depth

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.

ModeloptConfig DepthScenario

Bases: StrictModel

Show default config as JSON

Default config (JSON):

{
   "parent_checkpoint_identity": null,
   "hidden_width": null,
   "removals": [],
   "data_identity": null,
   "evaluator_revision": null,
   "metric": "lm_loss",
   "granularity": "subblock"
}

field data_identity: str [Required]

field evaluator_revision: str [Required]

field granularity: Granularity

field hidden_width: int [Required]

Constraints:

gt = 0

field metric: Literal['lm_loss']

field parent_checkpoint_identity: str [Required]

field removals: tuple[SublayerRemoval, ...]

property scenario_id: str

ModeloptConfig SublayerRemoval

Bases: StrictModel

Show default config as JSON

Default config (JSON):

{
   "layer_idx": null,
   "kind": null
}

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

field layer_idx: int [Required]

Constraints:

ge = 0

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]