Thanks, good question. Short answer: at one user it doesn't help. We measured it on an M5 Max with one server slot, the same n-gram drafting settings and prompts, toggling only GGML_METAL_TQD_EXT. Qwen3.6-35B-A3B came out at −0.9 % mean over 5 paired runs (range −7 % to +2.5 %), and Qwen3.8-Flash-Next at −1.1 % and −0.6 % over 2 pairs. That is level within run-to-run noise, and outputs were identical.
The reason is that with n-gram drafting the verify step is usually a ~65-token pass, which runs on the mat-mat path, not this kernel. Passes of 2–8 tokens were only about 0.4 % of mat-vec calls in a one-user session, so the kernel barely runs there. Our release note overstated its role in speculative verify at batch one; we've corrected it and this post. Where it pays is several users decoding at once (2–8 users, +8–24 % total decode).
For interactive use it's safe to leave on (no measurable cost, identical outputs), but don't expect lower per-token latency from it. GGML_METAL_TQD_EXT=0 turns it off if you want to A/B it yourself.