fix: forward warmup_ratio correctly to TrainingArguments

[Neonkraft]

Summary

Drop warmup_ratio from TrainingConfig and make warmup_steps a float. The original bug: warmup_ratio (e.g. 0.03) was being forwarded to warmup_steps (an int field in TrainingArguments), silently truncating to 0 and disabling warmup entirely. HuggingFace TrainingArguments already interprets a float value for warmup_steps as a ratio, so a single float field covers both use cases with no special handling needed.

Type of change

• Bug fix
• New feature
• Refactor
• Performance
• Documentation
• Maintenance

[KonstiNik]

Thanks for picking this up!

Tho, I'd argue we should follow HF's API more directly here. Maintaining warmup_ratio and warmup_steps as two separate fields for the same underlying concept doesn't really hold up — HF themselves consolidated this in 5.2 into a single warmup_steps: float where < 1 is interpreted as a ratio and >= 1 as absolute steps (training_args.py:788-793, training_args.py:2047-2054). One knob covers both use cases cleanly.

Concretely I suggest:

1. Drop warmup_ratio from TrainingConfig and rename to warmup_steps. The piece I think matters most: users need to be told what to put in this field, since accepting both ratios and absolute step counts is non-obvious from the name alone. Use the field(metadata=...) form (same pattern HF uses in TrainingArguments), and we can surface it in --help or auto-generated docs later:

warmup_steps: float = field(
    default=0.03,
    metadata={
        "help": (
            "Linear warmup duration. Values in [0, 1) are interpreted as a "
            "ratio of total training steps; values >= 1 are interpreted as an "
            "absolute number of steps; 0 disables warmup."
        )
    },
)

2. Forward only warmup_steps to TrainingArguments.
   Add a small shim in PostTrainingConfig.load (config.py:270, between OmegaConf.load and the schema merge) so old YAMLs don't break:

if training is not None and "warmup_ratio" in training:
    logger.warning(
        "training.warmup_ratio is deprecated; use training.warmup_steps "
        "(values < 1 are interpreted as a ratio). Auto-migrating."
    )
    training.warmup_steps = training.pop("warmup_ratio")

(Needs import logging + a module-level logger — neither exists in config.py today.)

[Neonkraft]

Done.

[KonstiNik]

Refactor looks great. One small question: the default went from 0.03 to 0.0, so configs that didn't set warmup explicitly will now get no warmup instead of 3%. Was the default change intentional, or worth flipping back to 0.03 to keep the old behavior?

[Neonkraft]

Was the default change intentional, or worth flipping back to 0.03

It was intentional. TRL's default is 0.0 so I thought it best to mimic it.