Dataset & DataLoader¶

class softs.SoftIterableDataset[source]¶

Infinite dataset yielding decoded tensor dicts.

Call set_model(model_id) to switch models at any time. DataLoader workers detect the change and discard pending work.

__init__(model_id, endpoint, batch_config, medium_cls, num_slots=8, retry_delay=0.01, request_timeout=60.0)[source]¶

Parameters:

model_id (str)
endpoint (str)
batch_config (BatchConfig)
num_slots (int)
retry_delay (float)
request_timeout (float | None)

property model_id: str¶

set_model(model_id)[source]¶

Parameters:: model_id (str)
Return type:: None

class softs.SoftDataLoader[source]¶

DataLoader with model switching support.

Usage:

loader = SoftDataLoader(model_id="teacher_v1", endpoint=ep.frontend,
                        batch_config=config, medium_cls=ShmMedium,
                        num_slots=8, batch_size=4)
for batch in loader:
    train(batch)

loader.set_model("teacher_v2")  # switch the requested product id

__init__(model_id, endpoint, batch_config, medium_cls, num_slots=8, retry_delay=0.01, request_timeout=60.0, **dataloader_kwargs)[source]¶

Parameters:

model_id (str)
endpoint (str)
batch_config (BatchConfig)
num_slots (int)
retry_delay (float)
request_timeout (float | None)

set_model(model_id)[source]¶

Parameters:: model_id (str)
Return type:: None

property model_id: str¶

class softs.Batch[source]¶

__init__(tensors, slot_ids=None, client=None)[source]¶

Parameters:

tensors (dict[str, Tensor])
slot_ids (list[int] | None)
client (Client | None)

keys()[source]¶

release()[source]¶

Return type:: None

softs.make_collate_fn(client, batch_config, auto_release=True)[source]¶

Parameters:

client (Client)
batch_config (BatchConfig)
auto_release (bool)

Return type:

Callable[[list[int]], Batch]