Unknown pre-tokenizer type: 'deepseek-r1-qwen'

[ElCondor1969]

Hi Jarrod.
I tried loading some DeepSeek Distil Qwen based models with Lumina; below you can find the list of models I tried to load:

1. uncensoredai_UncensoredLM-DeepSeek-R1-Distill-Qwen-14B-Q6_K_L.gguf
2. DeepSeek-R1-Distill-Qwen-32B-abliterated-Q4_0.gguf
3. DeepSeek-R1-Distill-Qwen-32B-Q4_1.gguf

But all of them, during loading, give the following error:

"print_info: file size = 11.22 GiB (6.78 BPW)
llama_model_load: error loading model: error loading model vocabulary: unknown pre-tokenizer type: 'deepseek-r1-qwen'"

Lumina Log.log

Not that maybe Lumina should be compiled with the latest version of Llama.cpp?
Thanks!

[jarroddavis68]

it's because llama.cpp needs to be updated to support those versions. I will be updating llama.cpp soon. It's tricky because it can break code (3 times have had to start over so far, and had to make some major changes even in Lumina, the last time I updated llama.cpp).

If you want to use those models NOW, just use AIToolkit, its designed in a way that I can frequently update llama.cpp more often.

[ElCondor1969]

Hi Jarrod.
Okay, I'll take your advice and dig deeper into AIToolkit.
Thanks for your valuable work.