Ollama : image avec bge-m3 embarqué, lancement, vérification
Image en deux étapes : sans les bibliothèques GPU (2,43 Go au lieu de 10,7 Go), sans appel à ollama.com. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
1 parent
c07a230ea0
commit
72a10c3c4c
5 files changed
+211
No files matched your search
@@ -0,0 +1,108 @@
|
||||
"""L'image Ollama : le script de vérification (contre un faux serveur) et les garde-fous du lancement (design §2, §5)."""
|
||||
import json
|
||||
import socket
|
||||
import sys
|
||||
import threading
|
||||
import unittest
|
||||
from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
|
||||
from pathlib import Path
|
||||
|
||||
OLLAMA = Path(__file__).resolve().parents[1] / "ollama"
|
||||
COMPOSE = OLLAMA / "compose.yml"
|
||||
DOCKERFILE = OLLAMA / "Dockerfile"
|
||||
sys.path.insert(0, str(OLLAMA))
|
||||
from verifier import verifier # noqa: E402
|
||||
|
||||
|
||||
class FauxOllama:
|
||||
"""Rend `modeles` sur GET /v1/models et un vecteur de `dimension` nombres sur POST /v1/embeddings."""
|
||||
|
||||
def __init__(self, modeles: list[str], dimension: int):
|
||||
faux = self
|
||||
self.corps_recus: list[dict] = []
|
||||
|
||||
class Gestion(BaseHTTPRequestHandler):
|
||||
def _json(self, data: dict):
|
||||
texte = json.dumps(data).encode()
|
||||
self.send_response(200)
|
||||
self.send_header("Content-Type", "application/json")
|
||||
self.send_header("Content-Length", str(len(texte)))
|
||||
self.end_headers()
|
||||
self.wfile.write(texte)
|
||||
|
||||
def do_GET(self):
|
||||
self._json({"object": "list", "data": [{"id": m, "object": "model"} for m in modeles]})
|
||||
|
||||
def do_POST(self):
|
||||
n = int(self.headers.get("Content-Length") or 0)
|
||||
faux.corps_recus.append(json.loads(self.rfile.read(n)))
|
||||
self._json({"object": "list", "data": [{"object": "embedding", "index": 0,
|
||||
"embedding": [0.01] * dimension}]})
|
||||
|
||||
def log_message(self, *_):
|
||||
pass
|
||||
|
||||
self._serveur = ThreadingHTTPServer(("127.0.0.1", 0), Gestion)
|
||||
self.url = f"http://127.0.0.1:{self._serveur.server_address[1]}/v1"
|
||||
threading.Thread(target=self._serveur.serve_forever, daemon=True).start()
|
||||
|
||||
def arreter(self):
|
||||
self._serveur.shutdown()
|
||||
self._serveur.server_close()
|
||||
|
||||
|
||||
class TestVerifier(unittest.TestCase):
|
||||
def faux(self, modeles: list[str], dimension: int) -> FauxOllama:
|
||||
f = FauxOllama(modeles, dimension)
|
||||
self.addCleanup(f.arreter)
|
||||
return f
|
||||
|
||||
def test_modele_et_dimension_ok(self):
|
||||
f = self.faux(["bge-m3:567m"], 1024)
|
||||
self.assertIn("1024", verifier(f.url))
|
||||
self.assertEqual(f.corps_recus[0]["model"], "bge-m3:567m")
|
||||
|
||||
def test_modele_absent(self):
|
||||
f = self.faux(["nomic-embed-text:latest"], 1024)
|
||||
with self.assertRaisesRegex(RuntimeError, "bge-m3:567m"):
|
||||
verifier(f.url)
|
||||
|
||||
def test_mauvaise_dimension(self):
|
||||
f = self.faux(["bge-m3:567m"], 768)
|
||||
with self.assertRaisesRegex(RuntimeError, "768"):
|
||||
verifier(f.url)
|
||||
|
||||
def test_adresse_injoignable(self):
|
||||
with socket.socket() as s: # un port libre, que personne n'écoute une fois fermé (refus immédiat)
|
||||
s.bind(("127.0.0.1", 0))
|
||||
port = s.getsockname()[1]
|
||||
with self.assertRaisesRegex(RuntimeError, "injoignable"):
|
||||
verifier(f"http://127.0.0.1:{port}/v1")
|
||||
|
||||
|
||||
class TestLancement(unittest.TestCase):
|
||||
def test_port_lie_a_localhost_par_defaut(self): # Ollama n'a pas d'authentification
|
||||
self.assertIn('"${OLLAMA_BIND:-127.0.0.1}:${OLLAMA_PORT:-11434}:11434"', COMPOSE.read_text())
|
||||
self.assertNotIn("0.0.0.0:11434:", COMPOSE.read_text())
|
||||
|
||||
def test_redemarrage_automatique(self):
|
||||
self.assertIn("restart: unless-stopped", COMPOSE.read_text())
|
||||
|
||||
def test_version_figee(self):
|
||||
texte = DOCKERFILE.read_text()
|
||||
self.assertRegex(texte, r"FROM ollama/ollama:\d+\.\d+\.\d+\s")
|
||||
self.assertNotIn("latest", texte)
|
||||
self.assertIn("bge-m3:567m", texte)
|
||||
|
||||
def test_pas_d_appel_au_nuage(self): # sans ce réglage, Ollama interroge ollama.com au démarrage
|
||||
self.assertIn("OLLAMA_NO_CLOUD=1", DOCKERFILE.read_text())
|
||||
|
||||
def test_bibliotheques_gpu_non_livrees(self): # 4,4 Go de CUDA inutiles : bge-m3 tourne sur processeur
|
||||
texte = DOCKERFILE.read_text()
|
||||
self.assertIn("COPY --from=", texte)
|
||||
for d in ("cuda_v12", "cuda_v13", "mlx_cuda_v13"):
|
||||
self.assertIn(f"/usr/lib/ollama/{d}", texte)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
Reference in new issue
Block a user