physicalai-runtime-configuring-inference-pipeline

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Chinese

Pipeline order: observation → preprocessors → runner → postprocessors → action output. See

docs/how-to/inference/configure-pre-post-processing.md

Core code:

src/physicalai/inference/component_factory.py

—

ComponentRegistry

instantiate_component

_MAX_COMPONENT_DEPTH

```
src/physicalai/inference/model.py
```
— builds processor chains from manifest specs.
Built-ins under
```
preprocessors/
```
and
```
postprocessors/
```
; runners under
```
runners/
```
.

管道顺序：observation → 预处理器 → runner → 后处理器 → 动作输出。详见

docs/how-to/inference/configure-pre-post-processing.md

。

核心代码：

src/physicalai/inference/component_factory.py

—

ComponentRegistry

instantiate_component

_MAX_COMPONENT_DEPTH

Use
class_path
+
init_args
for explicit classes:

yaml

preprocessors:
  - class_path: physicalai.inference.preprocessors.StatsNormalizer
    init_args:
      artifact: stats.safetensors

Add a new built-in processor:
- Implement subclass of
```
Preprocessor
```
  /
```
Postprocessor
```
  in the appropriate package.
- Register a short
```
type
```
  name in
```
component_factory
```
  if manifest-friendly aliases are needed.
- Add unit tests under
```
tests/unit/inference/preprocessors/
```
  or
```
postprocessors/
```
  .
- Done when: manifest using
```
type
```
  or
```
class_path
```
  instantiates in a minimal
```
InferenceModel
```
  test.
Nested components in
```
init_args
```
must stay within
```
_MAX_COMPONENT_DEPTH
```
; avoid cyclic specs.

对于显式类，使用
class_path
+
init_args
：

yaml

preprocessors:
  - class_path: physicalai.inference.preprocessors.StatsNormalizer
    init_args:
      artifact: stats.safetensors

bash

uv run pytest tests/unit/inference/preprocessors tests/unit/inference/postprocessors tests/unit/inference/test_manifest.py -q

bash

uv run pytest tests/unit/inference/preprocessors tests/unit/inference/postprocessors tests/unit/inference/test_manifest.py -q

Processor order matches training/export semantics (normalization before runner, denormalization after).
Artifact file names in manifests do not traverse paths (
```
..
```
, absolute paths).
New public processors appear in
```
docs/reference/inference-api.md
```
or how-to docs when user-visible.
Runner choice (
```
SinglePass
```
, chunking runners) stays consistent with
```
predict_action_chunk
```
vs
```
select_action
```
docs.