I have had DeepSeek write large poems for testing token generation speed.
Tonight, on a real project it was searching for relevant material when it found a note it thought might be relevant. That note was simply the poem.
It thought:
The river remembers was a poem-- probably output of testing...
I never connected testing to that task or mentioned it.
We've now reached up to 143 tok/s prose on Broadwell with significant gains in prefill speed and reductions in wall-time.
We now provide a docker image with configurable for speed, quality, or balance:
github.com/CurtisAccelerate/…
This update adds full 40-layer integer prefill preparation fusion, faster target execution and integrated draft sampling to our existing DeepSeek EXL3 serving stack. We retained our specialized expert shapes, split-attention and collective tuning, and Engram integration.
Selected improvements from @GNmac's recipe are stacked into that work, including short-row FP8 target paths and fast draft-head/sampling components. We also tested his adaptive drafting policy; our local results favored keeping fixed-five verification as the default. We don't claim a performance superiority.
Thanks @GNmac for building and sharing your recipe!
Pro Tip: How to Get Better While Using AI
1. First explain to the AI how you currently do your task/job.
2. Ask it how a a novice, journeyman, or expert would do the same task. Alternatively, ask it how to do the task in the most efficient way.
Ask the AI, first explain how you are using the computer to do a task then ask it how to do it most efficiently, how an expert would do it, etc.
Examples 1
I have dozens of build directories with runnable programs I want to examine. But, each build has many subfolders that include partials that I don't want to view. I explained that neither trying to navigate in the Windows file explorer NOR using the search filter was efficient for me.
To my surprise, it built a custom file explorer on the fly. This did not really help but I realized it would work if we added the ability to filter/expand by folder depth.
With those changes, I can now see all artifacts that are the final builds for all folders and advance by depth for additional artifacts.
Example 2
I explained that working in terminal required too much typing and asked how an expert terminal user would achieve the same task: it installed a bunch of extensions.
But, how to remember them: so, I had it print a cheat sheet of all my terminal commands every time I open my terminal.




