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.
More from coding & agent
- A single Codex task has been running autonomously for over 10 days — chrisalbon · 2026-11-22
- Full indie music video made locally on DGX Spark with Wan2GP + MiniMax H3 — Inevitable_Emu2722 · 2026-11-22
- The 'Agent Platform of 2027' blueprint: actors, RBAC filesystems, inbox UX — yawnxyz · 2026-11-22
- Open-source tool unlocks third-party models in Claude Desktop via endpoint mapping — pneuny · 2026-11-22
- Open-source Claude skill turns one image into a fully disassemblable 3D mechanical model — tinyfool · 2026-11-22
- Salesforce's release-review agent cut review turnaround from 24 hours to 12 seconds — shoumikgoswami · 2026-11-22