DFlash boosts Qwen3.5 on 8GB RTX

This is a strong proof that speculative decoding can matter even on VRAM-starved consumer GPUs, not just on large server cards. The win is practical, but it is also clearly configuration-sensitive and still tied to a bleeding-edge llama.cpp PR.

• –The result depends on careful tuning: `-ncmoe 34` and `--draft-max 6` were the sweet spots, while longer drafts reduced acceptance and hurt throughput.
• –The setup shows a useful pattern for oversized MoE models: keep the main model mostly off GPU, then use a small draft model to recover some decode speed.
• –The reported acceptance rate was extremely high at about 99.3%, which is why the speedup held up despite the modest 8GB card.
• –This reads more like an early benchmark win than a polished feature release, but it is a meaningful data point for local inference users trying to squeeze large Qwen models onto older hardware.