30 lines
1.3 KiB
Markdown
30 lines
1.3 KiB
Markdown
# Checkpoints
|
|
|
|
ChronoEdit supports **three** checkpoint formats, each optimized for a different use case.
|
|
|
|
### 1) DCP checkpoints
|
|
**DCP (Diffusion Checkpoint Package)** is a unified, sharded model bundle designed for **fast, distributed** saving/loading on multi-GPU/multi-node setups.
|
|
We use **DCP** as the **default training checkpoint** format due to speed and robustness in distributed environments.
|
|
|
|
### 2) Consolidated `.pth` checkpoints
|
|
We provide a script to convert a DCP checkpoint into a single, regular PyTorch **`.pth`** checkpoint (convenient for single-node inference or simple export).
|
|
|
|
**Example:**
|
|
```bash
|
|
python scripts/convert_distcp_to_pt.py \
|
|
"./checkpoints/ChronoEdit-14B/iter_000010000/model" \
|
|
"./checkpoints/ChronoEdit-14B/nvidia/chronoedit_14B"
|
|
```
|
|
scripts/run_inference.py accepts both DCP checkpoints and consolidated .pth checkpoints.
|
|
|
|
### 3) Diffusers checkpoints
|
|
For Hugging Face Diffusers workflows, convert from DCP to the Diffusers format (directory with model, scheduler, and config files).
|
|
|
|
**Example:**
|
|
```bash
|
|
python scripts/convert_distcp_to_diffusers.py \
|
|
--src "./checkpoints/ChronoEdit-14B/iter_000010000/model" \
|
|
--dst "./checkpoints/ChronoEdit-14B-Diffusers"
|
|
```
|
|
the Diffusers pipeline only accepts Diffusers-format checkpoints.
|