> This page is for version Nightly (default).
> For other versions, use one of these documentation indexes:
> - Nightly (default): https://docs.nvidia.com/nemo/automodel/nightly/llms.txt
> - Latest: https://docs.nvidia.com/nemo/automodel/latest/llms.txt
> - 0.7.0 · 26.10: https://docs.nvidia.com/nemo/automodel/v0.7/llms.txt
> - 0.5.0 · 26.06: https://docs.nvidia.com/nemo/automodel/v0.5/llms.txt
> - 0.4.0 · 26.04: https://docs.nvidia.com/nemo/automodel/v0.4/llms.txt

> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/nemo/automodel/llms.txt.

# nemo_automodel.components.models.praxis.config

Typed view of the Praxis checkpoint `config.json`.

The checkpoint config is a nested JSON document. The sections the model consumes are
parsed into frozen dataclasses with explicit fields; unknown keys raise so that a
checkpoint of another model variant is refused instead of silently misread.

## Module Contents

### Classes

| Name                                                                                         | Description                                                                     |
| -------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------- |
| [`ActionExpertConfig`](#nemo_automodel-components-models-praxis-config-ActionExpertConfig)   | `action_expert`: the small transformer that denoises the actions.               |
| [`ActionSpaceConfig`](#nemo_automodel-components-models-praxis-config-ActionSpaceConfig)     | `action_space`: the per-frame action row layout.                                |
| [`AttentionConfig`](#nemo_automodel-components-models-praxis-config-AttentionConfig)         | `backbone.attention`.                                                           |
| [`BackboneConfig`](#nemo_automodel-components-models-praxis-config-BackboneConfig)           | `backbone`: the two-tower video transformer.                                    |
| [`ChunkingConfig`](#nemo_automodel-components-models-praxis-config-ChunkingConfig)           | `chunking`: the window geometry in raw frames and latents.                      |
| [`ExpertRopeConfig`](#nemo_automodel-components-models-praxis-config-ExpertRopeConfig)       | `action_expert.rope`: the expert reuses the trunk rotary positions.             |
| [`FlexKernelOptions`](#nemo_automodel-components-models-praxis-config-FlexKernelOptions)     | `action_expert.flex_kernel_options`: pinned flex-attention template parameters. |
| [`MoeSectionConfig`](#nemo_automodel-components-models-praxis-config-MoeSectionConfig)       | `backbone.moe`: a DeepSeek-style sigmoid router with a shared expert.           |
| [`NormalizationConfig`](#nemo_automodel-components-models-praxis-config-NormalizationConfig) | `normalization`: the q01/q99 action statistics and the proprio mean/std.        |
| [`PraxisConfig`](#nemo_automodel-components-models-praxis-config-PraxisConfig)               | The parsed `config.json` of a Praxis single-arm checkpoint.                     |
| [`PraxisConfigError`](#nemo_automodel-components-models-praxis-config-PraxisConfigError)     | Raised when `config.json` is not a supported Praxis single-arm config.          |
| [`ProjectionConfig`](#nemo_automodel-components-models-praxis-config-ProjectionConfig)       | One entry of `input_projections` / `output_projections`.                        |
| [`RopeConfig`](#nemo_automodel-components-models-praxis-config-RopeConfig)                   | `backbone.rope`: integer-position rotary embedding of the trunk.                |
| [`TextEncoderConfig`](#nemo_automodel-components-models-praxis-config-TextEncoderConfig)     | `text_encoder`: the frozen Qwen text encoder and the unconditional caption.     |
| [`TowerConfig`](#nemo_automodel-components-models-praxis-config-TowerConfig)                 | `backbone.encoder` or `backbone.decoder`.                                       |
| [`VideoConfig`](#nemo_automodel-components-models-praxis-config-VideoConfig)                 | `video`: the canvas of the camera views and the VAE geometry.                   |

### Functions

| Name                                                                                 | Description                                                                 |
| ------------------------------------------------------------------------------------ | --------------------------------------------------------------------------- |
| [`_dataclass_from`](#nemo_automodel-components-models-praxis-config-_dataclass_from) | Build `cls` from `section` with its field names; unknown keys are an error. |
| [`_section`](#nemo_automodel-components-models-praxis-config-_section)               | -                                                                           |

