Fine-tune gotcha: HF export silently drops Qwen3's MTP head, slowing decode

AIFlow_ML · x · 2026-11-22

AIFlowML shares a subtle fine-tuning pitfall: after @catonooka modified Qwen3-27B, the Hugging Face export silently dropped the entire MTP head — the built-in drafter. Answers stay fine, only decode gets slower, so nobody notices.

The takeaway: diff the tensor names before and after every model export to catch missing components.

Original post →

More from coding & agent

coding & agent channel →