app.py

import gradio as gr
from huggingface_hub import hf_hub_download
from llama_cpp import Llama
import os

print("Lade DeepSeek-R1 GGUF...")

# Modell-Download
model_path = hf_hub_download(
    repo_id="unsloth/DeepSeek-R1-Distill-Qwen-1.5B-GGUF",
    filename="DeepSeek-R1-Distill-Qwen-1.5B-Q4_K_M.gguf"
)

# Initialisierung (n_ctx auf 1024 begrenzt für CPU-Stabilität)
llm = Llama(
    model_path=model_path,
    n_ctx=1024,
    n_threads=2,
    verbose=False
)

def respond(message, history):
    # DeepSeek Format: Wir starten direkt mit dem Denk-Tag
    prompt = f"User: {message}\nAssistant: <think>\n"
    
    response = ""
    # Streaming-Generierung
    stream = llm(
        prompt,
        max_tokens=512,
        stop=["User:", "<|endoftext|>"],
        stream=True,
        temperature=0.7
    )

    for chunk in stream:
        if "text" in chunk["choices"][0]:
            token = chunk["choices"][0]["text"]
            response += token
            yield response

# Gradio Interface
demo = gr.ChatInterface(
    fn=respond, 
    title="DeepSeek-R1 CPU (GGUF Safe Mode)",
    description="Läuft stabil auf Hugging Face CPU Hardware."
)

if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=7860)