### Data

[`FORMAT_VERSION`](#nemo_automodel-components-models-praxis-config-FORMAT_VERSION)

[`JsonScalar`](#nemo_automodel-components-models-praxis-config-JsonScalar)

[`JsonValue`](#nemo_automodel-components-models-praxis-config-JsonValue)

[`MODEL_TYPE`](#nemo_automodel-components-models-praxis-config-MODEL_TYPE)

[`_T`](#nemo_automodel-components-models-praxis-config-_T)

### API

```python
class nemo_automodel.components.models.praxis.config.ActionExpertConfig(
    dim: int,
    num_layers: int,
    num_heads: int,
    head_dim: int,
    trunk_dim: int,
    action_patch_dim: int,
    proprio_patch_dim: int,
    num_future_per_chunk: int,
    max_chunks: int,
    max_action_tokens: int,
    rope: nemo_automodel.components.models.praxis.config.ExpertRopeConfig,
    global_order: tuple[str, ...] = ('register', 'anchor', 'text'),
    tap_norm: bool = True,
    clean_only_conditioning: bool = False,
    rope_cross: bool = False,
    rope_self: bool = False,
    rope_base: float = 100.0,
    absolute_positions: bool = True,
    action_pos_table: bool = False,
    future_chunk_offset: typing.Literal['zero'] = 'zero',
    attention: typing.Literal['sdpa', 'flex_dynamic'] = 'sdpa',
    flex_kernel_options: nemo_automodel.components.models.praxis.config.FlexKernelOptions | None = None,
    head_plus_zero: bool = False,
    layout_stamp: int | None = None
)
```

Dataclass

`action_expert`: the small transformer that denoises the actions.

**`absolute_positions`** `bool = True`

---

**`action_patch_dim`** `int`

---

**`action_pos_table`** `bool = False`

---

**`attention`** `Literal['sdpa', 'flex_dynamic'] = 'sdpa'`

---

**`clean_only_conditioning`** `bool = False`

---

**`dim`** `int`

---

**`flex_kernel_options`** `FlexKernelOptions | None = None`

---

**`future_chunk_offset`** `Literal['zero'] = 'zero'`

---

**`global_order`** `tuple[str, ...] = ('register', 'anchor', 'text')`

---

**`head_dim`** `int`

---

**`head_plus_zero`** `bool = False`

---

**`layout_stamp`** `int | None = None`

---

**`max_action_tokens`** `int`

---

**`max_chunks`** `int`

---

**`num_future_per_chunk`** `int`

---

**`num_heads`** `int`

---

**`num_layers`** `int`

---

**`proprio_patch_dim`** `int`

---

**`rope`** `ExpertRopeConfig`

---

**`rope_base`** `float = 100.0`

---

**`rope_cross`** `bool = False`

---

**`rope_self`** `bool = False`

---

**`tap_norm`** `bool = True`

---

**`trunk_dim`** `int`

---

```python
nemo_automodel.components.models.praxis.config.ActionExpertConfig.from_json(
    section: collections.abc.Mapping[str, nemo_automodel.components.models.praxis.config.JsonValue]
) -> nemo_automodel.components.models.praxis.config.ActionExpertConfig
```

classmethod

```python
class nemo_automodel.components.models.praxis.config.ActionSpaceConfig(
    dim: int,
    layout: str,
    pose_layout: str
)
```

Dataclass

`action_space`: the per-frame action row layout.

**`dim`** `int`

---

**`layout`** `str`

---

**`pose_layout`** `str`

---

```python
class nemo_automodel.components.models.praxis.config.AttentionConfig(
    num_heads: int,
    num_kv_heads: int,
    head_dim: int,
    gated: bool = True,
    qk_norm_eps: float = 1e-06,
    qk_norm_compute_dtype: str = 'float32',
    rope_after_norm: bool = False,
    flex_block_size: int = 128,
    backend: typing.Literal['triton', 'flash'] = 'triton'
)
```

Dataclass

`backbone.attention`.

