From 5e26679d3fda2d70adee4d27d78d37c9c93f345b Mon Sep 17 00:00:00 2001
From: furionw <qiwa@nvidia.com>
Date: Tue, 15 Sep 2026 21:10:11 -0700
Subject: [PATCH 1/7] bench: compare native EC H2D overlap

Co-authored-by: OpenAI Codex <codex@openai.com>
Signed-off-by: furionw <qiwa@nvidia.com>
---
 benchmarks/multimodal/sweep/README.md         |  15 ++
 .../embedding_cache/vllm_serve.yaml           |  57 ++++----
 benchmarks/multimodal/sweep/orchestrator.py   |   8 +-
 .../multimodal/sweep/repetition_plan.py       |  27 ++++
 benchmarks/multimodal/sweep/server.py         |  12 +-
 .../workflows/run_vllm_ec_uuid_repetitions.sh | 131 ++++++++++++++----
 .../multimodal/sweep/workflows/vllm_serve.sh  | 131 ++++++++++++++++--
 .../multimodal/sweep/test_orchestrator.py     |   6 +
 .../multimodal/sweep/test_repetition_plan.py  |  53 +++++++
 .../multimodal/sweep/test_server.py           |  52 +++++++
 .../sweep/test_vllm_serve_workflow.py         | 107 ++++++++++++++
 11 files changed, 529 insertions(+), 70 deletions(-)
 create mode 100644 benchmarks/multimodal/sweep/repetition_plan.py
 create mode 100644 tests/benchmarks/multimodal/sweep/test_repetition_plan.py
 create mode 100644 tests/benchmarks/multimodal/sweep/test_server.py
 create mode 100644 tests/benchmarks/multimodal/sweep/test_vllm_serve_workflow.py

diff --git a/benchmarks/multimodal/sweep/README.md b/benchmarks/multimodal/sweep/README.md
index 40f57abe92d8..8eafc1f52555 100644
--- a/benchmarks/multimodal/sweep/README.md
+++ b/benchmarks/multimodal/sweep/README.md
@@ -57,6 +57,21 @@ configs:
     extra_args: [--no-enable-prefix-caching, --multimodal-embedding-cache-capacity-gb, "10"]
 ```
 
+### vLLM workflow environment
+
+| Variable | Purpose |
+|---|---|
+| `DYN_DISABLE_NSYS` | Set to `0` to profile the vLLM server; defaults to `1`. |
+| `DYN_NSYS_BIN` | Nsight Systems executable path. |
+| `DYN_NSYS_DIR` / `DYN_NSYS_TMPDIR` | Final report and temporary capture directories. |
+| `DYN_NSYS_TRACE` | Nsight trace domains; defaults to `cuda,nvtx`. |
+| `DYN_NSYS_OUTPUT_PREFIX` | Report prefix; the orchestrator appends the arm label. |
+| `DYN_SERVER_TERMINATE_TIMEOUT` | Orchestrator shutdown timeout; defaults to 300 seconds with profiling and 15 otherwise. |
+| `DYN_SERVER_SHUTDOWN_GRACE_SECONDS` | Wrapper grace period before SIGKILL; defaults to 150 seconds with profiling and 10 otherwise. |
+| `DYN_PYTHON` | Python executable used by the repetition wrapper. |
+| `VLLM_SOURCE_REVISION` | Required tested-vLLM revision recorded by the repetition wrapper. |
+
+
 ## CLI Overrides
 
 Any top-level YAML field can be overridden from the command line:
diff --git a/benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml b/benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml
index a87b92d6cd42..2f1fa76c0c33 100644
--- a/benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml
+++ b/benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml
@@ -1,7 +1,7 @@
 # SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
-# Qwen3.5-122B three-arm embedding-cache comparison on 2xH100.
+# Qwen3.5-122B two-arm native embedding-cache comparison on 2xH100.
