perf: multimodal benchmark sweep (#6795)

examples/backends/vllm/launch/vllm_serve_embedding_cache.sh

+#!/bin/bash
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+MODEL="Qwen/Qwen3-VL-30B-A3B-Instruct-FP8"
+CAPACITY_GB=10
+EXTRA_ARGS=()
+
+while [[ $# -gt 0 ]]; do
+    case "$1" in
+        --model)
+            MODEL="$2"; shift 2 ;;
+        --multimodal-embedding-cache-capacity-gb)
+            CAPACITY_GB="$2"; shift 2 ;;
+        *)
+            EXTRA_ARGS+=("$1"); shift ;;
+    esac
+done
+
+# Need vLLM main or v0.17+
+EC_ARGS=()
+if [[ "$CAPACITY_GB" != "0" ]]; then
+    EC_ARGS=(--ec-transfer-config "{
+        \"ec_role\": \"ec_both\",
+        \"ec_connector\": \"DynamoMultimodalEmbeddingCacheConnector\",
+        \"ec_connector_module_path\": \"dynamo.vllm.multimodal_utils.multimodal_embedding_cache_connector\",
+        \"ec_connector_extra_config\": {\"multimodal_embedding_cache_capacity_gb\": $CAPACITY_GB}
+    }")
+fi
+
+CUDA_VISIBLE_DEVICES=2 \
+vllm serve "$MODEL" \
+    --enable-log-requests \
+    --max-model-len 16384 \
+    --gpu-memory-utilization .9 \
+    "${EC_ARGS[@]}" \
+    "${EXTRA_ARGS[@]}"