Tuxedo 17 (en Ubuntu 24) + GIGABYTE AORUS RTX 5060 Ti AI Box Carte Graphique - 16GB GDDR7.

Tuxedo17

J’ai l’impression qu’il y a un bug ouvert : https://github.com/NVIDIA/open-gpu-kernel-modules/issues/979#issuecomment-4172636691 : RTX 5080 via Thunderbolt 5 eGPU: Hard lock on CUDA operations (nvidia-smi works at idle) .

Tuxedo17

Sans l’extenstion :

# llama-bench -m /models/gemma-3-1b-it-q4_k_m.gguf 
| model                          |       size |     params | backend    | threads |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: |
| gemma3 1B Q4_K - Medium        | 762.49 MiB |   999.89 M | CPU        |       8 |           pp512 |        379.86 ± 6.14 |
| gemma3 1B Q4_K - Medium        | 762.49 MiB |   999.89 M | CPU        |       8 |           tg128 |         49.11 ± 0.20 |

build: c3d47e696 (10030)
# nvidia-smi 
Thu Jul 16 11:00:14 2026       
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 580.159.04             Driver Version: 580.159.04     CUDA Version: 13.0     |
+-----------------------------------------+------------------------+----------------------+
| 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  NVIDIA GeForce RTX 3060 ...    Off |   00000000:01:00.0 Off |                  N/A |
| N/A   43C    P0             24W /  115W |       1MiB /   6144MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI              PID   Type   Process name                        GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|  No running processes found                                                             |
+-----------------------------------------------------------------------------------------+

Tuxedo17

Avec l’eGPU :

Il y a une erreur

 1261.324229] NVRM: GPU at PCI:0000:05:00: GPU-5d8d2eb1-648d-e52f-369a-26f078362f02
[ 1261.324235] NVRM: GPU Board Serial Number: 0
[ 1261.324235] NVRM: Xid (PCI:0000:05:00): 79, GPU has fallen off the bus.
[ 1261.324240] NVRM: GPU 0000:05:00.0: GPU has fallen off the bus.
[ 1261.324242] NVRM: GPU 0000:05:00.0: GPU serial number is 0.
[ 1261.324247] NVRM: kgspRcAndNotifyAllChannels_IMPL: RC all channels for critical error 79.
[ 1261.324254] NVRM: _threadNodeCheckTimeout: API_GPU_ATTACHED_SANITY_CHECK failed!
[ 1261.324263] NVRM: _threadNodeCheckTimeout: API_GPU_ATTACHED_SANITY_CHECK failed!
[ 1261.324268] NVRM: _threadNodeCheckTimeout: API_GPU_ATTACHED_SANITY_CHECK failed!
[ 1261.324272] NVRM: _threadNodeCheckTimeout: API_GPU_ATTACHED_SANITY_CHECK failed!
[ 1261.324388] NVRM: _issueRpcAndWait: rpcSendMessage failed with status 0x0000000f for fn 78 sequence 358!
[ 1261.324391] NVRM: nvCheckOkFailedNoLog: Check failed: GPU lost from the bus [NV_ERR_GPU_IS_LOST] (0x0000000F) returned from nvdEngineDumpCallbackHelper(pGpu, pPrbEnc, pNvDumpState, pEngineCallback) @ nv_debug_dump.c:273
[ 1261.324424] NVRM: _issueRpcAndWait: rpcSendMessage failed with status 0x0000000f for fn 78 sequence 359!

Tuxedo17

Mise à jours avec la dernière version :

# sudo apt-get purge nvidia-* 

# add-apt-repository ppa:graphics-drivers/ppa -y

# apt update 

# apt install nvidia-driver-595 -y

Tuxedo17

Ajout du fichier :

cat  /etc/modprobe.d/nvidia-graphics-drivers.conf 
options nvidia NVreg_EnableGpuFirmware=0

Désactiver le firmware GSP : Si vous utilisez une série RTX 40, parfois le firmware GSP provoque des erreurs NVRM. Empêchez cela en ajoutant NVreg_EnableGpuFirmware=0 à la configuration du pilote

