Je suis passé en version “Ubuntu 24.04.2 LTS” (le kernel est 6.8.0-60-generic)
Migration :
- Ubuntu 22.04.5 LTS ( kernel : 5.15.0-140-generic ) => Ubuntu 24.04.2 LTS (kernel est 6.8.0-60-generic)
- CUDA : 12.8.93 => 12.9
- Python : 3.11 => 3.12
Les cartes NVIDIA sont toujours visibles :
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 575.51.03 Driver Version: 575.51.03 CUDA Version: 12.9 |
|-----------------------------------------+------------------------+----------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 Quadro M5000 Off | 00000000:00:10.0 Off | Off |
| 39% 44C P8 14W / 150W | 5MiB / 8192MiB | 0% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+
| 1 Quadro M4000 Off | 00000000:00:11.0 Off | N/A |
| 49% 48C P8 14W / 120W | 5MiB / 8192MiB | 0% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+
J’ai refait un benchmark :
llm_benchmark run
-------Linux----------
{'id': '0', 'name': 'Quadro M5000', 'driver': '575.51.03',
'gpu_memory_total': '8192.0 MB', 'gpu_memory_free': '8110.0 MB',
'gpu_memory_used': '5.0 MB', 'gpu_load': '0.0%', 'gpu_temperature': '44.0°C'}
{'id': '1', 'name': 'Quadro M4000', 'driver': '575.51.03',
'gpu_memory_total': '8192.0 MB', 'gpu_memory_free': '8110.0 MB',
'gpu_memory_used': '5.0 MB', 'gpu_load': '0.0%', 'gpu_temperature': '48.0°C'}
At least two GPU cards
Total memory size : 119.03 GB
cpu_info: Intel(R) Xeon(R) CPU E5-2450 v2 @ 2.50GHz
gpu_info: Quadro M5000
Quadro M4000
os_version: Ubuntu 24.04.2 LTS
ollama_version: 0.9.0
----------
...
At least two GPU cards
{
"phi4:14b": "6.75",
"deepseek-r1:14b": "6.19",
"deepseek-r1:32b": "0.43",
"uuid": "2a3d3de2-5e53-5b28-a909-62559c5a817c",
"ollama_version": "0.9.0"
}
-------
Maintenant les grands modèles (deepseek-r1:32b) ne font plus planter le test …. qui dure 4 heures.
Misère.