[AMD][AgentX] dsv4-fp4-mi355x-sglang-agentic-mtp: reduce mem_fraction_static to 0.85 and chunked_prefill_size to 4096 to avoid runtime OOM

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: seungrokj

benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh

@@ -82,8 +82,8 @@ export AIPERF_HTTP_TCP_USER_TIMEOUT=1000000
 USE_SGLANG_ROUTER=false
 SGLANG_BACKEND_PORT="$PORT"
 ROUTER_LOG="$RESULT_DIR/router.log"
-MEM_FRACTION_STATIC=0.90
-CHUNKED_PREFILL_SIZE=8192
+MEM_FRACTION_STATIC=0.85
+CHUNKED_PREFILL_SIZE=4096
 PARALLEL_ARGS=(--tensor-parallel-size "$TP")
 if [ "$DP_ATTENTION" = "true" ]; then
     USE_SGLANG_ROUTER=true
@@ -98,7 +98,7 @@ if [ "$DP_ATTENTION" = "true" ]; then
     export SGLANG_DP_USE_REDUCE_SCATTER=1
     export GPU_MAX_HW_QUEUES=5
 
-    CHUNKED_PREFILL_SIZE=$((8192 * TP))
+    CHUNKED_PREFILL_SIZE=$((4096 * TP))
     PARALLEL_ARGS+=(
         --dp "$TP"
         --enable-dp-attention