Add recommended sampling defaults and reduce MoE GPU offload on dankserver

- frank-default-large: add TEMP=1.0, TOP_K=20, TOP_P=0.95, MIN_P=0.0
  (per HauhauCS Qwen3.6-27B model card, thinking-general mode)
- frank-default-small: add same sampling defaults
  (per HauhauCS Qwen3.6-35B-A3B model card)
- frank-default-small: reduce N_CPU_MOE from 33 to 34 (6 expert
  layers on GPU instead of 7) to prevent OOM during high-load sessions
- Both serve scripts pass --temp, --top-k, --top-p, --min-p to
  llama-server; clients can still override per-request via API

scripts/frank-default-large.env

@@ -33,5 +33,12 @@ CACHE_TYPE_V=q8_0
 # User-specified ubatch 4096 / batch 12288 is not feasible here (see bench).
 BATCH_SIZE=2048
 UBATCH_SIZE=512
+# Sampling settings — per HauhauCS Qwen3.6-27B model card.
+# Thinking-general: temp=1.0, top_p=0.95, top_k=20, min_p=0.
+# Clients may override per-request; these are the server-side defaults.
+TEMP=1.0
+TOP_K=20
+TOP_P=0.95
+MIN_P=0.0
 THREADS=12
 THREADS_HTTP=1

scripts/frank-default-small.env

@@ -8,15 +8,15 @@ PORT=7770
 # (Increased from 3×200k to add a 4th parallel stream.)
 CTX_SIZE=800000
 PARALLEL=4
-# Spec: ngram-mod. n_min=48, n_max=64, n_match=24 (defaults). 8 expert
+# Spec: ngram-mod. n_min=48, n_max=64, n_match=24 (defaults). 6 expert
-# layers back on GPU; rest on CPU. Batch 8192 / ubatch 2048.
+# layers back on GPU; 34 on CPU. Batch 8192 / ubatch 2048.
 SPEC_TYPE=ngram-mod
 NGRAM_MOD_N_MIN=48
 NGRAM_MOD_N_MAX=64
 NGRAM_MOD_N_MATCH=24
-# Expert offload: 33 of 40 layers keep MoE weights on CPU; remaining
+# Expert offload: 34 of 40 layers keep MoE weights on CPU; remaining
-# 7 layers have their expert weights resident on GPU for faster routing.
+# 6 layers have their expert weights resident on GPU for faster routing.
-N_CPU_MOE=33
+N_CPU_MOE=34
 N_GPU_LAYERS=all
 # KV cache on GPU, K quantized to q8_0, V stays f16 (user-locked decision).
 CACHE_TYPE_K=q8_0
@@ -31,5 +31,12 @@ UBATCH_SIZE=2048
 # CPU, so eager RAM load avoids page faults during decode (uses ~31GB RAM).
 NO_MMAP=1
 # Compute threads lowered to 7 so other processes on the box keep a core.
+# Sampling settings — per HauhauCS Qwen3.6-35B-A3B model card.
+# Thinking-general: temp=1.0, top_p=0.95, top_k=20, min_p=0.
+# Clients may override per-request; these are the server-side defaults.
+TEMP=1.0
+TOP_K=20
+TOP_P=0.95
+MIN_P=0.0
 THREADS=7
 THREADS_HTTP=8

scripts/serve-llama-frank-large.sh

@@ -37,6 +37,10 @@ fi
 : "${CACHE_TYPE_V:=q8_0}"
 : "${BATCH_SIZE:=2048}"
 : "${UBATCH_SIZE:=512}"
+: "${TEMP:=1.0}"
+: "${TOP_K:=20}"
+: "${TOP_P:=0.95}"
+: "${MIN_P:=0.0}"
 : "${THREADS:=12}"
 : "${THREADS_HTTP:=1}"
 
@@ -68,6 +72,10 @@ exec "$LLAMA_SERVER" \
     --parallel "$PARALLEL" \
     --batch-size "$BATCH_SIZE" \
     --ubatch-size "$UBATCH_SIZE" \
+    --temp "$TEMP" \
+    --top-k "$TOP_K" \
+    --top-p "$TOP_P" \
+    --min-p "$MIN_P" \
     --threads "$THREADS" \
     --threads-http "$THREADS_HTTP" \
     --alias "$ALIAS" \

scripts/serve-llama-frank-small.sh

@@ -32,12 +32,16 @@ fi
 : "${NGRAM_MOD_N_MATCH:=24}"
 : "${CACHE_TYPE_K:=q8_0}"
 : "${CACHE_TYPE_V:=f16}"
-: "${N_CPU_MOE:=40}"     # all 40 layers: MoE expert weights stay on CPU
+: "${N_CPU_MOE:=34}"     # 34 of 40 layers: MoE expert weights stay on CPU; 6 on GPU
 : "${N_GPU_LAYERS:=all}" # every non-expert tensor goes to GPU
 : "${BATCH_SIZE:=2048}"
 : "${UBATCH_SIZE:=512}"
 : "${CTX_CHECKPOINTS:=0}"
 : "${NO_MMAP:=0}"
+: "${TEMP:=1.0}"
+: "${TOP_K:=20}"
+: "${TOP_P:=0.95}"
+: "${MIN_P:=0.0}"
 : "${THREADS:=8}"        # 7700X physical cores
 : "${THREADS_HTTP:=8}"
 
@@ -70,6 +74,10 @@ exec "$LLAMA_SERVER" \
     --ubatch-size "$UBATCH_SIZE" \
     --ctx-checkpoints "$CTX_CHECKPOINTS" \
     $([ "$NO_MMAP" = "1" ] && printf -- "--no-mmap") \
+    --temp "$TEMP" \
+    --top-k "$TOP_K" \
+    --top-p "$TOP_P" \
+    --min-p "$MIN_P" \
     --threads "$THREADS" \
     --threads-http "$THREADS_HTTP" \
     --alias "$ALIAS" \