Missing length normalization for grpo

[stepartm]

The problem is in the function run_grpo_microbatch_train_step.
Its docstring implies that the function should expect constant_normalize_factor as an input argument(which is very convenient for ablation studies on length normalization):

assignment5-alignment/tests/adapters.py

Line 237 in a016aff

constant_normalize_factor: int | None, provided if we want to sum over

And yet there is no such argument in the function definition😢

[mmjwxbc]

In supervised fine‑tuning, should the loss be normalized by the number of valid tokens (i.e., divided by the effective token count)?