loader

YAML-driven configuration loading for fastgen distillation pipelines.

YAML is the first-class entry point for DMD configurations — the fastgen library does not expect callers to hand-build Python dicts. Typical usage:

from modelopt.torch.fastgen import DMDConfig, load_dmd_config

# (a) Load a built-in recipe by relative path
cfg = load_dmd_config("general/distillation/dmd2_qwen_image")

# (b) Load a user-provided file
cfg = load_dmd_config("/path/to/my_dmd.yaml")

# (c) Equivalent classmethod
cfg = DMDConfig.from_yaml("/path/to/my_dmd.yaml")

The loader resolves paths in two places, in order:

  1. modelopt_recipes/ (the built-in recipes package shipped with ModelOpt) — resolved via importlib.resources.files(). Suffixes .yml / .yaml may be omitted.

  2. The filesystem (absolute or working-directory-relative).

Suffixes .yml and .yaml are both accepted.

Functions

load_config

Load a YAML file and return the parsed mapping.

load_dmd_config

Load a YAML file and construct a DMDConfig.
load_config(config_file)

Load a YAML file and return the parsed mapping.

Mirrors modelopt.recipe._config_loader.load_config() in spirit but without the ExMy-num-bits post-processing that is specific to quantization recipes.

Parameters:

config_file (str | Path) – YAML path. Suffix is optional; resolution searches the built-in modelopt_recipes/ package first, then the filesystem.

Returns:

The parsed dictionary. An empty file yields {}.

Return type:

dict[str, Any]

load_dmd_config(config_file)

Load a YAML file and construct a DMDConfig.

The YAML is validated against DMDConfig’s Pydantic schema — unknown keys raise ValidationError.

Example YAML:

pred_type: flow
guidance_scale: 5.0
student_sample_steps: 2
gan_loss_weight_gen: 0.03
sample_t_cfg:
  time_dist_type: shifted
  t_list: [0.999, 0.833, 0.0]
ema:
  decay: 0.9999
Parameters:

config_file (str | Path)

Return type:

DMDConfig