frank-default-small: ncpu-moe=35(5GPU), ubatch=2048, ctx-checkpoints=16

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: 36 of 40 layers keep MoE weights on CPU; 4 on GPU.
-N_CPU_MOE=36
+# Expert offload: 35 of 40 layers keep MoE weights on CPU; 5 on GPU.
+N_CPU_MOE=35
 N_GPU_LAYERS=all
 
 # KV cache on GPU, both K and V at f16.
@@ -41,11 +41,11 @@ CACHE_TYPE_V=f16
 
 VERBOSITY=4
 # Store SWA/context checkpoints in RAM to hold peak GPU memory down.
-CTX_CHECKPOINTS=2
+CTX_CHECKPOINTS=16
 
-# Big prefill batches (ubatch 1536, batch 4096).
+# Big prefill batches (ubatch 2048, batch 4096).
 BATCH_SIZE=4096
-UBATCH_SIZE=1536
+UBATCH_SIZE=2048
 
 # Eager RAM load (MoE experts on CPU; ~31 GB RAM).
 NO_MMAP=1

scripts/serve-frank-small.sh

@@ -36,11 +36,11 @@ fi
 : "${CACHE_TYPE_K:=f16}"
 : "${CACHE_TYPE_V:=f16}"
 : "${VERBOSITY:=4}"
-: "${N_CPU_MOE:=36}"
+: "${N_CPU_MOE:=35}"
 : "${N_GPU_LAYERS:=all}"
 : "${BATCH_SIZE:=4096}"
-: "${UBATCH_SIZE:=1024}"
-: "${CTX_CHECKPOINTS:=2}"
+: "${UBATCH_SIZE:=2048}"
+: "${CTX_CHECKPOINTS:=16}"
 : "${NO_MMAP:=1}"
 : "${TEMPERATURE:=1.0}"
 : "${TOP_K:=20}"