Restore Model Training closing flow

Author: profvjreddi

book/quarto/contents/vol1/training/training.qmd

@@ -6819,10 +6819,6 @@ The iron law classifies a slow job as compute-bound (improve arithmetic or algor
 
 Training is where the iron law becomes a daily instrument. Mixed precision changes effective throughput and memory traffic; checkpointing exchanges memory for recomputation; scaling adds communication. The engineering task is to identify the dominant term and apply the least costly intervention that changes it.
 
-```{=latex}
-\newpage
-```
-
 ::: {.callout-chapter-connection title="From training to data selection"}
 
 Training produces learned parameters by spending compute on every example; the chapter's optimized run still requires `{python} GPT2SummaryChapterRecap.o_total_mem_gb_str` of memory. Once the training loop is efficient, the next question is whether every example earns that cost. Repeated or low-information samples consume the same forward pass, backward pass, optimizer update, and communication as examples that add coverage or correct a weakness.