| # Reassembling the checkpoint |
|
|
| The full training checkpoint `m1_mix_final_step50000.pt` (15.3 GB) is uploaded as **9 |
| byte-split parts**. The upload path (an egress gateway) both capped single files at |
| ~9.7 GB *and* throttled cumulative bytes per connection window, so the file is split into |
| small parts that each fit a window: the **first two parts are 4 GiB** (uploaded before the |
| throttle was understood) and the **remaining seven are ≤1 GiB**. Concatenating all parts |
| in sorted order reproduces the checkpoint **bit-for-bit**. |
|
|
| | part | size | |
| |-----------------------------------------|------------| |
| | `m1_mix_final_step50000.pt.part00` | 4.0 GiB | |
| | `m1_mix_final_step50000.pt.part01` | 4.0 GiB | |
| | `m1_mix_final_step50000.pt.part02` | 1.0 GiB | |
| | `m1_mix_final_step50000.pt.part03` | 1.0 GiB | |
| | `m1_mix_final_step50000.pt.part04` | 1.0 GiB | |
| | `m1_mix_final_step50000.pt.part05` | 1.0 GiB | |
| | `m1_mix_final_step50000.pt.part06` | 1.0 GiB | |
| | `m1_mix_final_step50000.pt.part07` | 1.0 GiB | |
| | `m1_mix_final_step50000.pt.part08` | ~0.25 GiB | |
|
|
| ## 1. Reassemble |
|
|
| ```bash |
| cat m1_mix_final_step50000.pt.part?? > m1_mix_final_step50000.pt |
| ``` |
|
|
| (`part??` expands to `part00 part01 … part08` in sorted order — the command is unchanged |
| regardless of how many parts there are.) |
|
|
| ## 2. Verify integrity |
|
|
| ```bash |
| sha256sum -c m1_mix_final_step50000.pt.sha256 |
| # expected: m1_mix_final_step50000.pt: OK |
| ``` |
|
|
| The expected SHA-256 of the full file is: |
|
|
| ``` |
| df149a0d27c85ad76217552c4005bee0094f6068fdb92bd9f6b256190f5aa09a |
| ``` |
|
|
| ## 3. Load |
|
|
| Load as usual — the deploy loader reads `payload["model_state_dict"]` and calls |
| `load_state_dict(..., strict=False)`. See the top-level `README.md` for the full |
| inference / evaluation instructions and the optional inference-only slimming snippet. |
|
|