 # Repeated image payloads carry stable UUIDs and are stripped after their first
 # use in each session. The mm processor cache is deliberately identical across
 # arms; only the post-encoder embedding-cache connector changes.
@@ -13,7 +13,7 @@ conversation_num: 30
 warmup_count: 2
 port: 8000
 timeout: 2400
-output_dir: /dynamo-tmp/logs/09-14/qwen35-122b-vllm-ec-uuid/default
+output_dir: /dynamo-tmp/logs/09-15/qwen35-122b-vllm-ec-h2d-overlap/default
 skip_plots: true
 restart_server_every_benchmark: true
 uuid_and_strip: true
@@ -79,32 +79,33 @@ configs:
       - --ec-transfer-config
       - '{"ec_connector":"ECCPUConnector","ec_role":"ec_both","ec_connector_extra_config":{"ec_cpu_bytes":4294967296}}'
 
-  - label: vllm-serve-dynamo-ec
-    workflow: benchmarks/multimodal/sweep/workflows/vllm_serve.sh
-    extra_args:
-      - --tensor-parallel-size
-      - "2"
-      - --quantization
-      - fp8
-      - --enable-expert-parallel
-      - --gpu-memory-utilization
-      - "0.90"
-      - --max-model-len
-      - "32768"
-      - --max-num-batched-tokens
-      - "32768"
-      - --max-num-seqs
-      - "50"
-      - --mm-encoder-tp-mode
-      - data
-      - --mm-processor-cache-gb
-      - "30"
-      - --limit-mm-per-prompt
-      - '{"image":5}'
-      - --no-enable-prefix-caching
-      - --no-enable-log-requests
-      - --multimodal-embedding-cache-capacity-gb
-      - "4"
+  # Dynamo EC is parked while this experiment isolates native vLLM EC.
+  # - label: vllm-serve-dynamo-ec
+  #   workflow: benchmarks/multimodal/sweep/workflows/vllm_serve.sh
+  #   extra_args:
+  #     - --tensor-parallel-size
+  #     - "2"
+  #     - --quantization
+  #     - fp8
+  #     - --enable-expert-parallel
+  #     - --gpu-memory-utilization
+  #     - "0.90"
+  #     - --max-model-len
+  #     - "32768"
+  #     - --max-num-batched-tokens
+  #     - "32768"
+  #     - --max-num-seqs
+  #     - "50"
+  #     - --mm-encoder-tp-mode
+  #     - data
+  #     - --mm-processor-cache-gb
+  #     - "30"
+  #     - --limit-mm-per-prompt
+  #     - '{"image":5}'
+  #     - --no-enable-prefix-caching
+  #     - --no-enable-log-requests
+  #     - --multimodal-embedding-cache-capacity-gb
+  #     - "4"
 
 # Historical cache-on/off matrix retained for reference. Keep it commented so
 # the canonical file continues to describe one runnable experiment.
diff --git a/benchmarks/multimodal/sweep/orchestrator.py b/benchmarks/multimodal/sweep/orchestrator.py
index 73c070433d01..5514f450a265 100644
--- a/benchmarks/multimodal/sweep/orchestrator.py
+++ b/benchmarks/multimodal/sweep/orchestrator.py
@@ -118,6 +118,10 @@ def _run_config(
 ) -> None:
     """Run all sweep values for a single benchmark config."""
     workflow_abs = _resolve_workflow(bench_cfg.workflow, repo_root)
+    arm_env_overrides = {
+        **env_overrides,
+        "DYN_BENCHMARK_ARM": bench_cfg.label,
+    }
     _print_banner(f"Config: {bench_cfg.label}", char="#")
 
