Can the MTP version currently be loaded in LM Studio?

#2
by cclongdevil - opened

First of all, thank you very much for your work; I have had a very good experience using QWEN3.6 27BV2, 35B-A3B, and GEMMA4!
I see that the GGUF version of MTP is now available, but I don't seem to be able to load it smoothly on LM Studio (even with v2.14.0). Is there any way to load it?
I am really looking forward to the more than twofold speedup that MTP will bring!
By the way, my English is not very good, so this text is the result of translating Chinese into English using your GEMMA4-26B-A4B GGUF.

Owner

First of all, thank you very much for your work; I have had a very good experience using QWEN3.6 27BV2, 35B-A3B, and GEMMA4!

Happy to hear that.

I see that the GGUF version of MTP is now available, but I don't seem to be able to load it smoothly on LM Studio (even with v2.14.0). Is there any way to load it?
I am really looking forward to the more than twofold speedup that MTP will bring!

Right now MTP support has not yet been implemented in llama.cpp, but it's supposed to be coming:

https://www.reddit.com/r/LocalLLaMA/comments/1t3guzw/llamacpp_mtp_support_now_in_beta/

I tested and it seems to work well with this branch: https://github.com/ggml-org/llama.cpp/pull/22673

To use it, start llama-server with --spec-type mtp --spec-draft-n-max 3

With Q8_0 My token generation went up significantly (~16 tps to ~25tps, though this highly depends on what the LLM is generating)

Works for me with that branch, but for some reason it's significantly slower than Unsloth's 27b-MTP release. Getting 35-45 TPS with this one (Q4_K_M, RTX 3090), vs 55-60 on Unsloth's (IQ4_NL quant).

I used llama.cpp to execute MTP, using 27b-q5-k-m, and on my 5090 laptop, it improved the speed from 33 tokens/s to 55 tokens/s. Thank you so much!

I used llama.cpp to execute MTP, using 27b-q5-k-m, and on my 5090 laptop, it improved the speed from 33 tokens/s to 55 tokens/s. Thank you so much!

Great to hear and you're welcome!

Sign up or log in to comment