List your company

LLM Training Optimization on Llama 3.2 3B

12.3Training Time Reduction (%)Kareus
-1.9481.7515.459.149Jan 25, 2026

Evaluation results

MethodLinks
KareusParallelism=TP8, µBatch Size=8, Sequence Length=4K2026.0112.319.6——
Nanobatching + PerseusParallelism=TP8, µBatch Size=8, Sequence Length=4K2026.018.515.5——
KareusParallelism=CP2TP4, µBatch Size=16, Sequence Length=4K2026.01816.2——
KareusParallelism=CP2TP4, µBatch Size=8, Sequence Length=8K2026.016.214.9——
KareusParallelism=CP2TP4, µBatch Size=8, Sequence Length=4K2026.015.214.4——
Nanobatching + PerseusParallelism=CP2TP4, µBatch Size=16, Sequence Length=4K2026.012.47——
Nanobatching + PerseusParallelism=CP2TP4, µBatch Size=8, Sequence Length=8K2026.011.57.3——
Megatron-LM + PerseusParallelism=CP2TP4, µBatch Size=8, Sequence Length=4K2026.010.27.4——
Nanobatching + PerseusParallelism=CP2TP4, µBatch Size=8, Sequence Length=4K2026.0107.3——
Megatron-LM + PerseusParallelism=TP8, µBatch Size=8, Sequence Length=4K2026.01-0.310——
Megatron-LM + PerseusParallelism=CP2TP4, µBatch Size=16, Sequence Length=4K2026.01-0.78.9——
Megatron-LM + PerseusParallelism=CP2TP4, µBatch Size=8, Sequence Length=8K2026.01-1.411.5——
KareusParallelism Strategy=TP8, Number of Microbatches=8, Sequence Length=4K2026.01——24.324
KareusParallelism Strategy=CP2TP4, Number of Microbatches=8, Sequence Length=4K2026.01——17.413.9
KareusParallelism Strategy=CP2TP4, Number of Microbatches=8, Sequence Length=8K2026.01——12.416.4
KareusParallelism Strategy=CP2TP4, Number of Microbatches=16, Sequence Length=4K2026.01——20.412.5
Nanobatching + PerseusParallelism Strategy=TP8, Number of Microbatches=8, Sequence Length=4K2026.01——2118.9
Nanobatching + PerseusParallelism Strategy=CP2TP4, Number of Microbatches=8, Sequence Length=4K2026.01———7
Nanobatching + PerseusParallelism Strategy=CP2TP4, Number of Microbatches=8, Sequence Length=8K2026.01——44.4
Nanobatching + PerseusParallelism Strategy=CP2TP4, Number of Microbatches=16, Sequence Length=4K2026.01——6.65.1