chore: frank-default-small — ngram-mod, drop DFlash draft

- spec-type → ngram-mod (no separate draft model)
- drop --model-draft, --spec-draft-n-max; add --spec-ngram-mod-n-min/max/match
- DFlash batch default fallback SPEC_TYPE → ngram-mod
- env: SPEC_TYPE=ngram-mod, NGRAM_MOD_N_MIN=48, N_MAX=64, N_MATCH=24
- DRAFT_FILE env var emptied (not used under ngram-mod)

scripts/serve-llama-frank-small.sh

@@ -1,6 +1,6 @@
 #!/usr/bin/env bash
 # Serve HauhauCS Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive:Q6_K_P
-# with a DFlash draft model on dankserver (RTX 3090, MoE experts on CPU).
+# with ngram-mod speculative decoding on dankserver (RTX 3090, MoE experts on CPU).
 # Runtime: upstream llama.cpp llama-server (b10108+).
 #
 # Tunables come from frank-default-small.env in this directory unless the
@@ -19,15 +19,17 @@ fi
 # --- defaults (used when the env file does not set them) -----------------
 : "${MODEL_DIR:=/home/dank/models}"
 : "${MODEL_FILE:=Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-Q6_K_P.gguf}"
-: "${DRAFT_FILE:=Qwen3.6-35B-A3B-DFlash-Q8_0.gguf}"
+: "${DRAFT_FILE:=}"   # not used — ngram-mod reuses the loaded model for drafts
 : "${MMPROJ_FILE:=mmproj-Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-f16.gguf}"
 : "${ALIAS:=frank-default-small}"
 : "${HOST:=0.0.0.0}"
 : "${PORT:=7770}"
 : "${CTX_SIZE:=262144}"
 : "${PARALLEL:=3}"
-: "${SPEC_TYPE:=draft-dflash}"
-: "${SPEC_N_MAX:=7}"
+: "${SPEC_TYPE:=ngram-mod}"
+: "${NGRAM_MOD_N_MIN:=48}"
+: "${NGRAM_MOD_N_MAX:=64}"
+: "${NGRAM_MOD_N_MATCH:=24}"
 : "${CACHE_TYPE_K:=q8_0}"
 : "${CACHE_TYPE_V:=f16}"
 : "${N_CPU_MOE:=40}"     # all 40 layers: MoE expert weights stay on CPU
@@ -41,10 +43,9 @@ fi
 
 LLAMA_SERVER="${LLAMA_SERVER:-/home/dank/llama.cpp/build/bin/llama-server}"
 MODEL="$MODEL_DIR/$MODEL_FILE"
-DRAFT="$MODEL_DIR/$DRAFT_FILE"
 MMPROJ="$MODEL_DIR/$MMPROJ_FILE"
 
-for f in "$MODEL" "$DRAFT" "$MMPROJ"; do
+for f in "$MODEL" "$MMPROJ"; do
     if [[ ! -f "$f" ]]; then
         echo "missing required file: $f" >&2
         exit 1
@@ -53,11 +54,12 @@ done
 
 exec "$LLAMA_SERVER" \
     --model "$MODEL" \
-    --model-draft "$DRAFT" \
     --mmproj "$MMPROJ" \
     --jinja \
     --spec-type "$SPEC_TYPE" \
-    --spec-draft-n-max "$SPEC_N_MAX" \
+    --spec-ngram-mod-n-min "$NGRAM_MOD_N_MIN" \
+    --spec-ngram-mod-n-max "$NGRAM_MOD_N_MAX" \
+    --spec-ngram-mod-n-match "$NGRAM_MOD_N_MATCH" \
     --cache-type-k "$CACHE_TYPE_K" \
     --cache-type-v "$CACHE_TYPE_V" \
     --ctx-size "$CTX_SIZE" \