Databricks reports 6x faster Qwen fine-tuning through kernel and batch changes

Databricks says it reduced fine-tuning time for Qwen 3.8 27B on a fixed dataset from 7 hours 44 minutes to 1 hour 17 minutes on AI Runtime. Its baseline already showed 98% GPU utilization, a measure that records whether the processor is active rather than how efficiently it works. The company combined small GPU operations into larger kernels and packed training examples to reduce unused padding. Installing fused kernels first cut the run to 2 hours 5 minutes; packing examples delivered the remaining improvement. Databricks warns that packing makes the input pipeline more complex and can introduce errors that affect model quality.