Based on advice from LocalLLama implemented a n+2 lookahead prediction for ssd streaming Qwen Flash Next - and it worked!

My project is optimising qwen flash next on Mac Mini m5 64gb through SSD streaming. I recently asked LocalLLama for guidance on how to optimise token output.

Today a popular method to do this is a lookahead guesser which fetches experts one layer ahead of the current layer. One popular suggestion was building a prediction model to start loading experts two layers ahead. This netted me +8% on chat and +4.2% on code.

The way the various optimisations stack on top of each other this may actually net alot more on a base vanilla setup.

https://github.com/skeggsguy/Flash-next-ssd

Pretty cool! Thanks reddit LocalLLama! I’ll go find the og post to link now.

Edit: Kudos to sdroege_ for og suggestion!