app.py

import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import json

BASE_MODEL = "mistralai/Mistral-7B-Instruct-v0.2"
ADAPTER    = "AmadouDiarouga/telma-mistral-telecom-gn"

print("Chargement tokenizer...")
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
tokenizer.pad_token    = tokenizer.eos_token
tokenizer.padding_side = "right"

print("Chargement modele base...")
base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    torch_dtype=torch.float32,
    device_map="cpu",
    low_cpu_mem_usage=True,
)

print("Chargement adapteur LoRA...")
model = PeftModel.from_pretrained(base_model, ADAPTER)
model.eval()
print("Modele pret !")

INSTRUCTION_FR = (
    "Tu es TELMA, un assistant vocal IA pour les telecoms en Guinee. "
    "Tu aides les clients d'Orange Guinee, MTN et Celcom. "
    "Reponds en francais simple, court et direct. "
    "Cite toujours le code USSD ou le numero quand c'est pertinent."
)

INSTRUCTION_PL = (
    "Be njahii e on ngondi e telecoms Guinee. "
    "Jaabii dow e Pulaar, yitere e laawol."
)

def repondre(question, langue="fr"):
    if not question.strip():
        return json.dumps({"error": "question vide"})

    instruction = INSTRUCTION_PL if langue == "poular" else INSTRUCTION_FR
    prompt = f"<s>[INST] {instruction}\n\n{question.strip()} [/INST]"

    inputs = tokenizer(prompt, return_tensors="pt")
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=200,
            temperature=0.3,
            do_sample=True,
            pad_token_id=tokenizer.eos_token_id,
            repetition_penalty=1.1,
        )

    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    reponse  = response.split("[/INST]")[-1].strip()

    return json.dumps({
        "question": question,
        "reponse": reponse,
        "langue": langue,
        "status": "ok"
    }, ensure_ascii=False)

demo = gr.Interface(
    fn=repondre,
    inputs=[
        gr.Textbox(label="Question", placeholder="Comment verifier mon solde Orange ?"),
        gr.Dropdown(choices=["fr", "poular"], value="fr", label="Langue"),
    ],
    outputs=gr.Textbox(label="Reponse JSON"),
    title="TELMA - Assistant Telecoms Guinee",
    description="API vocale pour Orange, MTN et Celcom en Guinee",
    examples=[
        ["Comment verifier mon solde Orange ?", "fr"],
        ["Je veux internet MTN pour une semaine", "fr"],
        ["Hol balansi Orange am ?", "poular"],
    ],
)

demo.launch(server_name="0.0.0.0", server_port=7860)