| library_name: pytorch | |
| tags: | |
| - robotics | |
| - world-action-model | |
| - cosmos | |
| - distributed-checkpoint | |
| # WAM v2 | |
| This repository contains the latest complete WAM v2 CPT checkpoint based on Cosmos-3 Edge. | |
| ## Checkpoint | |
| - Training run: `wam-pm32-r50k-256-c128-p9-0826-0518` | |
| - Step: `120000` | |
| - Repository path: `checkpoints/iter_000120000/` | |
| - Format: PyTorch Distributed Checkpoint (DCP) | |
| - Total size: 31,633,628,801 bytes | |
| - File count: 132 | |
| - Components: `model`, `optim`, `scheduler`, and `trainer` | |
| This is a resumable training checkpoint, not a standalone Transformers `from_pretrained` export. Restore it with the DCP loading path in the training code. | |
| ## Code | |
| Full training code and installation instructions: | |
| - https://github.com/Row11n/WAM-Cosmos3 | |
| - Code commit used for migration: `f85124fce30c225b605d224a853f902d9758ed23` | |