Update app.py
Browse files
app.py
CHANGED
|
@@ -5,17 +5,20 @@ from pydantic import BaseModel
|
|
| 5 |
from llama_cpp import Llama
|
| 6 |
from fastapi.middleware.cors import CORSMiddleware
|
| 7 |
|
| 8 |
-
|
| 9 |
-
|
|
|
|
| 10 |
|
| 11 |
-
print("--- Vérification
|
| 12 |
if not os.path.exists(MODEL_PATH):
|
| 13 |
-
print("
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
|
|
|
|
|
|
| 17 |
else:
|
| 18 |
-
print("Le modèle est
|
| 19 |
|
| 20 |
app = FastAPI()
|
| 21 |
|
|
@@ -27,10 +30,14 @@ app.add_middleware(
|
|
| 27 |
allow_headers=["*"],
|
| 28 |
)
|
| 29 |
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 34 |
|
| 35 |
class ChatRequest(BaseModel):
|
| 36 |
message: str
|
|
@@ -38,10 +45,18 @@ class ChatRequest(BaseModel):
|
|
| 38 |
|
| 39 |
@app.get("/")
|
| 40 |
def root_check():
|
| 41 |
-
return {"status": "BEGA AI API is running"}
|
| 42 |
|
| 43 |
@app.post("/api/chat")
|
| 44 |
def chat_endpoint(req: ChatRequest):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 45 |
prompt = (
|
| 46 |
"You are BEGA AI, a helpful AI assistant developed by BEGA Company. "
|
| 47 |
"Reply in the same language used in the user's message.\n\n"
|
|
@@ -52,11 +67,14 @@ def chat_endpoint(req: ChatRequest):
|
|
| 52 |
|
| 53 |
prompt += f"User: {req.message}\nAssistant:"
|
| 54 |
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
from llama_cpp import Llama
|
| 6 |
from fastapi.middleware.cors import CORSMiddleware
|
| 7 |
|
| 8 |
+
# Lien direct vers un GGUF Llama-3.1 8B optimisé pour les petites RAM
|
| 9 |
+
MODEL_URL = "https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF/resolve/main/Meta-Llama-3.1-8B-Instruct-Q2_K.gguf"
|
| 10 |
+
MODEL_PATH = "llama-3.1-8B-Q2_K.gguf"
|
| 11 |
|
| 12 |
+
print("--- Vérification du modèle Llama-3.1 8B ---")
|
| 13 |
if not os.path.exists(MODEL_PATH):
|
| 14 |
+
print("Téléchargement de Llama 8B (veuillez patienter quelques minutes)...")
|
| 15 |
+
try:
|
| 16 |
+
urllib.request.urlretrieve(MODEL_URL, MODEL_PATH)
|
| 17 |
+
print("Téléchargement réussi !")
|
| 18 |
+
except Exception as e:
|
| 19 |
+
print(f"Erreur lors du téléchargement : {e}")
|
| 20 |
else:
|
| 21 |
+
print("Le fichier du modèle Llama est détecté sur le disque.")
|
| 22 |
|
| 23 |
app = FastAPI()
|
| 24 |
|
|
|
|
| 30 |
allow_headers=["*"],
|
| 31 |
)
|
| 32 |
|
| 33 |
+
print("--- Chargement du modèle Llama en mémoire RAM ---")
|
| 34 |
+
llm = None
|
| 35 |
+
try:
|
| 36 |
+
# n_ctx=1024 réduit pour éviter de saturer la RAM avec un modèle 8B
|
| 37 |
+
llm = Llama(model_path=MODEL_PATH, n_ctx=1024, n_threads=2)
|
| 38 |
+
print("Modèle Llama correctement initialisé !")
|
| 39 |
+
except Exception as e:
|
| 40 |
+
print(f"Erreur critique lors de l'initialisation de Llama : {e}")
|
| 41 |
|
| 42 |
class ChatRequest(BaseModel):
|
| 43 |
message: str
|
|
|
|
| 45 |
|
| 46 |
@app.get("/")
|
| 47 |
def root_check():
|
| 48 |
+
return {"status": "BEGA AI API is running with Llama-3.1-8B"}
|
| 49 |
|
| 50 |
@app.post("/api/chat")
|
| 51 |
def chat_endpoint(req: ChatRequest):
|
| 52 |
+
global llm
|
| 53 |
+
if llm is None:
|
| 54 |
+
raise HTTPException(
|
| 55 |
+
status_code=500,
|
| 56 |
+
detail="Le modèle IA n'a pas pu être chargé."
|
| 57 |
+
)
|
| 58 |
+
|
| 59 |
+
# Prompt système adapté au format de Llama 3
|
| 60 |
prompt = (
|
| 61 |
"You are BEGA AI, a helpful AI assistant developed by BEGA Company. "
|
| 62 |
"Reply in the same language used in the user's message.\n\n"
|
|
|
|
| 67 |
|
| 68 |
prompt += f"User: {req.message}\nAssistant:"
|
| 69 |
|
| 70 |
+
try:
|
| 71 |
+
output = llm(
|
| 72 |
+
prompt,
|
| 73 |
+
max_tokens=250,
|
| 74 |
+
stop=["User:", "\nUser:", "Assistant:", "\nAssistant:"],
|
| 75 |
+
temperature=0.4
|
| 76 |
+
)
|
| 77 |
+
reply = output['choices'][0]['text'].strip()
|
| 78 |
+
return {"response": reply}
|
| 79 |
+
except Exception as e:
|
| 80 |
+
raise HTTPException(status_code=500, detail=str(e))
|