# openpi drop-in — INSTALL

These files let you **serve** (and retrain) the `pi05_g1_inspire_piston_lora_abs_st`
checkpoint in this repo with the [openpi](https://github.com/Physical-Intelligence/openpi)
codebase.

## Our fork state (be honest about this)

This checkpoint was produced from a **fork** of openpi:

```
upstream:  Physical-Intelligence/openpi @ 15a9616
           ("update output objects to support batching", #975)
+ 5 local commits (HEAD = 9088f209):
    99456de  scripts: in-process pi0.5 -> GR00T-WBC-Bridge ZMQ adapter (piston)
    b08d778  g1-inspire: land the G1+Inspire pi0.5 fine-tune baseline (redball + piston v1)
    2838d42  training: state-dropout (p=0.8) retrain config for the piston LoRA fine-tune
    4297d2c  training: absolute-action retrain config for the piston LoRA fine-tune
    9088f20  training: closure-observability retrain config (discrete state tokens + omission dropout)
```

`git diff --stat 15a9616 HEAD` touches exactly these files:

```
 scripts/zmq_adapter_pi05.py              | 376 +   (new — GR00T ZMQ serving adapter)
 src/openpi/policies/g1_inspire_policy.py |  94 +   (new — G1Inspire transforms)
 src/openpi/training/checkpoints.py       |   2 +-  (checkpoint retention, train-only)
 src/openpi/training/config.py            | 316 +   (piston configs + LeRobotG1InspireDataConfig + 2 DataConfig fields)
 src/openpi/training/data_loader.py       |  27 +   (state/token dropout injection, train-only)
 src/openpi/transforms.py                 | 102 +-  (StateDropout, StateTokenDropout, TokenizePrompt state-omission path)
```

## What SERVING actually needs (from a clean upstream 15a9616 checkout)

Good news: the `discrete_state_input=True` behaviour this checkpoint relies on
— discretizing the normalized 29-D proprio state into 256 bins and injecting it
into the prompt (`"Task: ..., State: <ints>;\nAction: "`) — is **already
upstream** in `src/openpi/models/tokenizer.py` and in `TokenizePrompt`
(the `discrete_state_input` flag is native pi0.5). So is the `state=None`
prompt format. You do **not** need any tokenizer patch to serve.

The load-bearing **local** pieces for serving are only:

1. **`src/openpi/policies/g1_inspire_policy.py`** — copy `g1_inspire_policy.py`
   from this folder. Defines `G1InspireInputs` / `G1InspireOutputs`
   (key remap, 29-D proprio slice, 30-D action truncation).
2. **`LeRobotG1InspireDataConfig`** — paste the class from `config_entry.py`
   part (2) into `src/openpi/training/config.py` and add
   `from openpi.policies import g1_inspire_policy` to its imports.
3. **The two `DataConfig` fields** — paste `state_dropout_prob` and
   `state_token_dropout_prob` from `config_entry.py` part (3) into the
   `DataConfig` dataclass. Both are train-only, but the config in (1) passes
   `state_token_dropout_prob=0.8`, so the field must **exist** or the config
   won't construct.
4. **The `TrainConfig` entry** — paste `config_entry.py` part (1) into the
   `_CONFIGS` list.

The `StateDropout` / `StateTokenDropout` transforms and the
`data_loader.transform_dataset` dropout injection are **train-only** — not
needed to serve, but you'll want them (from our fork) if you retrain.

The simplest path is to `git clone` our fork state (upstream @ `15a9616` +
cherry-pick `99456de..9088f20`) rather than hand-splice.

## Serving

See the repo `README.md` "Serving" section. In short, from an openpi checkout
with its `.venv`:

```bash
# ckpt_dir = the directory you downloaded this HF repo into (contains params/, assets/, _CHECKPOINT_METADATA)
XLA_PYTHON_CLIENT_PREALLOCATE=false .venv/bin/python scripts/zmq_adapter_pi05.py \
  --config  pi05_g1_inspire_piston_lora_abs_st \
  --ckpt_dir <download_dir> \
  --port 5555
```