A Record of the Troubles Encountered During the Installation of the H20 Graphics Card Driver

23 2 月, 2025 109点热度 0人点赞 0条评论

背景：

春节假期结束后，各项工作主要围绕 DeepSeek-R1 各版本的本地部署展开，比往年更忙一些。同时，还对 ComfyUI 和 Janus_pro 提供简单支持，和之前经验一样，大显存，大虚拟内存才是王道。好在春节期间已有相关经验，因此整体进展还算顺利。然而，一个遗留问题始终困扰着我：某物理节点的显卡是 H20，但之前通过 gpu-operator 安装的驱动版本为 535-5.15.0-88，显卡被识别为 NVIDIA-Graphics-Device，而非 H20。尽管应用仍能运行，但总感觉不太对劲。我本以为周末花一小时更新驱动就能轻松解决，没想到却掉进了一个大坑。

过程：

问题初现

这里主要记录过程中的报错及解决办法

第一轮，直接尝试了直接安装最新版本的 CUDA Toolkit

cuda_12.6.0_560.28.03_linux.run

安装时遇到典型的告警

WARNING: An NVIDIA kernel module 'nvidia-uvm' appears to be already loaded in your kernel.

1	WARNING: An NVIDIA kernel module 'nvidia-uvm' appears to be already loaded in your kernel.

解决这个比较直接：停止节点调度，停止 containerd 、kubelet

尝试 rmmod nvidia-uvm 提示n多个进程被占用

强制删除几轮

lsof -w /dev/nvidia* |awk '{print $2}'|xargs kill -9

1	lsof -w /dev/nvidia* \|awk '{print $2}'\|xargs kill -9

后面还有 nvidia模块之类。

然后，继续安装 cuda_12.6.0_560.28.03_linux.run，安装完成，显卡型号H20 识别正常。

nvidia-smi
Sat Feb 22 12:00:29 2025
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 560.28.03              Driver Version: 560.28.03      CUDA Version: 12.6     |
|-----------------------------------------+------------------------+----------------------+
| 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 H20                     Off |   00000000:0F:00.0 Off |                    0 |
| N/A   31C    P0            113W /  500W |       1MiB /  97871MiB |      0%      Default |
|                                         |                        |             Disabled |
+-----------------------------------------+------------------------+----------------------+
|   1  NVIDIA H20                     Off |   00000000:34:00.0 Off |                    0 |
| N/A   29C    P0            111W /  500W |       1MiB /  97871MiB |      0%      Default |
|                                         |                        |             Disabled |
+-----------------------------------------+------------------------+----------------------+
|   2  NVIDIA H20                     Off |   00000000:48:00.0 Off |                    0 |
| N/A   32C    P0            116W /  500W |       1MiB /  97871MiB |      0%      Default |
|                                         |                        |             Disabled |
+-----------------------------------------+------------------------+----------------------+
|   3  NVIDIA H20                     Off |   00000000:5A:00.0 Off |                    0 |
| N/A   30C    P0            111W /  500W |       1MiB /  97871MiB |      0%      Default |
|                                         |                        |             Disabled |
+-----------------------------------------+------------------------+----------------------+
|   4  NVIDIA H20                     Off |   00000000:87:00.0 Off |                    0 |
| N/A   31C    P0            117W /  500W |       1MiB /  97871MiB |      0%      Default |
|                                         |                        |             Disabled |
+-----------------------------------------+------------------------+----------------------+
|   5  NVIDIA H20                     Off |   00000000:AE:00.0 Off |                    0 |
| N/A   28C    P0            111W /  500W |       1MiB /  97871MiB |      0%      Default |
|                                         |                        |             Disabled |
+-----------------------------------------+------------------------+----------------------+
|   6  NVIDIA H20                     Off |   00000000:C2:00.0 Off |                    0 |
| N/A   31C    P0            114W /  500W |       1MiB /  97871MiB |      0%      Default |
|                                         |                        |             Disabled |
+-----------------------------------------+------------------------+----------------------+
|   7  NVIDIA H20                     Off |   00000000:D7:00.0 Off |                    0 |
| N/A   29C    P0            116W /  500W |       1MiB /  97871MiB |      0%      Default |
|                                         |                        |             Disabled |
+-----------------------------------------+------------------------+----------------------+