     # Collect pending runs, skipping those with existing results.
@@ -151,7 +155,7 @@ def _run_config(
             workflow_script=workflow_abs,
             model=config.model,
             extra_args=bench_cfg.extra_args,
-            env_overrides=env_overrides,
+            env_overrides=arm_env_overrides,
         )
 
     try:
@@ -166,7 +170,7 @@ def _run_config(
                     workflow_script=workflow_abs,
                     model=config.model,
                     extra_args=bench_cfg.extra_args,
-                    env_overrides=env_overrides,
+                    env_overrides=arm_env_overrides,
                 )
 
             try:
diff --git a/benchmarks/multimodal/sweep/repetition_plan.py b/benchmarks/multimodal/sweep/repetition_plan.py
new file mode 100644
index 000000000000..bdb6ef77a44a
--- /dev/null
+++ b/benchmarks/multimodal/sweep/repetition_plan.py
@@ -0,0 +1,27 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+from typing import Any
+
+
+def balanced_config_orders(
+    configs: Sequence[dict[str, Any]], repetitions: int
+) -> list[list[dict[str, Any]]]:
+    """Return deterministic rotations followed by their reverse orders."""
+    if not configs:
+        raise ValueError("At least one benchmark config is required")
+    if repetitions < 1:
+        raise ValueError("repetitions must be positive")
+
+    labels = [config["label"] for config in configs]
+    if len(labels) != len(set(labels)):
+        raise ValueError("Benchmark config labels must be unique")
+
+    rotations = [
+        list(configs[index:]) + list(configs[:index]) for index in range(len(configs))
+    ]
+    balanced = rotations + [list(reversed(order)) for order in rotations]
+    return [balanced[index % len(balanced)] for index in range(repetitions)]
diff --git a/benchmarks/multimodal/sweep/server.py b/benchmarks/multimodal/sweep/server.py
index a8f1ae4ec378..2036df7ae1d6 100644
--- a/benchmarks/multimodal/sweep/server.py
+++ b/benchmarks/multimodal/sweep/server.py
@@ -21,6 +21,7 @@ class ServerManager:
     def __init__(self, port: int = 8000, timeout: int = 600) -> None:
         self.port = port
         self.timeout = timeout
+        self.terminate_timeout = 15.0
         self._process: Optional[subprocess.Popen] = None
 
     @property
@@ -50,6 +51,15 @@ def start(
         env = os.environ.copy()
         if env_overrides:
             env.update(env_overrides)
+        env["DYN_HTTP_PORT"] = str(self.port)
+        default_terminate_timeout = (
+            300.0 if env.get("DYN_DISABLE_NSYS", "1") != "1" else 15.0
+        )
+        self.terminate_timeout = float(
+            env.get("DYN_SERVER_TERMINATE_TIMEOUT", default_terminate_timeout)
+        )
+        if self.terminate_timeout <= 0:
+            raise ValueError("DYN_SERVER_TERMINATE_TIMEOUT must be positive")
 
         print(f"Launching: {' '.join(cmd)}", flush=True)
         self._process = subprocess.Popen(
@@ -111,7 +121,7 @@ def stop(self) -> None:
                 pass
 
         try:
-            self._process.wait(timeout=15)
+            self._process.wait(timeout=self.terminate_timeout)
         except subprocess.TimeoutExpired:
             try:
                 os.killpg(pid, signal.SIGKILL)
diff --git a/benchmarks/multimodal/sweep/workflows/run_vllm_ec_uuid_repetitions.sh b/benchmarks/multimodal/sweep/workflows/run_vllm_ec_uuid_repetitions.sh
