Update entrypoint.sh

entrypoint.sh

@@ -11,7 +11,7 @@ REASONING_MODE="${REASONING_MODE:-off}"
 CACHE_TYPE_K="${CACHE_TYPE_K:-q8_0}"
 CACHE_TYPE_V="${CACHE_TYPE_V:-q8_0}"
 FLASH_ATTN="${FLASH_ATTN:-on}"
-SPEC_TYPE="${SPEC_TYPE:-ngram-cache}"
+SPEC_TYPE="${SPEC_TYPE:-draft-mtp}"
 
 # NOTE: the b8840 CPU build does not support draft-mtp (needs a newer
 # llama.cpp build than this image ships). ngram-cache is a self-speculation
@@ -32,4 +32,5 @@ exec /app/llama-server \
   --cache-type-v "$CACHE_TYPE_V" \
   --reasoning "$REASONING_MODE" \
   -fa "$FLASH_ATTN" \
-  $SPEC_FLAGS
+  $SPEC_FLAGS \
+   --spec-draft-n-max 4