ling trial: layer ngram-mod under draft-mtp; add ngram args to serve script

scripts/frank-default-large.ling.env

@@ -21,8 +21,12 @@ PORT=7770
 CTX_SIZE=262144
 PARALLEL=2
 # MTP is bundled in the Ling GGUF itself (block 42) - no separate draft file.
-SPEC_TYPE=draft-mtp
+# Layering ngram-mod under MTP (draft-mtp,ngram-mod) re-drafts via ngram.
+SPEC_TYPE=draft-mtp,ngram-mod
 SPEC_DRAFT_N_MAX=3
+SPEC_NGRAM_MOD_N_MIN=48
+SPEC_NGRAM_MOD_N_MAX=64
+SPEC_NGRAM_MOD_N_MATCH=24
 # Vulkan exposes the larger heap (UMA + GTT) vs ROCm. Use Vulkan0.
 DEVICE=Vulkan0
 N_GPU_LAYERS=all

scripts/serve-ling-flash.sh

@@ -34,6 +34,9 @@ fi
 : "${PARALLEL:=2}"
 : "${SPEC_TYPE:=draft-mtp}"         # MTP block is bundled in the model GGUF
 : "${SPEC_DRAFT_N_MAX:=3}"
+: "${SPEC_NGRAM_MOD_N_MIN:=48}"
+: "${SPEC_NGRAM_MOD_N_MAX:=64}"
+: "${SPEC_NGRAM_MOD_N_MATCH:=24}"
 : "${DEVICE:=Vulkan0}"
 : "${N_GPU_LAYERS:=all}"
 : "${BATCH_SIZE:=512}"
@@ -62,6 +65,9 @@ exec "$LLAMA_SERVER" \
     --model "$MODEL" \
     --spec-type "$SPEC_TYPE" \
     --spec-draft-n-max "$SPEC_DRAFT_N_MAX" \
+    --spec-ngram-mod-n-min "$SPEC_NGRAM_MOD_N_MIN" \
+    --spec-ngram-mod-n-max "$SPEC_NGRAM_MOD_N_MAX" \
+    --spec-ngram-mod-n-match "$SPEC_NGRAM_MOD_N_MATCH" \
     --device "$DEVICE" \
     --gpu-layers "$N_GPU_LAYERS" \
     --flash-attn auto \