An EAGLE draft module trained on general chat data is d...

When a target model is updated, the draft module must be re-aligned to match the new output distribution. The immediate solution is to perform a short fine-tuning run (e.g., 200M tokens) of the EAGLE module using outputs from the updated target model. To prevent future regressions, the draft module’s training should be synchronized with the target model’s update pipeline. While the fix is being deployed, falling back to direct decoding for the affected traffic slice is recommended to maintain latency SLOs.