pytrio.ServiceClient

class ServiceClient:
    def __init__(
        self,
        api_key: str | None = None,
    ):

ServiceClient is the main entry point for PyTRIO APIs. It creates:

TrainingClient instances for model training.
SamplingClient instances for generation and inference.
RestClient instances for REST operations.

import pytrio as trio

client = trio.ServiceClient()

training_client = client.create_lora_training_client(base_model="Qwen/Qwen3.5-4B")
sampling_client = client.create_sampling_client(base_model="Qwen/Qwen3.5-4B")
rest_client = client.create_rest_client()

Parameters

Parameter	Type	Default	Description
`api_key`	`str \| None`	`None`	API key. If omitted, PyTRIO reads the current process configuration, which can come from the local login state, `PYTRIO_API_KEY`, or `pytrio.configure(api_key=...)`; see Environment Configuration

During initialization, ServiceClient verifies login, opens the socket connection, and fetches the available model list. Each ServiceClient copies the current process configuration when construction begins; later calls to pytrio.configure(...) do not change an existing client's server URL, timeout, or similar settings.

Methods

get_supported_models

def get_supported_models(self) -> list[str]

Return the models currently available to the account.

Returns

list[str] - model names, for example ["Qwen/Qwen3.5-4B"].

Example

models = client.get_supported_models()
print(models)  # ["Qwen/Qwen3.5-4B", ...]

create_lora_training_client

def create_lora_training_client(
    self,
    base_model: str,
    rank: int = 32,
    seed: int | None = None,
    train_mlp: bool = True,
    train_attn: bool = True,
    train_unembed: bool = True,
) -> TrainingClient

Create a TrainingClient for LoRA fine-tuning.

Parameters

Parameter	Type	Default	Description
`base_model`	`str`	-	Base model, for example `"Qwen/Qwen3.5-4B"`
`rank`	`int`	`32`	LoRA rank, from 4 to 64
`seed`	`int \| None`	`None`	Initialization seed
`train_mlp`	`bool`	`True`	Whether to train MLP layers
`train_attn`	`bool`	`True`	Whether to train attention layers
`train_unembed`	`bool`	`True`	Whether to train the `lm_head` layer

Returns

TrainingClient - client with the active training state.

Example

training_client = client.create_lora_training_client(
    base_model="Qwen/Qwen3.5-4B",
    rank=16,
    train_unembed=False,
)

create_sampling_client

def create_sampling_client(
    self,
    base_model: str,
    model_path: str | None = None,
) -> SamplingClient

Create a SamplingClient for generation and inference.

Parameters

Parameter	Type	Default	Description
`base_model`	`str`	`""`	Base model, for example `"Qwen/Qwen3.5-4B"`
`model_path`	`str \| None`	`None`	LoRA model checkpoint path URL to load during initialization

Returns

SamplingClient - sampling client.

Examples

sampling_client = client.create_sampling_client(base_model="Qwen/Qwen3.5-4B")

sampling_client = client.create_sampling_client(
    base_model="Qwen/Qwen3.5-4B",
    model_path="/path/to/checkpoint",
)

create_rest_client

def create_rest_client(self) -> RestClient

Create a RestClient for REST API operations, such as listing adapters and querying checkpoints.

Returns

RestClient - REST client.

Example

rest_client = client.create_rest_client()
checkpoint_list = rest_client.list_user_checkpoints().result()

create_training_client_from_state

def create_training_client_from_state(self, path: str) -> TrainingClient

Create a TrainingClient from a saved adapter or checkpoint without restoring optimizer state.

Parameters

Parameter	Type	Description
`path`	`str`	Adapter or checkpoint path

Returns

TrainingClient - client with restored model state.

Example

training_client = client.create_training_client_from_state(
    path="/path/to/checkpoint"
)

create_training_client_from_state_with_optimizer

def create_training_client_from_state_with_optimizer(self, path: str) -> TrainingClient

Create a TrainingClient from a train checkpoint and restore optimizer state for resume training.

Parameters

Parameter	Type	Description
`path`	`str`	Train checkpoint path

Returns

TrainingClient - client with restored model and optimizer state.

Example

training_client = client.create_training_client_from_state_with_optimizer(
    path="/path/to/checkpoint"
)

Async Methods

create_lora_training_client_async

async def create_lora_training_client_async(
    self,
    base_model: str,
    rank: int = 32,
    seed: int | None = None,
    train_mlp: bool = True,
    train_attn: bool = True,
    train_unembed: bool = True,
) -> TrainingClient

Asynchronously create a TrainingClient for LoRA fine-tuning. Parameters and return value are the same as create_lora_training_client.

Parameters

Parameter	Type	Default	Description
`base_model`	`str`	-	Base model, for example `"Qwen/Qwen3.5-4B"`
`rank`	`int`	`32`	LoRA rank, from 4 to 64
`seed`	`int \| None`	`None`	Initialization seed
`train_mlp`	`bool`	`True`	Whether to train MLP layers
`train_attn`	`bool`	`True`	Whether to train attention layers
`train_unembed`	`bool`	`True`	Whether to train the `lm_head` layer

Returns

TrainingClient - client with the active training state.

Example

training_client = await client.create_lora_training_client_async(
    base_model="Qwen/Qwen3.5-4B",
    rank=16,
    train_unembed=False,
)

create_sampling_client_async

async def create_sampling_client_async(
    self,
    base_model: str,
    model_path: str | None = None,
) -> SamplingClient

Asynchronously create a SamplingClient for generation and inference. Parameters and return value are the same as create_sampling_client.

Parameters

Parameter	Type	Default	Description
`base_model`	`str`	`""`	Base model, for example `"Qwen/Qwen3.5-4B"`
`model_path`	`str \| None`	`None`	LoRA model checkpoint path URL to load during initialization

Returns

SamplingClient - sampling client.

Examples

sampling_client = await client.create_sampling_client_async(base_model="Qwen/Qwen3.5-4B")

sampling_client = await client.create_sampling_client_async(
    base_model="Qwen/Qwen3.5-4B",
    model_path="/path/to/checkpoint",
)

create_training_client_from_state_async

async def create_training_client_from_state_async(self, path: str) -> TrainingClient

Asynchronously create a TrainingClient from a saved adapter or checkpoint without restoring optimizer state. Parameters and return value are the same as create_training_client_from_state.

Parameters

Parameter	Type	Description
`path`	`str`	Adapter or checkpoint path

Returns

TrainingClient - client with restored model state.

Example

training_client = await client.create_training_client_from_state_async(
    path="/path/to/checkpoint"
)

create_training_client_from_state_with_optimizer_async

async def create_training_client_from_state_with_optimizer_async(self, path: str) -> TrainingClient

Asynchronously create a TrainingClient from a train checkpoint and restore optimizer state for resume training. Parameters and return value are the same as create_training_client_from_state_with_optimizer.

Parameters

Parameter	Type	Description
`path`	`str`	Train checkpoint path

Returns

TrainingClient - client with restored model and optimizer state.

Example

training_client = await client.create_training_client_from_state_with_optimizer_async(
    path="/path/to/checkpoint"
)

On this page