> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://fr.nvidia-localization.ferndocs.com/fr/nemo/rl/latest/nemo-rl-models-generation-vllm-vllm-generation/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://fr.nvidia-localization.ferndocs.com/_mcp/server.

# nemo\_rl.models.generation.vllm.vllm\_generation

## Contenu du module

### Classes

| Nom                                                                         | Description                                                              |
| --------------------------------------------------------------------------- | ------------------------------------------------------------------------ |
| [`VllmGeneration`](#nemorlmodelsgenerationvllmvllmgenerationvllmgeneration) | Classe de base abstraite définissant l'interface pour les politiques RL. |

### Données

`TOP_K_THRESHOLD`
`TOP_P_THRESHOLD`

### API

```python
nemo_rl.models.generation.vllm.vllm_generation.TOP_K_THRESHOLD
```

**Valeur**: `8000`

```python
nemo_rl.models.generation.vllm.vllm_generation.TOP_P_THRESHOLD
```

**Valeur**: `0.99`

```python
class nemo_rl.models.generation.vllm.vllm_generation.VllmGeneration(cluster: nemo_rl.distributed.virtual_cluster.RayVirtualCluster, config: nemo_rl.models.generation.vllm.config.VllmConfig, name_prefix: str = 'vllm_policy', workers_per_node: typing.Optional[typing.Union[int, list[int]]] = None)
```

**Bases**: `nemo_rl.models.generation.interfaces.GenerationInterface`

```python
_get_tied_worker_bundle_indices(cluster: nemo_rl.distributed.virtual_cluster.RayVirtualCluster) -> list[tuple[int, list[int]]]
```

Calculer les indices des groupes de workers pour les workers parallèles tensor et pipeline.

Gère les groupes de placement unifiés (pour le parallélisme de modèle inter-nœuds) et
les groupes de placement par nœud (pour le parallélisme de modèle local au nœud).

```python
_report_device_id() -> list[list[str]]
```

Rapporter l'ID du dispositif des workers vllm.

```python
_post_init()
```

```python
init_collective(
    ip: str,
    port: int,
    world_size: int
) -> list[ray.ObjectRef]
```

Initialiser la communication collective.

\[... Rest of the document remains the same, only text translations have been applied ...]