Why your quantized LLM loses its MTP heads and how to keep them
The article discusses the challenges faced when quantizing models with multi-token prediction (MTP) heads. It highlights how these heads can be inadvertently dropped during the conversion process, leading to unexpected performance issues. The author provides a step-by-step solution to ensure MTP heads are preserved during quantization.
- ▪MTP heads are auxiliary components that predict future tokens in parallel, enhancing model performance.
- ▪Many quantization toolchains are not designed to recognize MTP heads, leading to their silent removal during conversion.
- ▪The author suggests a workflow that includes inventorying MTP heads before conversion and modifying the converter's allowlist to prevent loss.
DEV.to (Top) files mainly under programming. We currently carry 4,924 of its stories.
Story provenance
Source · retrieval · rights · ranking — open for full record
inspect →
Attribution is not the same as permission. This drawer separates discovery metadata, excerpts, WeSearch-generated summaries, reuse status, and whether the publisher receives the visit. Nothing here claims a legal grant the publisher has not made.
Record
| Original publisher | DEV.to (Top) |
| Canonical URL | https://dev.to/alanwest/why-your-quantized-llm-loses-its-mtp-heads-and-how-to-keep-them-m7h |
| Publication time | Wed, 27 May 2026 16:00:08 +0000 |
| Retrieval time | 2026-05-27T16:08:01.779Z |
| Last seen | 2026-05-27T16:08:01.779Z |
| Headline source | Publisher (no WeSearch rewrite) |
| Excerpt source | publisher body |
| Excerpt method | First ~120 words (~800 chars) of extracted publisher body, fair-use limited. |
| Summary | WeSearch · cerebras-chat (WeSearch summarizer) |
| Summary source text | contentText |
| Citation coverage | Summary is a WeSearch-generated derivative; primary citation is the original publisher URL. |
| Cluster | pr5PoymWEbVt |
| Cluster logic | Grouped by semantic title/content similarity across sources within a rolling window. Same-publisher template collisions are excluded from coverage comparison. |
| Ranking reason | Story pages are not engagement-ranked. Hub feeds use recency, with optional source-diversified chronological ordering (cap consecutive stories per source). No personalized ranking. |
| Publisher visit | Yes — open original |
| Substitutes article? | No — link-out required for full text |
Rights status (four layers)
WeSearch handling by dimension
| Indexing | May the item be indexed (stored, ranked, made findable)? | Allowed |
| Snippet | May a short excerpt of the publisher's text be shown? | Allowed |
| AI summary | May WeSearch generate its own short summary of the article? | Limited |
| Retrieval / RAG | May the content be exposed for third-party retrieval-augmented generation? | Not asserted |
| Model training | May the content be used to train AI models? | Not asserted |
| Commercial reuse | May the content be reused commercially? | Not permitted |
Basis: Derived from the published RSS/Atom feed. Contact: [email protected]. Reviewed: 2026-07-24.
Opening excerpt (first ~120 words) tap to expand
try { if(localStorage) { let currentUser = localStorage.getItem('current_user'); if (currentUser) { currentUser = JSON.parse(currentUser); if (currentUser.id === 3834047) { document.getElementById('article-show-container').classList.add('current-user-is-article-author'); } } } } catch (e) { console.error(e); } Alan West Posted on May 27 Why your quantized LLM loses its MTP heads and how to keep them #machinelearning #llm #python #quantization The frustrating problem Last month a teammate pinged me with a classic head-scratcher. He'd taken a base model with multi-token prediction (MTP) heads, ran it through a standard quantization pipeline to ship a smaller GGUF for edge inference, and the latency numbers came back worse than expected.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at DEV.to (Top).
Discussion
0 commentsMore from DEV.to (Top)
