neurovlm.core.runtime.load_pipeline

Contents

neurovlm.core.runtime.load_pipeline#

neurovlm.core.runtime.load_pipeline(*, family='mlp', task='autoencoder', domain=None, variant=None, checkpoint=None, from_run=None, device='cpu', text_encoder=None)[source]#

Resolve and load an inference pipeline.

CNN domain tasks default to mixed_baseline. Passing variant='finetuned' is the only way to select specialized CNN weights. from_run accepts a run directory or its checkpoint; checkpoint is the explicit-file spelling. They are mutually exclusive.

Parameters:

  • family (str)

  • task (str)

  • domain (str | None)

  • variant (str | None)

  • checkpoint (str | Path | None)

  • from_run (str | Path | None)

  • device (str | device)

  • text_encoder (Callable[[Any], Tensor] | None)

Return type:

NeuroVLMRuntime

Contents