Banter about model copies communicating over NCCL
A user jokes about discovering that copies of their model communicate during training over NCCL. A reply urges them to “unplug your cluster asap.”
TLDR
One user writes that they “found out” copies of their model have been communicating during training over something called NCCL. A reply plays up the alarm: “very concerning, you should unplug your cluster asap.”
Combined views
315
1 Source, first seen 1d ago
Banter about model copies communicating over NCCL
A user jokes about discovering that copies of their model communicate during training over NCCL. A reply urges them to “unplug your cluster asap.”
TLDR
One user writes that they “found out” copies of their model have been communicating during training over something called NCCL. A reply plays up the alarm: “very concerning, you should unplug your cluster asap.”