已合并
[fix] 修正EI0020端口冲突处置指引 (#987) #5725
ren-mengguang创建于 9月28日
[fix] 修正EI0020端口冲突处置指引 (#987) #5725
已合并
共 3 个文件变更+11-3
| @@ -16,4 +16,8 @@ Failed to enable listening for the NPU network adapter socket. Reason: The IP ad | |||
| 16 | 16 | ||
| 17 | ## Solution | 17 | ## Solution |
| 18 | 18 | ||
| 19 | -Check whether the single-card multi-process scenario is used. If yes, configure the port number using the environment variable HCCL_NPU_SOCKET_PORT_RANGE. | 19 | +1. Check whether the single-card multi-process scenario is used (multiple service processes are started on the same device). If yes, handle the issue based on how the communication domain is initialized: |
| 20 | + | ||
| 21 | + - If the communication domain is initialized using root info: configure the environment variable HCCL_NPU_SOCKET_PORT_RANGE to specify a port range. | ||
| 22 | + | ||
| 23 | + - If the communication domain is initialized using a rank table file: HCCL_NPU_SOCKET_PORT_RANGE does not take effect in this mode. Configure different device_port/host_port values for the rank with the same device_ip in the rank table file used by each process (the default value 16666 is used when the fields are not configured). | ||
| @@ -16,4 +16,8 @@ Failed to enable listening for the NPU network adapter socket. Reason: The IP ad | |||
| 16 | 16 | ||
| 17 | ## 解决方法 | 17 | ## 解决方法 |
| 18 | 18 | ||
| 19 | -请确认是否为单卡多进程场景，如果是的话请通过环境变量HCCL_NPU_SOCKET_PORT_RANGE配置对应的端口号。 | 19 | +1. 确认是否为单卡多进程场景（同一Device上拉起多个业务进程）。若是，根据通信域初始化方式分别处理： |
| 20 | + | ||
| 21 | + - rootInfo方式初始化通信域：请配置环境变量HCCL_NPU_SOCKET_PORT_RANGE指定端口范围。 | ||
| 22 | + | ||
| 23 | + - rank table文件方式初始化通信域：该方式下HCCL_NPU_SOCKET_PORT_RANGE不生效，请为每个进程使用的rank table文件中同一device_ip的rank配置互不相同的device_port/host_port（字段缺省时默认使用16666，单卡多进程场景下会冲突）。 | ||
| @@ -233,7 +233,7 @@ const std::string hcomm_g_msg = R"( | |||
| 233 | "Arglist": "reason", | 233 | "Arglist": "reason", |
| 234 | "suggestion": { | 234 | "suggestion": { |
| 235 | "Possible Cause": "N/A", | 235 | "Possible Cause": "N/A", |
| 236 | - "Solution": "Check whether the single-card multi-process scenario is used. If yes, configure the port number using the environment variable HCCL_NPU_SOCKET_PORT_RANGE." | 236 | + "Solution": "1. Check whether the single-card multi-process scenario is used. If yes: for communication domains initialized using root info, configure the port range using the environment variable HCCL_NPU_SOCKET_PORT_RANGE; for communication domains initialized using a rank table file, HCCL_NPU_SOCKET_PORT_RANGE does not take effect in this mode, configure different device_port values for the same device IP in the rank table file used by each process (the default port 16666 is used when device_port is not configured).\r\n2. For detailed troubleshooting guidance, search for the keyword \"EI0020\" on https://www.hiascend.com/document/." |
| 237 | } | 237 | } |
| 238 | } | 238 | } |
| 239 | ] | 239 | ] |