Browse Source

fix(mgb/opr-mm): fix group manager sync error

GitOrigin-RevId: 344aae534c
tags/v1.0.0-rc1
Megvii Engine Team 4 years ago
parent
commit
be5dde9da8
1 changed files with 1 additions and 1 deletions
  1. +1
    -1
      src/opr-mm/impl/group_manager.cpp

+ 1
- 1
src/opr-mm/impl/group_manager.cpp View File

@@ -71,7 +71,6 @@ void GroupInfo::add_opr(const std::string& key, size_t nr_expected_devices,
}
m_opr_infos.push_back({comp_node_hash, is_root, rank});
m_nr_registered_devs++;
m_count++;
if (m_nr_registered_devs > nr_expected_devices) {
mgb_log_error(
"too many opr registered with key %s, expected %zu, actual %u",
@@ -84,6 +83,7 @@ void GroupInfo::add_opr(const std::string& key, size_t nr_expected_devices,
if (m_nr_expected_devs == m_nr_registered_devs) {
sort_opr_infos();
gen_infos_from_opr_infos();
m_count = m_nr_registered_devs;
m_register_cv.notify_all();
} else {
m_register_cv.wait(lk,


Loading…
Cancel
Save