cosmos_battery / README.md
m1ku2's picture
upload README
c1df75b verified
|
Raw
History Blame Contribute Delete
2.46 kB
# cosmos_battery
This directory is the staging area for the inference-only package of the Piper14 Cosmos3 battery policy run.
- run: `battery_piper14_cosmos3_nano_20000step_4gpu_b8_acc1_offline`
- source checkpoint format: DCP directory checkpoints
- exported model directories: `12k/`, `16k/`, `18k/`, `20k/`
Use this folder as the package root when preparing an export for handoff or upload. It keeps the inference-facing model directories and the provenance copies that explain where they came from in one place.
## What belongs in the inference-only package
Required:
- exactly one or more exported model directories produced by `cosmos_framework.scripts.export_model`
- `provenance/config.yaml`
- `provenance/job_env.yaml`
- `provenance/launch_info.yaml`
- this README plus `package_manifest.yaml`
## What does not belong here by default
Do not package these into the inference-only artifact unless a downstream environment explicitly requires them:
- raw DCP checkpoint directories under `reports/.../checkpoints/`
- full training dataset
- local Qwen snapshot cache
- local `Wan2.2_VAE.pth`
## Why the source checkpoints are directories
This training run uses distributed Cosmos Framework checkpointing. Each training checkpoint is a directory with sharded state, not a single `.pt` file. The intended inference flow is to export those DCP checkpoints into inference-only model directories and ship the exported directories instead of the raw training checkpoints.
## Package layout
```text
cosmos_battery/
README.md
package_manifest.yaml
provenance/
README.md
config.yaml
job_env.yaml
launch_info.yaml
12k/
16k/
18k/
20k/
```
Notes:
- `12k/`, `16k/`, `18k/`, and `20k/` are complete exported inference checkpoints.
- Each of those directories should contain `config.json`, `checkpoint.json`, `model.safetensors.index.json`, and seven `model-*.safetensors` shards.
## Runtime expectation outside this package
The downstream inference side still needs the local Piper14 registration code from this repository:
- `piper_cosmos/cosmos3/domain.py`
- `piper_cosmos/cosmos3/action_policy_piper14_nano.py`
Those files are runtime code dependencies, not part of this `cosmos_battery/` packaging bundle.
## How to use this folder
1. Verify the chosen checkpoint directory contains exporter-produced inference artifacts.
2. Keep the copied provenance files under `provenance/` with the exported model directory.