| tags: | |
| - cua | |
| - megatron-bridge | |
| - checkpoint | |
| - atomic | |
| # CUA Atomic Function, iter 422 | |
| This repository contains the Megatron model-weight initialization checkpoint used by the Complex Function training run. The checkpoint was imported on DGX from the 3x-full Function Hugging Face export at `iter_0000422`. An import/export comparison found all 738 tensors (9,319,730,176 tensor bytes) equal to that source export. | |
| The local checkpoint directory is named `iter_0000000` because Complex training started a new optimization run. It does not contain the original 3x-full optimizer, scheduler, or RNG state. It is suitable as the exact model-weight starting point for Complex Function, not as a resume point for the original 3x-full training run. | |