refactor(char-causal-lm): drop redundant gradient checkpointing overrides

[AmitMY]

Audit item #7, part two. Stacked on #10 — based on that branch so its four new tests run against this deletion. Merge #10 first; GitHub will retarget this to main automatically.

-    def gradient_checkpointing_enable(self, gradient_checkpointing_kwargs=None):
-        """Enable gradient checkpointing on the underlying model."""
-        self.model.gradient_checkpointing_enable(gradient_checkpointing_kwargs=gradient_checkpointing_kwargs)
-
-    def gradient_checkpointing_disable(self):
-        """Disable gradient checkpointing on the underlying model."""
-        self.model.gradient_checkpointing_disable()

Both only forwarded to self.model — which is exactly what PreTrainedModel already does. Its own implementation walks submodules and reaches the wrapped model without help.

Verified, not assumed

This is why #10 came first. With the overrides deleted, all four of those tests still pass on both CI transformers versions:

	transformers 4.57.6	transformers 5.14.1
enable reaches wrapped model	✅	✅
disable reaches wrapped model	✅	✅
gradient_checkpointing_kwargs form	✅	✅
forward + backward produces gradients	✅	✅

245 passed on both. ruff clean.

If you would rather keep the explicit delegation as documentation of intent, that is a fair call — closing this costs nothing, and #10's tests are worth having either way.

🤖 Generated with Claude Code

---

Note

Low Risk
Small deletion of pass-through methods; behavior is covered by existing gradient-checkpointing tests on the stacked PR.

Overview
Removes gradient_checkpointing_enable and gradient_checkpointing_disable from CharacterCausalLMWrapper because they only delegated to the inner self.model.

PreTrainedModel already enables and disables gradient checkpointing across submodules, and the wrapper still sets supports_gradient_checkpointing = True, so training behavior is unchanged. Companion tests on the stacked branch cover both transformers versions.

^{Reviewed by Cursor Bugbot for commit 389f7b1. Bugbot is set up for automated code reviews on this repo. Configure here.}

[AmitMY]

Closing (auto-closed when #10's base branch was deleted, and that is the right outcome): keeping the explicit gradient_checkpointing_enable/disable delegation.

The measurement stands — PreTrainedModel's own implementation does reach the wrapped model on both transformers 4.57.6 and 5.14.1 — but relying on that means depending on transformers' submodule walk continuing to work through this wrapper in future versions. The explicit forwarding states the intent for 7 lines. Not worth the coupling.

The four tests from #10 are merged and remain valuable either way: they pin delegation behaviour that was previously untested, so a future transformers change that breaks it fails loudly instead of silently disabling checkpointing mid-training.