Tuxedo17

Lien à voir : https://github.com/mmhorda/ai-box-rtx-5060ti-egpu-guide

Tuxedo17

Téléchargement des modèles :

# du -sh /models/*
769M    /models/gemma-3-1b-it-q4_k_m.gguf
889M    /models/mmproj-BF16(1).gguf
862M    /models/mmproj-BF16.gguf
885M    /models/mmproj-F16.gguf
5,3G    /models/Qwen3.5-9B-Q4_K_M.gguf
16G     /models/Qwen3.6-27B-Q4_K_M.gguf
21G     /models/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf

Tuxedo17

Check :

# boltctl list
 ● GIGABYTE AORUS RTX506T AI BOX
   ├─ type:          peripheral
   ├─ name:          AORUS RTX506T AI BOX
   ├─ vendor:        GIGABYTE
   ├─ uuid:          008b38f1-ba7e-8780-ffff-ffffffffffff
   ├─ generation:    USB4
   ├─ status:        authorized
   │  ├─ domain:     d02c5899-4b21-8780-ffff-ffffffffffff
   │  ├─ rx speed:   40 Gb/s = 2 lanes * 20 Gb/s
   │  ├─ tx speed:   40 Gb/s = 2 lanes * 20 Gb/s
   │  └─ authflags:  none
   ├─ authorized:    jeu. 16 juil. 2026 11:33:13
   ├─ connected:     jeu. 16 juil. 2026 11:33:13
   └─ stored:        ven. 05 juin 2026 15:47:50
      ├─ policy:     iommu
      └─ key:        no

# boltctl authorize 008b38f1-ba7e-8780-ffff-ffffffffffff

# nvidia-smi 
Thu Jul 16 13:35:11 2026       
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 595.71.05              Driver Version: 595.71.05      CUDA Version: 13.2     |
+-----------------------------------------+------------------------+----------------------+
| 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  NVIDIA GeForce RTX 3060 ...    Off |   00000000:01:00.0 Off |                  N/A |
| N/A   40C    P0            752W /  115W |       1MiB /   6144MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI              PID   Type   Process name                        GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|  No running processes found                                                             |
+-----------------------------------------------------------------------------------------+

J’ai l’erreur :

[   15.841164] NVRM: The NVIDIA GPU 0000:05:00.0 (PCI ID: 10de:2d04)
               NVRM: installed in this system requires use of the NVIDIA open kernel modules.
[   15.841198] NVRM: GPU 0000:05:00.0: RmInitAdapter failed! (0x22:0x56:1017)
[   15.841201] NVRM: GPU 0000:05:00.0: rm_init_adapter failed, device minor number 1

Tuxedo17

Je fais donc :

# apt-get install nvidia-open

Tuxedo17

Nouveau test :

# llama-bench -m /models/gemma-3-1b-it-q4_k_m.gguf 
| model                          |       size |     params | backend    | threads |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: |
| gemma3 1B Q4_K - Medium        | 762.49 MiB |   999.89 M | CPU        |       8 |           pp512 |       371.45 ± 41.49 |
| gemma3 1B Q4_K - Medium        | 762.49 MiB |   999.89 M | CPU        |       8 |           tg128 |         50.31 ± 0.11 |

build: c3d47e696 (10030)
# llama-bench -m /models/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf 
| model                          |       size |     params | backend    | threads |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: |
| qwen35moe 35B.A3B Q4_K - Medium |  20.60 GiB |    34.66 B | CPU        |       8 |           pp512 |         82.60 ± 2.10 |
| qwen35moe 35B.A3B Q4_K - Medium |  20.60 GiB |    34.66 B | CPU        |       8 |           tg128 |         12.67 ± 0.35 |

build: c3d47e696 (10030)

Tuxedo17

Rien n’est visible :

