fix(model): effective-rank and dormant-ratio logging call sites raise TypeError

[Anai-Guo]

Two call sites in the "logging for analysis" path raise TypeError the moment the metric is enabled. Both are single-token mismatches against helpers that the sibling file already calls correctly.

1. world_model.py calls the wrong *_effective_rank

lzero/model/utils.py has two similarly named helpers:

def compute_effective_rank(singular_values: np.ndarray) -> float:            # :38
def calculate_effective_rank(model, inputs, representation_layer_name) -> float:   # :83

world_model.py imports the first one and calls it with the second one's arguments:

# lzero/model/unizero_world_models/world_model.py:2525
e_rank_last_linear = compute_effective_rank(
    self.tokenizer.encoder, inputs, representation_layer_name="last_linear"
)

world_model_multitask.py:1878 — the sibling implementation of the same block — already does it right:

e_rank_last_linear = calculate_effective_rank(
    self.tokenizer.encoder[encoder_index], inputs, representation_layer_name="last_linear")

So this changes the import and the two call sites to calculate_effective_rank. compute_effective_rank is not used anywhere else in world_model.py, and is still used internally by calculate_effective_rank (utils.py:133).

2. muzero.py passes percentage= to calculate_dormant_ratio

def calculate_dormant_ratio(model, inputs, dormant_threshold=1e-2, target_modules=...)   # utils.py:160

muzero.py:434 and :505 pass percentage=self._cfg.dormant_threshold. The config key is already called dormant_threshold; only the keyword is stale. world_model.py:2507 and world_model_multitask.py:1869 both use dormant_threshold=.

This is gated on if self._cfg.calculate_dormant_ratio: (default False), which is why it has stayed unnoticed — but turning that analysis flag on crashes the learn step immediately.

Verification

lzero/model/utils.py was loaded verbatim via importlib (not through the lzero package, so no ditk needed) and both call sites were replayed against a small nn.Module on CPU:

compute_effective_rank   (singular_values: numpy.ndarray) -> float
calculate_effective_rank (model, inputs, representation_layer_name: str) -> float
calculate_dormant_ratio  (model, inputs, dormant_threshold: float = 0.01, target_modules=...) -> Dict[str, float]

=== world_model.py:2525 — effective rank of a named layer ===
  compute_effective_rank  (on main): TypeError: compute_effective_rank() got an unexpected keyword argument 'representation_layer_name'
  calculate_effective_rank (patched): -> 7.2696

=== muzero.py:433 — dormant ratio ===
  percentage=          (on main): TypeError: calculate_dormant_ratio() got an unexpected keyword argument 'percentage'
  dormant_threshold=  (patched): -> global=52.3438

Both patched calls return a real number rather than merely binding, so the helpers do run end to end (forward hook fires, SVD computed).

No functional change anywhere else; compute_average_weight_magnitude and the surrounding logging are untouched.

🤖 Generated with Claude Code