): """ Loads and caches the LLM agent. This function runs on a GPU. """ global agent if agent is None: print("--- Loading model and agent for the first time ---") # Download the model to a persistent cache model_path = snapshot_download(MODEL_ID) tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True) model = AutoModelForCausalLM.from_pretrained( model_path, device_map="auto", torch_dtype="auto" # Recommended for modern GPUs ) llm_pipeline = pipeline( "text-generation", model=model, tokenizer=tokenizer, return_full_text=False, max_new_tokens=1024, ) tools = [Tool(name=t, description=server.get_schema(t)["description"], inputs=server.get_schema(t)["parameters"], function=lambda **kwargs, t=t: server.invoke(t, **kwargs)) for t in server.list_tools()] SYSTEM_PROMPT = textwrap.dedent(""" You are an expert technical editor. You must use your tools to answer the user's request. All file paths are relative to the '/workspace' directory. Always verify file contents with `read_lines` or `search_in_file` before patching. """) agent = Agent( llm_pipeline=llm_pipeline, tools=tools, system_prompt=SYSTEM_PROMPT, max_steps=10, memory=True ) print("--- Agent loaded successfully ---") return agent # --- 3. Gradio Chat Application --- # Using gr.ChatInterface for a cleaner UI setup with gr.Blocks(theme=gr.themes.Soft(), css="footer {display: none !important}") as demo: gr.Markdown("# ZeroGPU Text-Editing Agent 📝") gr.Markdown( """ Chat with this AI agent to perform complex edits on text documents in the workspace. **Note:** The first request will have a delay as the model is loaded onto the GPU. """ ) chatbot = gr.Chatbot(height=600, label="Agent Chat") async def chat_interaction(message, history): history.append([message, None]) yield "", history # Immediately show user message # 1. Get the agent (this triggers the GPU) current_agent = get_agent() # 2. Stream the response response = "" async for step in current_agent.astream(message): if isinstance(step, str): response = step history[-1][1] = response yield "", history gr.ChatInterface( fn=chat_interaction, chatbot=chatbot, fill_height=False ) # Add a sample file to the workspace for easy testing with open(ROOT / "sample.txt", "w") as f: f.write("This is a sample file for testing the ZeroGPU agent.") demo.queue().launch()