**`backend`** `Literal['triton', 'flash'] = 'triton'`

---

**`flex_block_size`** `int = 128`

---

**`gated`** `bool = True`

---

**`head_dim`** `int`

---

**`num_heads`** `int`

---

**`num_kv_heads`** `int`

---

**`qk_norm_compute_dtype`** `str = 'float32'`

---

**`qk_norm_eps`** `float = 1e-06`

---

**`rope_after_norm`** `bool = False`

---

```python
class nemo_automodel.components.models.praxis.config.BackboneConfig(
    dim: int,
    timestep_dim: int,
    num_registers: int,
    attention: nemo_automodel.components.models.praxis.config.AttentionConfig,
    rope: nemo_automodel.components.models.praxis.config.RopeConfig,
    moe: nemo_automodel.components.models.praxis.config.MoeSectionConfig,
    dense_ffn_hidden: int,
    encoder: nemo_automodel.components.models.praxis.config.TowerConfig,
    decoder: nemo_automodel.components.models.praxis.config.TowerConfig,
    modulation_dtype: typing.Literal['bfloat16', 'float32', 'mixed'] = 'mixed',
    norm_eps: float = 1e-06,
    norm_compute_dtype: str = 'float32',
    layout: typing.Literal['chunk_causal'] = 'chunk_causal'
)
```

Dataclass

`backbone`: the two-tower video transformer.

**`attention`** `AttentionConfig`

---

**`decoder`** `TowerConfig`

---

**`dense_ffn_hidden`** `int`

---

**`dim`** `int`

---

**`encoder`** `TowerConfig`

---

**`layout`** `Literal['chunk_causal'] = 'chunk_causal'`

---

**`modulation_dtype`** `Literal['bfloat16', 'float32', 'mixed'] = 'mixed'`

---

**`moe`** `MoeSectionConfig`

---

**`norm_compute_dtype`** `str = 'float32'`

---

**`norm_eps`** `float = 1e-06`

---

**`num_registers`** `int`

---

**`rope`** `RopeConfig`

---

**`timestep_dim`** `int`

---

```python
nemo_automodel.components.models.praxis.config.BackboneConfig.from_json(
    section: collections.abc.Mapping[str, nemo_automodel.components.models.praxis.config.JsonValue]
) -> nemo_automodel.components.models.praxis.config.BackboneConfig
```

classmethod

```python
class nemo_automodel.components.models.praxis.config.ChunkingConfig(
    n_chunks: int,
    action_chunk: int,
    action_frames_per_token: int,
    latents_per_chunk: int,
    past_frames: int,
    future_frames: int,
    window_frames: int,
    chunk_geometry: tuple[tuple[int, int], ...],
    grid_hz: float,
    global_frame_anchor: bool = True,
    contiguous_chunks: bool = True,
    context_format: str = 'vxt',
    proprio_frames: int = 1
)
```

Dataclass

`chunking`: the window geometry in raw frames and latents.

**`action_chunk`** `int`

---

**`action_frames_per_token`** `int`

---

**`chunk_geometry`** `tuple[tuple[int, int], ...]`

---

**`context_format`** `str = 'vxt'`

---

**`contiguous_chunks`** `bool = True`

---

**`future_frames`** `int`

---

**`global_frame_anchor`** `bool = True`

---

**`grid_hz`** `float`

---

**`latents_per_chunk`** `int`

---

**`n_chunks`** `int`

---

**`past_frames`** `int`

---

**`proprio_frames`** `int = 1`

---

**`window_frames`** `int`

---

```python
nemo_automodel.components.models.praxis.config.ChunkingConfig.from_json(
    section: collections.abc.Mapping[str, nemo_automodel.components.models.praxis.config.JsonValue]
) -> nemo_automodel.components.models.praxis.config.ChunkingConfig
```

classmethod

```python
class nemo_automodel.components.models.praxis.config.ExpertRopeConfig(
    kind: typing.Literal['trunk'],
    dim: int,
    effective_dim: int,
    base: int,
    table_size: int,
    order: typing.Literal['before_norm']
)
```

Dataclass

`action_expert.rope`: the expert reuses the trunk rotary positions.