# nvidia-smi 
Thu Jul 16 14:23:40 2026       
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 580.159.04             Driver Version: 580.159.04     CUDA Version: 13.0     |
+-----------------------------------------+------------------------+----------------------+
| 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  NVIDIA GeForce RTX 3060 ...    Off |   00000000:01:00.0 Off |                  N/A |
| N/A   45C    P8             11W /  115W |       1MiB /   6144MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+
|   1  NVIDIA GeForce RTX 5060 Ti     Off |   00000000:05:00.0 Off |                  N/A |
|  0%   45C    P8              5W /  180W |       2MiB /  16311MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI              PID   Type   Process name                        GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|  No running processes found                                                             |
+-----------------------------------------------------------------------------------------+
# gpustat 
tuxedo-17                              Thu Jul 16 14:23:45 2026  580.159.04
[0] NVIDIA GeForce RTX 3060 Laptop GPU | 45°C,   0 % |     0 /  6144 MB |
[1] NVIDIA GeForce RTX 5060 Ti         | 45°C,   0 % |     1 / 16311 MB |

Tuxedo17

CPU capability :

# nvidia-smi --query-gpu=compute_cap --format=csv
compute_cap
8.6
12.0

Tuxedo17

Nouveau build : ( avec DCMAKE_CUDA_ARCHITECTURES=86 )

#  apt install -y nvidia-cuda-toolkit

#  cd llama.cpp

#  export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64:/usr/local/cuda/extras/CUPTI/lib64

# export PATH=$PATH:$CUDA_HOME/bin

# cmake -B build -DGGML_CUDA=ON  -DCMAKE_CUDA_COMPILER=`which nvcc` -DLLAMA_CURL=ON -DCMAKE_CUDA_ARCHITECTURES=86

# cmake --build build -j$(nproc)

Tuxedo17

Info chouchou : https://github.com/NVIDIA/nvbench/discussions/129

So for your example of a 3090, the compute capability is 8.6. So to build both PTX and SASS for that GPU you would specify -DCMAKE_CUDA_ARCHITECTURES=86.

Tuxedo17

Ajout de nvidia-dkms ( pour noyau personnalisé )

# apt-get install nvidia-dkms-595

A suivre.

Tuxedo17

Je recommance :

# nvidia-smi 
Thu Jul 16 15:09:01 2026       
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 580.159.04             Driver Version: 580.159.04     CUDA Version: 13.0     |
+-----------------------------------------+------------------------+----------------------+
| 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  NVIDIA GeForce RTX 3060 ...    Off |   00000000:01:00.0 Off |                  N/A |
| N/A   40C    P0             25W /  115W |       1MiB /   6144MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI              PID   Type   Process name                        GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|  No running processes found                                                             |
+-----------------------------------------------------------------------------------------+

# apt-get install nvidia-dkms-580

Tuxedo17

Ce message a été supprimé !

Tuxedo17

Uname & dkms

# uname -a
Linux tuxedo-17 6.17.0-122035-tuxedo #35~24.04.1tux1 SMP PREEMPT_DYNAMIC Mon Jun 15 14:27:53 UTC 2026 x86_64 x86_64 x86_64 GNU/Linux

# dkms status
evdi/1.14.14+dfsg, 6.17.0-122035-tuxedo, x86_64: installed
nvidia/580.95.05, 6.17.0-122035-tuxedo, x86_64: installed
r8125/9.011.00, 6.17.0-122035-tuxedo, x86_64: installed
tuxedo-drivers/4.22.3, 6.17.0-122035-tuxedo, x86_64: installed

Tuxedo17

Installation de chronium :

# apt-get install chromium-browser

Tuxedo17

Sans l’egpu :

#  llama-bench -m /models/gemma-3-1b-it-q4_k_m.gguf 
ggml_cuda_init: found 1 CUDA devices (Total VRAM: 5806 MiB):
  Device 0: NVIDIA GeForce RTX 3060 Laptop GPU, compute capability 8.6, VMM: yes, VRAM: 5806 MiB
| model                          |       size |     params | backend    | ngl |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |

| gemma3 1B Q4_K - Medium        | 762.49 MiB |   999.89 M | CUDA       |  -1 |           pp512 |    11293.53 ± 542.50 |
| gemma3 1B Q4_K - Medium        | 762.49 MiB |   999.89 M | CUDA       |  -1 |           tg128 |        224.54 ± 0.68 |

build: c3d47e696 (10030)

NodeBB

Tuxedo 17 (en Ubuntu 24) + GIGABYTE AORUS RTX 5060 Ti AI Box Carte Graphique - 16GB GDDR7.