DFlash or DSpark for MiniCPM-o ?
https://developer.nvidia.com/blog/boost-inference-performance-up-to-15x-on-nvidia-blackwell-using-dflash-speculative-decoding/
=>
https://huggingface.co/collections/z-lab/dflash
=>
https://huggingface.co/z-lab/Qwen3-8B-DFlash-b16
Would it be a valid option and would you plan to, train a DFlash drafter for MiniCPM-o 4.5 to even have a higher performance of the model on for example the DGX Spark or mobile device with enough RAM, without losing quality?
As Qwen3 8B is part of your architecture, I thought maybe with your training data for MiniCPM-o you could make a DFlash drafter as well and add it to the pipeline for even faster responses 🚀🎉
I love the mindset behind your approach.
I think it can be a solid approach for adding sensory data support to existing power houses, like you do as well with MiniCPM-Robot. Very smart concept!
Imagine MiniCPM-o-Robot 🤗 robot talks back and moves(= output) based on sensory data, video, audio, text, smart brain. You just tell the intent and show it, and the robot sees and reasons, remembers and does.
And the DFlash or DSpark drafter makes the response faster and costing less energy to compute.
Thank you for sharing!