perf: frank-default-small batch 3072 / ubatch 1536 + ctx-checkpoints

Bigger logical/physical batches for prefill throughput. Enable
--ctx-checkpoints to park attention checkpoints in RAM, holding peak GPU
memory down so the larger ubatch fits the 24GB VRAM alongside 3x262144 KV
without OOM. Weight split unchanged (experts CPU, non-experts GPU).

scripts/frank-default-small.env

@@ -15,5 +15,11 @@ CACHE_TYPE_V=f16
 N_CPU_MOE=40
 # Every non-expert tensor (attention, norms, embedding, LM head) goes to GPU.
 N_GPU_LAYERS=all
+# Larger batches for prefill throughput. The bigger physical batch grows GPU
+# activation buffers; ctx-checkpoints parks attention checkpoints in RAM to
+# hold peak GPU memory down so the larger ubatch fits without OOM.
+BATCH_SIZE=3072
+UBATCH_SIZE=1536
+CTX_CHECKPOINTS=4
 THREADS=8
 THREADS_HTTP=8

scripts/serve-llama-frank-small.sh

@@ -32,6 +32,9 @@ fi
 : "${CACHE_TYPE_V:=f16}"
 : "${N_CPU_MOE:=40}"     # all 40 layers: MoE expert weights stay on CPU
 : "${N_GPU_LAYERS:=all}" # every non-expert tensor goes to GPU
+: "${BATCH_SIZE:=2048}"
+: "${UBATCH_SIZE:=512}"
+: "${CTX_CHECKPOINTS:=0}"
 : "${THREADS:=8}"        # 7700X physical cores
 : "${THREADS_HTTP:=8}"
 
@@ -60,6 +63,9 @@ exec "$LLAMA_SERVER" \
     --parallel "$PARALLEL" \
     --n-cpu-moe "$N_CPU_MOE" \
     --gpu-layers "$N_GPU_LAYERS" \
+    --batch-size "$BATCH_SIZE" \
+    --ubatch-size "$UBATCH_SIZE" \
+    --ctx-checkpoints "$CTX_CHECKPOINTS" \
     --threads "$THREADS" \
     --threads-http "$THREADS_HTTP" \
     --alias "$ALIAS" \