LLM Training on Llama 70B Emulation 3.3
9.3Time ReductionKareus
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Kareus# Microbatches=16, Throughput Configuration=max-throughput2026.01 | 9.3 | 20.2 | |
| Kareus# Microbatches=32, Throughput Configuration=max-throughput2026.01 | 9.2 | 20 | |
| Kareus# Microbatches=64, Throughput Configuration=max-throughput2026.01 | 9.1 | 19.8 | |
| Kareus# Microbatches=128, Throughput Configuration=max-throughput2026.01 | 9.1 | 19.7 | |
| Megatron-LM + Perseus# Microbatches=16, Throughput Configuration=max-throughput2026.01 | 0 | 15 | |
| Megatron-LM + Perseus# Microbatches=32, Throughput Configuration=max-throughput2026.01 | 0 | 14.3 | |
| Megatron-LM + Perseus# Microbatches=64, Throughput Configuration=max-throughput2026.01 | 0 | 13.8 | |
| Megatron-LM + Perseus# Microbatches=128, Throughput Configuration=max-throughput2026.01 | 0 | 13.5 |