depth

Classes

DepthScenario
SublayerRemoval

Functions

launch_iterative_depth_automodel	Create the full removal trajectory, resuming at candidate granularity.
run_depth_mip_scenarios	Run all scenario MIPs, then realize/evaluate the winners in one parent sweep.

class DepthScenario

Bases: StrictModel

data_identity: str

evaluator_revision: str

granularity: Granularity

hidden_width: int

metric: Literal['lm_loss']

model_config = {'extra': 'forbid', 'frozen': True}

Configuration for the model, should be a dictionary conforming to [ConfigDict][pydantic.config.ConfigDict].

parent_checkpoint_identity: str

removals: tuple[SublayerRemoval, ...]

property scenario_id: str

class SublayerRemoval

Bases: StrictModel

kind: Literal['block', 'attention', 'mamba', 'ffn', 'moe']

layer_idx: int

model_config = {'extra': 'forbid', 'frozen': True}

Configuration for the model, should be a dictionary conforming to [ConfigDict][pydantic.config.ConfigDict].

launch_iterative_depth_automodel(hydra_cfg)

Create the full removal trajectory, resuming at candidate granularity.

Parameters:

hydra_cfg (DictConfig)

Return type:

dict[str, Any]

run_depth_mip_scenarios(cfg)

Run all scenario MIPs, then realize/evaluate the winners in one parent sweep.

Parameters:

cfg (DictConfig)

Return type:

list[str]