**`base`** `int`

---

**`dim`** `int`

---

**`effective_dim`** `int`

---

**`kind`** `Literal['trunk']`

---

**`order`** `Literal['before_norm']`

---

**`table_size`** `int`

---

```python
class nemo_automodel.components.models.praxis.config.FlexKernelOptions(
    fwd_BLOCK_M: int,
    fwd_BLOCK_N: int,
    num_warps: int,
    num_stages: int
)
```

Dataclass

`action_expert.flex_kernel_options`: pinned flex-attention template parameters.

**`fwd_BLOCK_M`** `int`

---

**`fwd_BLOCK_N`** `int`

---

**`num_stages`** `int`

---

**`num_warps`** `int`

---

```python
class nemo_automodel.components.models.praxis.config.MoeSectionConfig(
    num_routed_experts: int,
    top_k_routed: int,
    expert_hidden: int,
    num_experts: int | None = None,
    shared_expert: bool = True,
    top_k: int | None = None,
    router_activation: typing.Literal['sigmoid'] = 'sigmoid',
    router_dtype: str = 'float32',
    routed_scaling_factor: float = 2.5,
    no_expert_slots: bool = False,
    routed_weight_layout: typing.Literal['expert_first'] = 'expert_first',
    dispatch_chunk_size: int | None = None
)
```

Dataclass

`backbone.moe`: a DeepSeek-style sigmoid router with a shared expert.

**`dispatch_chunk_size`** `int | None = None`

---

**`expert_hidden`** `int`

---

**`no_expert_slots`** `bool = False`

---

**`num_experts`** `int | None = None`

---

**`num_routed_experts`** `int`

---

**`routed_scaling_factor`** `float = 2.5`

---

**`routed_weight_layout`** `Literal['expert_first'] = 'expert_first'`

---

**`router_activation`** `Literal['sigmoid'] = 'sigmoid'`

---

**`router_dtype`** `str = 'float32'`

---

**`shared_expert`** `bool = True`

---

**`top_k`** `int | None = None`

---

**`top_k_routed`** `int`

---

```python
nemo_automodel.components.models.praxis.config.MoeSectionConfig.__post_init__() -> None
```

```python
class nemo_automodel.components.models.praxis.config.NormalizationConfig(
    domain: str,
    action_q01: tuple[float, ...],
    action_q99: tuple[float, ...],
    q99_normed_dims: tuple[int, ...],
    proprio_mean: tuple[float, ...],
    proprio_std: tuple[float, ...],
    action_formula: str,
    proprio_formula: str
)
```

Dataclass

`normalization`: the q01/q99 action statistics and the proprio mean/std.

**`action_formula`** `str`

---

**`action_q01`** `tuple[float, ...]`

---

**`action_q99`** `tuple[float, ...]`

---

**`domain`** `str`

---

**`proprio_formula`** `str`

---

**`proprio_mean`** `tuple[float, ...]`

---

**`proprio_std`** `tuple[float, ...]`

---

**`q99_normed_dims`** `tuple[int, ...]`

---

```python
class nemo_automodel.components.models.praxis.config.PraxisConfig(
    backbone: nemo_automodel.components.models.praxis.config.BackboneConfig,
    action_expert: nemo_automodel.components.models.praxis.config.ActionExpertConfig,
    chunking: nemo_automodel.components.models.praxis.config.ChunkingConfig,
    video: nemo_automodel.components.models.praxis.config.VideoConfig,
    text_encoder: nemo_automodel.components.models.praxis.config.TextEncoderConfig,
    action_space: nemo_automodel.components.models.praxis.config.ActionSpaceConfig,
    normalization: nemo_automodel.components.models.praxis.config.NormalizationConfig,
    format_version: int = FORMAT_VERSION,
    model_type: str = MODEL_TYPE,
    embodiment: typing.Literal['single_arm'] = 'single_arm',
    torch_dtype: str = 'bfloat16',
    fp32_tensors: tuple[str, ...] = ('*.ffn.expert_bias',),
    pose_frame: typing.Literal['flange'] = 'flange'
)
```

Dataclass

The parsed `config.json` of a Praxis single-arm checkpoint.