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

nvidia-smi

Sat Feb 22 12:00:29 2025

+-----------------------------------------------------------------------------------------+

| NVIDIA-SMI 560.28.03 Driver Version: 560.28.03 CUDA Version: 12.6 |

|-----------------------------------------+------------------------+----------------------+

| 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 H20 Off | 00000000:0F:00.0 Off | 0 |

| N/A 31C P0 113W / 500W | 1MiB / 97871MiB | 0% Default |

| | | Disabled |

+-----------------------------------------+------------------------+----------------------+

| 1 NVIDIA H20 Off | 00000000:34:00.0 Off | 0 |

| N/A 29C P0 111W / 500W | 1MiB / 97871MiB | 0% Default |

| | | Disabled |

+-----------------------------------------+------------------------+----------------------+

| 2 NVIDIA H20 Off | 00000000:48:00.0 Off | 0 |

| N/A 32C P0 116W / 500W | 1MiB / 97871MiB | 0% Default |

| | | Disabled |

+-----------------------------------------+------------------------+----------------------+

| 3 NVIDIA H20 Off | 00000000:5A:00.0 Off | 0 |

| N/A 30C P0 111W / 500W | 1MiB / 97871MiB | 0% Default |

| | | Disabled |

+-----------------------------------------+------------------------+----------------------+

| 4 NVIDIA H20 Off | 00000000:87:00.0 Off | 0 |

| N/A 31C P0 117W / 500W | 1MiB / 97871MiB | 0% Default |

| | | Disabled |

+-----------------------------------------+------------------------+----------------------+

| 5 NVIDIA H20 Off | 00000000:AE:00.0 Off | 0 |

| N/A 28C P0 111W / 500W | 1MiB / 97871MiB | 0% Default |

| | | Disabled |

+-----------------------------------------+------------------------+----------------------+

| 6 NVIDIA H20 Off | 00000000:C2:00.0 Off | 0 |

| N/A 31C P0 114W / 500W | 1MiB / 97871MiB | 0% Default |

| | | Disabled |

+-----------------------------------------+------------------------+----------------------+

| 7 NVIDIA H20 Off | 00000000:D7:00.0 Off | 0 |

| N/A 29C P0 116W / 500W | 1MiB / 97871MiB | 0% Default |

| | | Disabled |

+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+

| Processes: |

| GPU GI CI PID Type Process name GPU Memory |

| ID ID Usage |

|=========================================================================================|

| No running processes found |

+-----------------------------------------------------------------------------------------+

问题来了， gpu-opearator 下 nvidia-cuda-validator-**反复重启

日志类似

Failed to allocate device vector A (error code system not yet initialized)!

[Vector addition of 50000 elements]

后面陆续试了几个 cuda 工具集，但 pod依旧报错。没办法，我只能拉上同事一起排查问题。

深入排查

我们首先试对着 CUDA Toolkit 官方文档查看兼容性，发现当前版本的 CUDA Toolkit 并不完全支持 H20 。

https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html

然后看了一下几个版本 cuda的 gpu表

/usr/local/cuda-12.8/bin/nvcc --list-gpu-arch

/usr/local/cuda-12.6/bin/nvcc --list-gpu-arch

/usr/local/cuda-12.4/bin/nvcc --list-gpu-arch

除开 12.8 版本支持了 100及以上compute_100、compute_101、compute_102

其他版本只支持到 compute_90 ，我理解为版本较低的 nvcc 不支持 H20卡。

然后检查驱动 https://www.nvidia.com/en-us/drivers/

Recommended/Certified H20 仅有4个版本认证过。

Driver Version: 550.144.03，CUDA Toolkit: 12.4，Release Date: Thu Jan 16, 2025
Driver Version: 535.230.02，CUDA Toolkit: 12.2，Release Date: Thu Jan 16, 2025
Driver Version: 550.127.08，CUDA Toolkit: 12.4，Release Date: Tue Nov 19, 2024
Driver Version: 535.216.03，CUDA Toolkit: 12.2，Release Date: Tue Nov 19, 2024

到这里发现踩坑严重了, 这让我意识到，之前尝试的驱动版本可能并不适合 H20 显卡。

验证与解决

为了进一步验证问题，同事建议我在节点上编译并运行 vectorAdd 示例程序。我们先后尝试了 1.24 和 1.28 两个版本的 CUDA 样例，结果都报错：

https://github.com/NVIDIA/cuda-samples

报错大概如下

~/124/cuda-samples-12.4/bin/x86_64/linux/release
./vectorAdd
[Vector addition of 50000 elements]
Failed to allocate device vector A (error code CUDA driver version is insufficient for CUDA runtime version)!
或者报错  Failed to allocate device vector A (error code system not yet initialized)!

