Skip to content

speculators.cli.train

Train command — thin typer wrapper around TrainConfig argument resolution.

Functions:

train_command 

train_command(ctx: Context) -> None

Train a speculator model.

All arguments are forwarded to the training config resolver (use --help to see every available flag). Pass --config with a YAML file, or use --key=value flags directly.

 Examples: speculators train --config config.yaml speculators train --verifier-name-or-path meta-llama/... \ --data-path ./output --speculator-type eagle3 torchrun --nproc_per_node=4 -m speculators.train \ --config config.yaml

Source code in speculators/cli/train.py 
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
def train_command(ctx: typer.Context) -> None:
    """Train a speculator model.

    All arguments are forwarded to the training config resolver
    (use ``--help`` to see every available flag). Pass ``--config``
    with a YAML file, or use ``--key=value`` flags directly.

    \b
    Examples:
        speculators train --config config.yaml
        speculators train --verifier-name-or-path meta-llama/... \\
            --data-path ./output --speculator-type eagle3
        torchrun --nproc_per_node=4 -m speculators.train \\
            --config config.yaml
    """
    from speculators.train.cli import main  # noqa: PLC0415
    from speculators.train.config import TrainConfig  # noqa: PLC0415

    main(TrainConfig.resolve(ctx.args))