From 42e2aad3a606c469ac5796728007fd2900169259 Mon Sep 17 00:00:00 2001
From: Mikei386 <44135113+Mikei386@users.noreply.github.com>
Date: Fri, 21 Aug 2026 13:18:45 +0200
Subject: [PATCH] Use embedded MTP tensor for IQ4-MIX profiles

---
 .env.example               | 1 -
 compose.yaml               | 4 ----
 config/install.env.example | 5 ++---
 install.sh                 | 5 +----
 4 files changed, 3 insertions(+), 12 deletions(-)

diff --git a/.env.example b/.env.example
index 9c87776..8212eed 100644
--- a/.env.example
+++ b/.env.example
@@ -12,7 +12,6 @@ MEDIUM_MODEL_FILE=qwen/Qwen3.8-27B-UD-IQ4_XS.gguf
 LONG_MODEL_FILE=qwen/Qwen3.8-27B-UD-IQ4_XS.gguf
 EXPERIMENTAL_MODEL_FILE=qwen/Qwen3.8-27B-UD-IQ4_XS.gguf
 VISION_PROJECTOR_FILE=qwen/mmproj-BF16.gguf
-MTP_MODEL_FILE=qwen/mtp-Qwen3.8-27B-Q4_0.gguf
 
 FAST_CONTEXT=76800
 MEDIUM_CONTEXT=94208
diff --git a/compose.yaml b/compose.yaml
index 6863b78..15f3ec3 100644
--- a/compose.yaml
+++ b/compose.yaml
@@ -87,8 +87,6 @@ services:
       - none
       - --spec-type
       - draft-mtp
-      - --model-draft
-      - "/models/${MTP_MODEL_FILE:?MTP_MODEL_FILE is required}"
       - --spec-draft-n-max
       - "2"
       - --spec-draft-type-k
@@ -217,8 +215,6 @@ services:
       - none
       - --spec-type
       - draft-mtp
-      - --model-draft
-      - "/models/${MTP_MODEL_FILE:?MTP_MODEL_FILE is required}"
       - --spec-draft-n-max
       - "2"
       - --spec-draft-type-k
diff --git a/config/install.env.example b/config/install.env.example
index 8403d60..98f77cd 100644
--- a/config/install.env.example
+++ b/config/install.env.example
@@ -43,9 +43,8 @@ EXPERIMENTAL_MODEL_SHA256=40fac4050e940397dbf13087afd50f4734a11805bf9d65ef8ddd74
 VISION_PROJECTOR_FILE=qwen/mmproj-BF16.gguf
 VISION_PROJECTOR_URL=https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/resolve/main/mmproj-BF16.gguf
 VISION_PROJECTOR_SHA256=83ee4f4f205fa514161778c41df1ea14144faa0f713510893b63c2395f5c2d53
-MTP_MODEL_FILE=qwen/mtp-Qwen3.8-27B-Q4_0.gguf
-MTP_MODEL_URL=https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/resolve/main/MTP/mtp-Qwen3.8-27B-Q4_0.gguf
-MTP_MODEL_SHA256=50d9ce5a6da381bbcfb31061cf73df94a90e6faf8efeddee379a9cb8f1501c6e
+# FAST/LONG use the MTP tensor embedded in the IQ4-MIX GGUF. A separate
+# draft-model artifact is neither downloaded nor passed to llama-server.
 
 FAST_CONTEXT=76800
 MEDIUM_CONTEXT=94208
diff --git a/install.sh b/install.sh
index 3ea4016..c1432cd 100755
--- a/install.sh
+++ b/install.sh
@@ -42,8 +42,7 @@ required=(AI_HOSTNAME ADMIN_USER AI_BIND_ADDRESS MODEL_DIR FAST_MODEL_FILE
   FAST_MODEL_URL FAST_MODEL_SHA256 MEDIUM_MODEL_FILE MEDIUM_MODEL_URL
   MEDIUM_MODEL_SHA256 LONG_MODEL_FILE LONG_MODEL_URL LONG_MODEL_SHA256
   EXPERIMENTAL_MODEL_FILE EXPERIMENTAL_MODEL_URL EXPERIMENTAL_MODEL_SHA256
-  VISION_PROJECTOR_FILE VISION_PROJECTOR_URL VISION_PROJECTOR_SHA256
-  MTP_MODEL_FILE MTP_MODEL_URL MTP_MODEL_SHA256)
+  VISION_PROJECTOR_FILE VISION_PROJECTOR_URL VISION_PROJECTOR_SHA256)
 for name in "${required[@]}"; do
   [[ -n "${!name:-}" ]] || die "Pflichtwert $name fehlt."
 done
@@ -202,7 +201,6 @@ MEDIUM_MODEL_FILE=$MEDIUM_MODEL_FILE
 LONG_MODEL_FILE=$LONG_MODEL_FILE
 EXPERIMENTAL_MODEL_FILE=$EXPERIMENTAL_MODEL_FILE
 VISION_PROJECTOR_FILE=$VISION_PROJECTOR_FILE
-MTP_MODEL_FILE=$MTP_MODEL_FILE
 FAST_CONTEXT=${FAST_CONTEXT:-76800}
 MEDIUM_CONTEXT=${MEDIUM_CONTEXT:-94208}
 LONG_CONTEXT=${LONG_CONTEXT:-131072}
@@ -245,7 +243,6 @@ $MEDIUM_MODEL_FILE|$MEDIUM_MODEL_URL|$MEDIUM_MODEL_SHA256
 $LONG_MODEL_FILE|$LONG_MODEL_URL|$LONG_MODEL_SHA256
 $EXPERIMENTAL_MODEL_FILE|$EXPERIMENTAL_MODEL_URL|$EXPERIMENTAL_MODEL_SHA256
 $VISION_PROJECTOR_FILE|$VISION_PROJECTOR_URL|$VISION_PROJECTOR_SHA256
-$MTP_MODEL_FILE|$MTP_MODEL_URL|$MTP_MODEL_SHA256
 EOF
 }