happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

hotfix, rdma keepalive based on tcp connection

This commit is contained in:
Valeriy Dubov 2026-03-16 13:46:44 +03:00
parent 759c28de3a
commit 22c038ac84
1 changed files with 15 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
ggml/src/ggml-rpc/ggml-rpc.cpp
View File

@ -547,14 +547,21 @@ static inline bool rdma_poll(struct ibv_cq * cq, struct ibv_wc * wc, int tcp_fd
}
if (n < 0) return false;
if ((s & 0xFFFFF) == 0 && s > 0) {
if (tcp_fd >= 0 && tcp_peer_closed(tcp_fd)) {
return false;
}
struct timespec now;
clock_gettime(CLOCK_MONOTONIC_COARSE, &now);
if (now.tv_sec - t0.tv_sec >= 30) {
GGML_LOG_ERROR("RDMA CQ poll timeout\n");
return false;
if (tcp_fd >= 0) {
// TCP FD available: use it as the authoritative liveness check.
// Abort immediately on peer close; otherwise wait indefinitely
// (the connection may be idle between inference requests).
if (tcp_peer_closed(tcp_fd)) {
return false;
}
} else {
// No TCP FD: fall back to a hard 30s timeout as safety net.
struct timespec now;
clock_gettime(CLOCK_MONOTONIC_COARSE, &now);
if (now.tv_sec - t0.tv_sec >= 30) {
GGML_LOG_ERROR("RDMA CQ poll timeout\n");
return false;
}
}
}
}