~/124/cuda-samples-12.4/bin/x86_64/linux/release

./vectorAdd

[Vector addition of 50000 elements]

Failed to allocate device vector A (error code CUDA driver version is insufficient for CUDA runtime version)!

或者报错 Failed to allocate device vector A (error code system not yet initialized)!

正常日志应该是这样

./vectorAdd
[Vector addition of 50000 elements]
Copy input data from the host memory to the CUDA device
CUDA kernel launch with 196 blocks of 256 threads
Copy output data from the CUDA device to the host memory
Test PASSED
Done

./vectorAdd

[Vector addition of 50000 elements]

Copy input data from the host memory to the CUDA device

CUDA kernel launch with 196 blocks of 256 threads

Copy output data from the CUDA device to the host memory

Test PASSED

Done

这进一步确认了问题是由于驱动版本与 CUDA 运行时版本不匹配导致的，与 k8s 环境无关。

另外，我们还注意到内核日志中有以下报错：

[Sat Feb 22 15:25:54 2025] NVRM: _knvlinkCheckFabricCliqueId: GPU 0 failed to get fabric clique Id: 0x55
[Sat Feb 22 15:25:54 2025] NVRM: _knvlinkCheckFabricCliqueId: GPU 1 failed to get fabric clique Id: 0x55

1 2	[Sat Feb 22 15:25:54 2025] NVRM: _knvlinkCheckFabricCliqueId: GPU 0 failed to get fabric clique Id: 0x55 [Sat Feb 22 15:25:54 2025] NVRM: _knvlinkCheckFabricCliqueId: GPU 1 failed to get fabric clique Id: 0x55

机房的同事指出，这是因为没有安装 NVIDIA-fabricmanager。

解决：

有了这些线索，问题的解决思路逐渐清晰。我们首先按照官方推荐的方式卸载了之前安装的各种驱动：

/usr/local/cuda-X.Y/bin/cuda-uninstaller

发现可以一次全选

CUDA Uninstaller                                                               │
│   [X] CUDA_Toolkit_12.5                                                      │
│   [X] CUDA_Demo_Suite_12.6                                                   │
│   [X] CUDA_Toolkit_12.4                                                      │
│   [X] CUDA_Demo_Suite_12.5                                                   │
│   [X] CUDA_Documentation_12.5                                                │
│   [X] CUDA_Demo_Suite_12.8                                                   │
│   [X] CUDA_Demo_Suite_12.4                                                   │
│   [X] CUDA_Documentation_12.8                                                │
│   [X] CUDA_Documentation_12.6                                                │
│   [X] CUDA_Documentation_12.4                                                │
│   [X] CUDA_Toolkit_12.6                                                      │
│   [X] CUDA_Toolkit_12.8

CUDA Uninstaller │

│ [X] CUDA_Toolkit_12.5 │

│ [X] CUDA_Demo_Suite_12.6 │

│ [X] CUDA_Toolkit_12.4 │

│ [X] CUDA_Demo_Suite_12.5 │

│ [X] CUDA_Documentation_12.5 │

│ [X] CUDA_Demo_Suite_12.8 │

│ [X] CUDA_Demo_Suite_12.4 │

│ [X] CUDA_Documentation_12.8 │

│ [X] CUDA_Documentation_12.6 │

│ [X] CUDA_Documentation_12.4 │

│ [X] CUDA_Toolkit_12.6 │

│ [X] CUDA_Toolkit_12.8

然后，我们根据 NVIDIA 官方驱动下载页面和 CUDA Toolkit 官方文档确定了 H20 显卡支持的 CUDA 和驱动版本。最终选择了 cuda_12.4.1_550.54.15_linux.run，并搭配认证过的 NVIDIA-Linux-x86_64-550.144.03.run 驱动进行安装。

安装完成后，我们还安装了 nvidia-fabricmanager 和 nvidia-compute-utils：版本一定要配对。

apt install nvidia-fabricmanager-550 nvidia-fabricmanager-dev-550
apt install nvidia-compute-utils-550  -y
systemctl enable --now nvidia-fabricmanager (如果版本对不上，这里会没法启动）