test: Size the HSTU KV cache for a CI slice of a node

FlexKV's client waits 30s for its server, which is ready only once the host
cache is prepared. The stock 19GB sizing overruns that on the CPUs a CI job
gets, so serve from a 6GB copy of the config. The export applies the same
sizing, and the GPU page count the exported package depends on is untouched.

Author: yinggeh

qa/L0_torch_aoti_hstu/test.sh

@@ -61,7 +61,18 @@ BACKENDS=${BACKENDS:="pytorch"}
 export BACKENDS
 
 export FLEXKV_LOG_LEVEL=${FLEXKV_LOG_LEVEL:="WARNING"}
-export KVCACHE_MANAGER_CONFIG_FILE=${KVCACHE_MANAGER_CONFIG_FILE:=${AOTI_DIR}/kvcache_cpp_runtime.yaml}
+
+# The KV-cache client waits 30s for its server, and the server is ready only
+# once it has prepared the host cache, which the stock 19GB sizing does not
+# manage in time on a CI slice of a node. The export applies the same sizing,
+# and the GPU-side page count the exported package depends on is untouched.
+HOST_CAPACITY_PER_LAYER=${HSTU_HOST_CAPACITY_PER_LAYER:=2147483648}
+CPU_CACHE_GB=${HSTU_CPU_CACHE_GB:=6}
+KVCACHE_CONFIG="${TESTDIR}/kvcache_cpp_runtime.yaml"
+sed -e "s/^host_capacity_per_layer: .*/host_capacity_per_layer: ${HOST_CAPACITY_PER_LAYER}/" \
+    -e "s/^cpu_cache_gb: .*/cpu_cache_gb: ${CPU_CACHE_GB}/" \
+    ${AOTI_DIR}/kvcache_cpp_runtime.yaml > ${KVCACHE_CONFIG}
+export KVCACHE_MANAGER_CONFIG_FILE=${KVCACHE_MANAGER_CONFIG_FILE:=${KVCACHE_CONFIG}}
 
 KVCACHE_PID=0