SPEC= (no draft): decode +46%, prefill +52%, ttft -35% vs MTP-2 baseline. Spec decode is NET-NEGATIVE on this endpoint for real chat+reasoning traffic at both depths; the bench-window win did not transfer.

Result: {"status":"keep","decode_tokens_per_sec":9.01,"prefill_tokens_per_sec":21088.7,"ttft_ms":851.8}
2026-10-01 21:00:19 -07:00
parent 5df6ef3166
commit 6513672f31
1 changed files with 1 additions and 1 deletions
@@ -85,7 +85,7 @@ FRANK_LARGE_UTIL=0.45
 # The gufo-era law "spec decode is absent" was a frank-primary finding; this
 # endpoint's traffic is decode-dominated thinking, which is the regime where
 # spec wins (prefill costs ~55% in every arm).
-FRANK_LARGE_SPEC=mtp
+FRANK_LARGE_SPEC=
 FRANK_LARGE_SPEC_DIR=/home/dank/models/Qwen3.5-9B
 FRANK_LARGE_SPEC_TOKENS=2