refactor(char-causal-lm): drop unreachable GenerationConfig construction

[AmitMY]

Small one — the cheap wins from the audit are mostly spent, so this is a narrow dead-branch removal.

What

-        if generation_config is None:
-            generation_config = GenerationConfig()
         max_new_tokens = getattr(generation_config, 'max_new_tokens', None) or 50
         min_new_tokens = getattr(generation_config, 'min_new_tokens', None) or 0

_get_generation_params built a throwaway GenerationConfig when none was passed, then immediately read it back through getattr(..., None). But:

GenerationConfig().max_new_tokens          # None
getattr(None, 'max_new_tokens', None)      # None

Both paths land on the same or 50 / or 0 fallbacks, so the constructed object could never influence the result. The None check was doing nothing.

Verification

The existing tests already pin this from both sides, so no new test was needed and none changed:

• test_none_config — _get_generation_params(None) == (50, 0)
• test_empty_config — _get_generation_params(GenerationConfig()) == (50, 0)

241 passed, ruff clean. The GenerationConfig import stays — still used by the type annotations.

Not a perf claim

It does remove an allocation from the default generate() path (2.74 µs → 0.06 µs for that call), but that is noise against the hundreds of model forwards a generation runs. Mentioning it only so the number is on record; the reason to merge is the dead branch, not the microseconds.

🤖 Generated with Claude Code

---

Note

Low Risk
Narrow refactor with no logic change; generation limits still default to 50/0 when unset.

Overview
Removes a dead branch in CharacterCausalLMWrapper._get_generation_params that instantiated GenerationConfig() when the argument was None. The helper already reads limits via getattr(..., None) or <default>, so a fresh default config (also None for those fields) never changed the returned (50, 0) values.

The docstring now states that a missing config is equivalent to a default one for those fallbacks. Behavior is unchanged; existing TestGetGenerationParams cases (None vs empty GenerationConfig()) still cover both paths.

^{Reviewed by Cursor Bugbot for commit 97d3d19. Bugbot is set up for automated code reviews on this repo. Configure here.}