Benchmark simple d’Ollama

import time
import requests
import json

def benchmark_ollama(server, model, prompt, num_runs=5):
    times = []
    tokens_per_sec = []

    for i in range(num_runs):
        start = time.time()

        response = requests.post(server,
            json={
                'model': model,
                'prompt': prompt,
                'stream': False
            })

        end = time.time()
        data = response.json()

        times.append(end - start)
        if 'eval_duration' in data:
            # eval_duration est en nanosecondes
            tokens = data.get('eval_count', 0)
            duration_sec = data['eval_duration'] / 1e9
            tokens_per_sec.append(tokens / duration_sec if duration_sec > 0 else 0)

    print(f"Serveur: {server}")
    print(f"Modèle: {model}")
    print(f"Temps moyen: {sum(times)/len(times):.2f}s")
    print(f"Tokens/sec moyen: {sum(tokens_per_sec)/len(tokens_per_sec):.2f}")
    print(f"Min: {min(times):.2f}s, Max: {max(times):.2f}s")

# Utilisation
benchmark_ollama('http://localhost:11434/api/generate', 'mistral', 'Explique-moi la physique quantique', num_runs=3)
#benchmark_ollama('http://X.X.X.X:11434/api/generate', 'mistral', 'Explique-moi la physique quantique', num_runs=3)