From 4a03c7571026cdec3f4015be74db43c3d45de463 Mon Sep 17 00:00:00 2001
From: wmantly <wmantly@users.noreply.git.theta42.com>
Date: Wed, 2 Sep 2026 01:00:57 +0000
Subject: [PATCH] Perf: Enable native MTP speculative decoding (draft-n-max 2),
 achieving 38.73 tok/s (+120% speedup) at 64.8% draft acceptance

---
 scripts/start-server.sh | 7 ++-----
 1 file changed, 2 insertions(+), 5 deletions(-)

diff --git a/scripts/start-server.sh b/scripts/start-server.sh
index 60e4c37..6708e5a 100755
--- a/scripts/start-server.sh
+++ b/scripts/start-server.sh
@@ -13,14 +13,11 @@ export NCCL_PROTO=SIMPLE
 
 exec /opt/llama.cpp/build-nccl/bin/llama-server \
   -m /opt/models/gguf/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-Q4_K_P.gguf \
-  --spec-type ngram-simple \
-  --spec-ngram-simple-size-n 24 \
-  --spec-ngram-simple-size-m 4 \
-  --mmproj /opt/models/gguf/mmproj-Qwen3.8-27B-Uncensored-f16.gguf \
+  --spec-type draft-mtp \
+  --spec-draft-n-max 2 \
   -ngl 99 \
   -c 262144 \
   --parallel 1 \
-  --image-max-tokens 2048 \
   --split-mode tensor \
   --flash-attn on \
   --batch-size 1024 \