Flag for minimum speculative decoding probability

[Nabokov86]

Thank you for adding MTP support! It works great!

Could you also add the --spec-draft-p-min (minimum speculative decoding probability) flag from llama.cpp? For me it speeds up generation quite a bit if set > 0.5.

[Nabokov86]

On llama.cpp with the same prompt and k1:

Flag Setting	Performance
--spec-draft-p-min 0.1	19.95 tokens/sec
--spec-draft-p-min 0.75	22.74 tokens/sec

[LostRuins]

Hmm should be possible. i'll keep it in mind!