index 4e5d795f3486..f1b6a43a1eda 100755
--- a/benchmarks/multimodal/sweep/workflows/run_vllm_ec_uuid_repetitions.sh
+++ b/benchmarks/multimodal/sweep/workflows/run_vllm_ec_uuid_repetitions.sh
@@ -5,8 +5,11 @@
 set -euo pipefail
 
 CONFIG="${1:-benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml}"
-OUTPUT_BASE="${2:-/dynamo-tmp/logs/09-14/qwen35-122b-vllm-ec-uuid}"
-REPETITIONS="${3:-3}"
+OUTPUT_BASE="${2:-/dynamo-tmp/logs/09-15/qwen35-122b-vllm-ec-h2d-overlap}"
+REPETITIONS="${3:-4}"
+PYTHON_BIN="${DYN_PYTHON:-python}"
+
+: "${VLLM_SOURCE_REVISION:?VLLM_SOURCE_REVISION must identify the tested vLLM commit}"
 
 if [[ ! "$REPETITIONS" =~ ^[1-9][0-9]*$ ]]; then
     echo "REPETITIONS must be a positive integer, got: $REPETITIONS" >&2
@@ -14,11 +17,12 @@ if [[ ! "$REPETITIONS" =~ ^[1-9][0-9]*$ ]]; then
 fi
 
 mkdir -p "$OUTPUT_BASE"
-python - "$OUTPUT_BASE/run_metadata.json" "$CONFIG" "$OUTPUT_BASE" "$REPETITIONS" <<'PY'
+"$PYTHON_BIN" - "$OUTPUT_BASE/run_metadata.json" "$CONFIG" "$OUTPUT_BASE" "$REPETITIONS" <<'PY'
 import datetime
 import json
 import os
 import pathlib
+import shutil
 import subprocess
 import sys
 from importlib.metadata import version
@@ -27,6 +31,8 @@ import dynamo._core
 import vllm
 import yaml
 
+from benchmarks.multimodal.sweep.repetition_plan import balanced_config_orders
+
 output = pathlib.Path(sys.argv[1])
 config_path = pathlib.Path(sys.argv[2])
 output_base = pathlib.Path(sys.argv[3])
@@ -36,11 +42,10 @@ configs = config["configs"]
 if not configs:
     raise ValueError(f"No configs found in {config_path}")
 
-rotations = [configs[index:] + configs[:index] for index in range(len(configs))]
-balanced_orders = rotations + [list(reversed(order)) for order in rotations]
 arm_orders = []
-for iteration in range(1, repetitions + 1):
-    ordered_configs = balanced_orders[(iteration - 1) % len(balanced_orders)]
+for iteration, ordered_configs in enumerate(
+    balanced_config_orders(configs, repetitions), start=1
+):
     labels = [item["label"] for item in ordered_configs]
     arm_orders.append(labels)
     iteration_config = dict(config)
@@ -50,20 +55,21 @@ for iteration in range(1, repetitions + 1):
     )
 
 metadata = {
-    "profile": "122b",
-    "model": "Qwen/Qwen3.5-122B-A10B-FP8",
-    "preset": "dl-H100x2",
-    "tp": 2,
-    "arms": ["vllm-serve", "vllm-serve-native-ec", "vllm-serve-dynamo-ec"],
+    "model": config["model"],
+    "arms": [item["label"] for item in configs],
     "arm_orders": arm_orders,
-    "ec_capacity_gb": 4,
-    "nvtx": False,
-    "uuid_and_strip": True,
+    "tensor_parallel_sizes": {},
+    "ec_cpu_capacity_bytes": {},
+    "nvtx": config.get("env", {}).get("DYN_DISABLE_NSYS", "1") != "1",
+    "uuid_and_strip": config.get("uuid_and_strip", False),
     "aiperf_version": subprocess.check_output(
         ["aiperf", "--version"], text=True
     ).strip(),
     "vllm_file": vllm.__file__,
     "vllm_version": vllm.__version__,
+    "vllm_executable": shutil.which("vllm"),
+    "vllm_source_revision": os.environ["VLLM_SOURCE_REVISION"],
+    "python_executable": sys.executable,
     "dynamo_core_file": dynamo._core.__file__,
     "dynamo_version": version("ai-dynamo"),
     "dynamo_runtime_version": version("ai-dynamo-runtime"),
@@ -71,37 +77,112 @@ metadata = {
     "container_image_digest": os.environ["CONTAINER_IMAGE_DIGEST"],
     "container_image_file": os.environ["CONTAINER_IMAGE_FILE"],
