fix(frank-default-large): enable flash attention by pairing V cache to q8_0

Root cause of ~1 t/s long-context decode: llama.cpp build has
GGML_CUDA_FA_ALL_QUANTS=OFF, and FA requires K type == V type.
Mixed cache q8_0/f16 silently disabled FA, collapsing decode at
200k ctx. Set V to q8_0 (halves V memory too) and retune
threads 12/1 for 16 Zen5 cores.

scripts/frank-default-large.env

@@ -20,9 +20,12 @@ SPEC_NGRAM_MOD_N_MATCH=24
 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).
+# KV on GPU, quantized (K q8_0, V q8_0). V MUST match K: this llama.cpp
+# build has GGML_CUDA_FA_ALL_QUANTS=OFF, and Flash Attention requires
+# K type == V type. Mixed q8_0/f16 silently disables FA, collapsing
+# long-context decode to ~1 t/s at 200k ctx. q8_0 V halves V memory too.
 CACHE_TYPE_K=q8_0
-CACHE_TYPE_V=f16
+CACHE_TYPE_V=q8_0
 # Batch sizing constrained by ROCm memory at 602100 ctx (3x200700):
 # KV ~50GB + model 29.77GB leave only ~16GB for compute graphs. The compute
 # graph scales with UBATCH, so ubatch must stay ~512; ubatch 2048 OOMs even
@@ -30,5 +33,5 @@ CACHE_TYPE_V=f16
 # User-specified ubatch 4096 / batch 12288 is not feasible here (see bench).
 BATCH_SIZE=2048
 UBATCH_SIZE=512
-THREADS=8
-THREADS_HTTP=8
+THREADS=12
+THREADS_HTTP=1

scripts/serve-llama-frank-large.sh

@@ -34,11 +34,11 @@ fi
 : "${DEVICE:=ROCm0}"
 : "${N_GPU_LAYERS:=all}"
 : "${CACHE_TYPE_K:=q8_0}"
-: "${CACHE_TYPE_V:=f16}"
+: "${CACHE_TYPE_V:=q8_0}"
 : "${BATCH_SIZE:=2048}"
 : "${UBATCH_SIZE:=512}"
-: "${THREADS:=8}"
-: "${THREADS_HTTP:=8}"
+: "${THREADS:=12}"
+: "${THREADS_HTTP:=1}"
 
 LLAMA_SERVER="${LLAMA_SERVER:-/home/dank/llama.cpp/build/bin/llama-server}"
 MODEL="$MODEL_DIR/$MODEL_FILE"