[qgallouedec]

Both trainers resolve the same per-model logit post-processing (logit_scale / output_multiplier / final_logit_softcapping) for their chunked loss, but reached it two different ways. DistillationTrainer uses config.get_text_config(); SFT branched on an is_vlm flag.

• _patch_chunked_ce_lm_head now reads model.config.get_text_config(), so the resolution block is word-for-word identical to the one in distillation_trainer.py.
• is_vlm stays — it still gates two transformers < 5.0.0 fallbacks (VLMs set base_model_prefix = "", and take config-level MoE aux-loss parameters) — but its docstring no longer claims a config-reading role it no longer has.
• TestPatchChunkedCELMHead + TestChunkedCrossEntropyLoss: 79 passed.

---

Note

Low Risk
Narrow refactor of config resolution for chunked NLL training; behavior should improve consistency for VLMs without touching auth or data paths.

Overview
SFT chunked-loss patching now resolves logit post-processing (logit_scale / output_multiplier, final_logit_softcapping, and MoE output_router_logits) via model.config.get_text_config() instead of choosing between top-level model.config and model.config.text_config when is_vlm is true. That matches DistillationTrainer’s chunked-loss path so both trainers read the same config surface for VLMs and text-only models.

The is_vlm argument is unchanged for behavior: it still only drives transformers < 5.0.0 fallbacks (VLM backbone via self.model and config-level MoE aux-loss fields). Its docstring no longer describes config selection, since that is handled uniformly by get_text_config(). Inline comments were updated to document the logit_scale / output_multiplier fallback semantics in line with distillation.

^{Reviewed by Cursor Bugbot for commit f4fe9e5. Bugbot is set up for automated code reviews on this repo. Configure here.}

[bot-ci-comment]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[cmpatino]

LGTM!