Request: DFlash / DFlash2 draft checkpoint for Qwen/Qwen3.8-Flash-Next #176

Hi z-lab team, thanks for DFlash and the new DFlash2 drafters.

Would you consider a drafter for Qwen/Qwen3.8-Flash-Next (qwen4_exp: 176B total, ~6B active, 512 experts x 48 layers, hidden 2560, GDN + QSA hybrid attention, same 248,320-token vocabulary as Qwen3.8-27B)? I couldn't find one on the Hub, and it's the model a lot of people run locally right now because it fits consumer hardware with expert offloading.

My setup: 2x RTX 3090 with Strata, an engine that keeps most experts in system RAM and streams them. Today it drafts with the model's own MTP head (about 2 tokens per verify window, ~127 tok/s on IQ3_S). A Strata contributor is planning DFlash support (Niko1221/Strata#1161), but there is no drafter for this model to plug in.

One data point that may matter for the block size: in an offloaded MoE engine, verification isn't nearly free the way it is on an H200. I measured ~6 ms fixed + ~4 ms per verified token per window (normal ~2.5-token windows: 15.7 ms; forced 8-token windows: 37.5 ms). So a drafter trained for blocks of 8 or less, with high acceptance in the first few positions, would fit this kind of hardware better than block 16.

Happy to test a preview and report acceptance and throughput once it can run on my side.