**`action_expert`** `ActionExpertConfig`

---

**`action_space`** `ActionSpaceConfig`

---

**`backbone`** `BackboneConfig`

---

**`chunking`** `ChunkingConfig`

---

**`embodiment`** `Literal['single_arm'] = 'single_arm'`

---

**`format_version`** `int = FORMAT_VERSION`

---

**`fp32_tensors`** `tuple[str, ...] = ('*.ffn.expert_bias',)`

---

**`model_type`** `str = MODEL_TYPE`

---

**`normalization`** `NormalizationConfig`

---

**`pose_frame`** `Literal['flange'] = 'flange'`

---

**`text_encoder`** `TextEncoderConfig`

---

**`torch_dtype`** `str = 'bfloat16'`

---

**`video`** `VideoConfig`

---

```python
nemo_automodel.components.models.praxis.config.PraxisConfig.from_json(
    config: collections.abc.Mapping[str, nemo_automodel.components.models.praxis.config.JsonValue]
) -> nemo_automodel.components.models.praxis.config.PraxisConfig
```

classmethod

Parse and validate a checkpoint config document.

```python
nemo_automodel.components.models.praxis.config.PraxisConfig.from_pretrained(
    path: str | pathlib.Path
) -> nemo_automodel.components.models.praxis.config.PraxisConfig
```

classmethod

Read and parse `<path>/config.json`.

```python
class nemo_automodel.components.models.praxis.config.PraxisConfigError()
```

**Bases:** `ValueError`

Raised when `config.json` is not a supported Praxis single-arm config.

```python
class nemo_automodel.components.models.praxis.config.ProjectionConfig(
    c_in: int | None = None,
    c_out: int | None = None,
    move_dim: bool = False,
    timestep_modulated: bool = False
)
```

Dataclass

One entry of `input_projections` / `output_projections`.

**`c_in`** `int | None = None`

---

**`c_out`** `int | None = None`

---

**`move_dim`** `bool = False`

---

**`timestep_modulated`** `bool = False`

---

```python
class nemo_automodel.components.models.praxis.config.RopeConfig(
    base: int = 26000000,
    dim: int = 128,
    effective_dim: int = 64,
    max_seq_len: int = 90000000,
    efficient_cache_max: int = 1000000,
    max_intra: int = 360,
    max_inter: int = 129600,
    chunked_video_layout: bool = True,
    n_chunks: int | None = None,
    time_scale: float = 1.0,
    n_tiles: int | None = None,
    ref_placement: typing.Literal['group'] = 'group'
)
```

Dataclass

`backbone.rope`: integer-position rotary embedding of the trunk.

**`base`** `int = 26000000`

---

**`chunked_video_layout`** `bool = True`

---

**`dim`** `int = 128`

---

**`effective_dim`** `int = 64`

---

**`efficient_cache_max`** `int = 1000000`

---

**`max_inter`** `int = 129600`

---

**`max_intra`** `int = 360`

---

**`max_seq_len`** `int = 90000000`

---

**`n_chunks`** `int | None = None`

---

**`n_tiles`** `int | None = None`

---

**`ref_placement`** `Literal['group'] = 'group'`

---

**`time_scale`** `float = 1.0`

---

```python
class nemo_automodel.components.models.praxis.config.TextEncoderConfig(
    model: str,
    hidden_layer: int,
    pad_to_multiple_of: int,
    trim_padding: bool,
    hidden_dim: int,
    uncond: typing.Literal['blank_caption'] = 'blank_caption',
    blank_caption: str = ' '
)
```

Dataclass

`text_encoder`: the frozen Qwen text encoder and the unconditional caption.

**`blank_caption`** `str = ' '`

---

**`hidden_dim`** `int`

---

**`hidden_layer`** `int`

---

**`model`** `str`

---

**`pad_to_multiple_of`** `int`

---

**`trim_padding`** `bool`

---

**`uncond`** `Literal['blank_caption'] = 'blank_caption'`

---

```python
class nemo_automodel.components.models.praxis.config.TowerConfig(
    depth: int,
    dense_blocks: tuple[int, ...],
    text_in: nemo_automodel.components.models.praxis.config.ProjectionConfig | None,
    video_in: nemo_automodel.components.models.praxis.config.ProjectionConfig,
    video_out: nemo_automodel.components.models.praxis.config.ProjectionConfig,
    tap_blocks: tuple[int, ...] = (),
    modulation_dim: int | None = None,
    unpatch_factor: tuple[int, int, int] = (1, 1, 1),
    positive_zero_residual_add: bool = True
)
```

Dataclass

`backbone.encoder` or `backbone.decoder`.

