frank-default-small: ncpu-moe=38 (2 experts on GPU)

scripts/frank-default-small.env

@@ -31,8 +31,8 @@ NGRAM_MOD_N_MIN=48
 NGRAM_MOD_N_MAX=64
 NGRAM_MOD_N_MATCH=24
 
-# Expert offload: 32 of 40 layers keep MoE weights on CPU; 8 on GPU.
-N_CPU_MOE=32
+# Expert offload: 38 of 40 layers keep MoE weights on CPU; 2 on GPU.
+N_CPU_MOE=38
 N_GPU_LAYERS=all
 
 # KV cache on GPU, both K and V at f16.

scripts/serve-frank-small.sh

@@ -36,7 +36,7 @@ fi
 : "${CACHE_TYPE_K:=f16}"
 : "${CACHE_TYPE_V:=f16}"
 : "${VERBOSITY:=4}"
-: "${N_CPU_MOE:=32}"
+: "${N_CPU_MOE:=38}"
 : "${N_GPU_LAYERS:=all}"
 : "${BATCH_SIZE:=4096}"
 : "${UBATCH_SIZE:=2048}"