Using `--spec-draft-type-k q4_0 --spec-draft-type-v q4_0` might actually decrease your context size! With quantized spec draft, my context size is 83200. Without it (i.e. using the default fp16 spec draft), context size increased to 91648. I reported this in a llama.cpp discussion and am17an (the GOAT behind MTP in llama.cpp) confirmed my findings as expected: https://github.com/ggml-org/llama.cpp/discussions/24102 Edit: I am using a 3090 for inference. This might or might not apply to...