A GGUF MTP Extract and Merge Tool

[FNsi]

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High

GGUF MTP (Multi-Token Prediction) Extract and Merge Tool
Extract NextN/MTP layers from a model that has them and insert to model without them, use for add mtp layer to unsloth finetuned model or existed gguf without mtp layer. Should work with Qwen3.6-27B, didn't test anything else..
Usage:
Extract MTP from source model and merge into target model
python gguf_mtp_extract_merge.py extract-merge
~/source-with-mtp.gguf
~/target-without-mtp.gguf
~/output-with-mtp.gguf
Just extract MTP tensors to a standalone file
python gguf_mtp_extract_merge.py extract
~/source-with-mtp.gguf
~/mtp-layers.gguf
Just merge extracted MTP tensors into target
python gguf_mtp_extract_merge.py merge
~/mtp-layers.gguf
~/target-without-mtp.gguf
~/output-with-mtp.gguf

AI usage: mostly generate by Kimi k2.6

[ikawrakow]

I haven't tested this, hopefully it works and is useful to people.

[FNsi]

I haven't tested this, hopefully it works and is useful to people.

Thanks for that, I also only tested the 27b version since I do only have one 27B mtp and one GRM (finetuned model without mtp. (Edit: after merge it) Accept rate with draft-max 4 works fine just like origin);

But maybe better keep it as draft? Simply think users here can use it even it's a draft, also because...I guess this tool is not needed anymore in very short future, like weeks......