Demystifying Workload Imbalances in Large Transformer Model Training over Variable-length Sequences

Open in new window