Mycroft adds collective-communication-level tracing to NCCL so that slow or stuck data transfers in LLM training can be detected and traced to likely faulty ranks in seconds.
In 22nd USENIX Symposium on Networked Systems Design and Implementation (NSDI 25)
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.DC 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Mycroft: Tracing Dependencies in Collective Communication Towards Reliable LLM Training
Mycroft adds collective-communication-level tracing to NCCL so that slow or stuck data transfers in LLM training can be detected and traced to likely faulty ranks in seconds.