deploy: frank-default-large final config — ROCm + draft-dflash, 3x200700 ctx (bench-fastest)

scripts/frank-default-large.env

@@ -4,25 +4,23 @@ MODEL_DIR=/home/dank/models
 ALIAS=frank-default-large
 HOST=0.0.0.0
 PORT=7770
-# 6 parallel slots x 65536 ctx each = 393216 total context pool.
-CTX_SIZE=393216
-PARALLEL=6
-# Spec: DFlash draft model layered with ngram-mod. --spec-type accepts a
-# comma-separated list; llama.cpp chains the implementations (verified in
-# common/speculative.cpp). --spec-draft-n-max must stay below the DFlash
-# block size (BF16 sidecar = 8 token block), so 8 is the ceiling.
-SPEC_TYPE=draft-dflash,ngram-mod
+# 3 parallel slots x 200700 ctx each = 602100 total context pool.
+CTX_SIZE=602100
+PARALLEL=3
+# Spec: DFlash draft model. Bench (4k, ROCm) showed draft-dflash fastest
+# generation (20.2 t/s): DFlash+ngram 16.9, ngram-only 8.7. ngram layering
+# reduced throughput, so it is not layered in production.
+SPEC_TYPE=draft-dflash
 SPEC_DRAFT_N_MAX=8
 SPEC_NGRAM_MOD_N_MIN=48
 SPEC_NGRAM_MOD_N_MAX=64
 SPEC_NGRAM_MOD_N_MATCH=24
-# Backend device for offload, from --list-devices. Pick the faster of:
-#   ROCm0    (HIP / ROCm backend)
-#   Vulkan0  (RADV Vulkan backend)
+# Backend device from --list-devices. ROCm wins on generation; Vulkan renders
+# DFlash spec unusable (NONE-op abort) and gen is slower on ngram (6.9 t/s).
 DEVICE=ROCm0
 # Dense model (Qwen3.6-27B, not MoE): no MoE CPU offload. Full offload.
 N_GPU_LAYERS=all
-# KV on GPU, quantized (K q8_0, V f16); mirrors the small service profile.
+# KV on GPU, quantized (K q8_0, V f16).
 CACHE_TYPE_K=q8_0
 CACHE_TYPE_V=f16
 BATCH_SIZE=2048

scripts/serve-llama-frank-large.sh

@@ -24,9 +24,9 @@ fi
 : "${ALIAS:=frank-default-large}"
 : "${HOST:=0.0.0.0}"
 : "${PORT:=7770}"
-: "${CTX_SIZE:=393216}"
-: "${PARALLEL:=6}"
-: "${SPEC_TYPE:=draft-dflash,ngram-mod}"
+: "${CTX_SIZE:=602100}"
+: "${PARALLEL:=3}"
+: "${SPEC_TYPE:=draft-dflash}"
 : "${SPEC_DRAFT_N_MAX:=8}"
 : "${SPEC_NGRAM_MOD_N_MIN:=48}"
 : "${SPEC_NGRAM_MOD_N_MAX:=64}"