Post

Here is the "fun" kronk has to deal with for you. ndraft=3 for MTP works great for Qwen/Gemma but not for Ornith. Ornith works better with ndraft=2. llama.cpp hardcodes 3. Kronk uses an adaptive algorithm that can downgrade from 3 to 2 based on sampling, so Ornith runs faster.

Just downloaded the latest version of the @ornith_ 1.5 35B-A3B model and guess what? The model is running better at ndraft=3 now. Hahaha Always look for new uploads of your favorite models.