diff --git a/gpustack_runtime/deployer/cdi/ascend.py b/gpustack_runtime/deployer/cdi/ascend.py index b00c634..da33483 100644 --- a/gpustack_runtime/deployer/cdi/ascend.py +++ b/gpustack_runtime/deployer/cdi/ascend.py @@ -43,6 +43,16 @@ https://gitcode.com/Ascend/mind-cluster/blob/master/docs/en/scheduling/references/appendix.md. """ +_HCCN_CONF_PATH = "/etc/hccn.conf" +""" +The device NIC addresses, as configured on the host. A transport that builds +RoCE endpoints reads them from this file or from hccn_tool, which lives in a +driver directory none of the mounts below covers; without either, the engine +fails to initialize with "Failed to get device ip from hccn.conf and +hccn_tool". A container that talks to another node over Device RoCE therefore +needs this file, and one that does not is unaffected by its presence. +""" + _A5_RANKTABLE_VERSION = "2.0" """ The ranktable version the A5 generation requires. @@ -160,6 +170,7 @@ def generate( mount_paths = [ _HCCL_RANKTABLE_PATH, + _HCCN_CONF_PATH, "/usr/local/Ascend/driver/topo", "/usr/local/Ascend/driver/lib64", "/usr/local/Ascend/driver/include",