**`dense_blocks`** `tuple[int, ...]`

---

**`depth`** `int`

---

**`modulation_dim`** `int | None = None`

---

**`positive_zero_residual_add`** `bool = True`

---

**`tap_blocks`** `tuple[int, ...] = ()`

---

**`text_in`** `ProjectionConfig | None`

---

**`unpatch_factor`** `tuple[int, int, int] = (1, 1, 1)`

---

**`video_in`** `ProjectionConfig`

---

**`video_out`** `ProjectionConfig`

---

```python
nemo_automodel.components.models.praxis.config.TowerConfig.from_json(
    section: collections.abc.Mapping[str, nemo_automodel.components.models.praxis.config.JsonValue],
    where: str
) -> nemo_automodel.components.models.praxis.config.TowerConfig
```

classmethod

```python
class nemo_automodel.components.models.praxis.config.VideoConfig(
    height: int,
    width: int,
    views: tuple[str, ...],
    layout: str,
    tile_bounds_rows_cols: dict[str, tuple[int, int, int, int]],
    vae: str,
    vae_encode_chunk_frames: int,
    latent_channels: int,
    vae_stride: tuple[int, int, int],
    patch_size: tuple[int, int, int],
    tile_height: int = 256,
    tile_width: int = 416,
    canvas: str = 'tiles',
    keep_aspect: bool = True,
    antialias: bool = False,
    aspect_tol: float = 0.03,
    vae_encode_impl: str = 'megakernel',
    vae_megakernel_chunk_frames: int | None = None
)
```

Dataclass

`video`: the canvas of the camera views and the VAE geometry.

**`antialias`** `bool = False`

---

**`aspect_tol`** `float = 0.03`

---

**`canvas`** `str = 'tiles'`

---

**`height`** `int`

---

**`keep_aspect`** `bool = True`

---

**`latent_channels`** `int`

---

**`layout`** `str`

---

**`patch_size`** `tuple[int, int, int]`

---

**`tile_bounds_rows_cols`** `dict[str, tuple[int, int, int, int]]`

---

**`tile_height`** `int = 256`

---

**`tile_width`** `int = 416`

---

**`vae`** `str`

---

**`vae_encode_chunk_frames`** `int`

---

**`vae_encode_impl`** `str = 'megakernel'`

---

**`vae_megakernel_chunk_frames`** `int | None = None`

---

**`vae_stride`** `tuple[int, int, int]`

---

**`views`** `tuple[str, ...]`

---

**`width`** `int`

---

```python
nemo_automodel.components.models.praxis.config.VideoConfig.from_json(
    section: collections.abc.Mapping[str, nemo_automodel.components.models.praxis.config.JsonValue]
) -> nemo_automodel.components.models.praxis.config.VideoConfig
```

classmethod

```python
nemo_automodel.components.models.praxis.config._dataclass_from(
    cls: type[nemo_automodel.components.models.praxis.config._T],
    section: collections.abc.Mapping[str, nemo_automodel.components.models.praxis.config.JsonValue],
    where: str
) -> nemo_automodel.components.models.praxis.config._T
```

Build `cls` from `section` with its field names; unknown keys are an error.

```python
nemo_automodel.components.models.praxis.config._section(
    parent: collections.abc.Mapping[str, nemo_automodel.components.models.praxis.config.JsonValue],
    name: str
) -> collections.abc.Mapping[str, nemo_automodel.components.models.praxis.config.JsonValue]
```

```python
nemo_automodel.components.models.praxis.config.FORMAT_VERSION = 2
```

```python
nemo_automodel.components.models.praxis.config.JsonScalar = str | int | float | bool | None
```

```python
nemo_automodel.components.models.praxis.config.JsonValue = JsonScalar | list['JsonValue'] | dict[str, 'JsonValue']
```

```python
nemo_automodel.components.models.praxis.config.MODEL_TYPE = 'praxis'
```

```python
nemo_automodel.components.models.praxis.config._T = TypeVar('_T')
```