     "harness_revision": os.environ["HARNESS_REVISION"],
-    "dataset": "/dynamo-tmp/data/30u_8t_5w_8000word_base64_uuid_seed42.jsonl",
-    "concurrency": 30,
+    "dataset": config["input_files"][0],
+    "concurrency": config["concurrencies"][0],
     "utc_start_time": datetime.datetime.now(datetime.timezone.utc).isoformat(),
 }
+for arm in configs:
+    args = arm.get("extra_args", [])
+    label = arm["label"]
+    if "--tensor-parallel-size" in args:
+        index = args.index("--tensor-parallel-size")
+        metadata["tensor_parallel_sizes"][label] = int(args[index + 1])
+    if "--ec-transfer-config" in args:
+        index = args.index("--ec-transfer-config")
+        ec_config = json.loads(args[index + 1])
+        capacity = ec_config.get("ec_connector_extra_config", {}).get("ec_cpu_bytes")
+        if capacity is not None:
+            metadata["ec_cpu_capacity_bytes"][label] = int(capacity)
 output.write_text(json.dumps(metadata, indent=2, sort_keys=True) + "\n")
 PY
 
+arms_raw="$("$PYTHON_BIN" - "$CONFIG" <<'PY'
+import pathlib
+import sys
+
+import yaml
+
+config = yaml.safe_load(pathlib.Path(sys.argv[1]).read_text())
+for item in config["configs"]:
+    print(item["label"])
+PY
+)"
+mapfile -t arms <<< "$arms_raw"
+if [[ ${#arms[@]} -eq 0 || -z "${arms[0]}" ]]; then
+    echo "No benchmark arms found in $CONFIG" >&2
+    exit 2
+fi
+
+shape_raw="$("$PYTHON_BIN" - "$CONFIG" <<'PY'
+import json
+import pathlib
+import sys
+from collections import defaultdict
+
+import yaml
+
+config = yaml.safe_load(pathlib.Path(sys.argv[1]).read_text())
+dataset = pathlib.Path(config["input_files"][0])
+seen = defaultdict(set)
+included_sessions = set()
+conversation_limit = config.get("conversation_num")
+content = 0
+stripped = 0
+for line_index, line in enumerate(dataset.open()):
+    item = json.loads(line)
+    session_id = item.get("session_id", f"row-{line_index}")
+    if session_id not in included_sessions:
+        if (
+            conversation_limit is not None
+            and len(included_sessions) >= conversation_limit
+        ):
+            continue
+        included_sessions.add(session_id)
+    session_seen = seen[session_id]
+    for image_uuid in item.get("image_uuids", []):
+        if image_uuid in session_seen:
+            stripped += 1
+        else:
+            session_seen.add(image_uuid)
+            content += 1
+
+print(dataset.stem.replace(" ", "_"))
+print(f"concurrency{config['concurrencies'][0]}")
+print(content)
+print(stripped)
+PY
+)"
+mapfile -t sweep_shape <<< "$shape_raw"
+dataset_tag="${sweep_shape[0]}"
+sweep_tag="${sweep_shape[1]}"
+expected_content="${sweep_shape[2]}"
+expected_stripped="${sweep_shape[3]}"
+
 for ((iteration = 1; iteration <= REPETITIONS; iteration++)); do
     iteration_config="$OUTPUT_BASE/config-rep-$iteration.yaml"
-    iteration_order="$(python - "$iteration_config" <<'PY'
+    iteration_order="$("$PYTHON_BIN" - "$iteration_config" <<'PY'
+import pathlib
 import sys
 
 import yaml
 
-config = yaml.safe_load(open(sys.argv[1]))
+config = yaml.safe_load(pathlib.Path(sys.argv[1]).read_text())
 print(" -> ".join(item["label"] for item in config["configs"]))
 PY
 )"
     echo "[sweep] ITERATION_ORDER_${iteration}=${iteration_order}"
-    python -m benchmarks.multimodal.sweep \
+    "$PYTHON_BIN" -m benchmarks.multimodal.sweep \
         --config "$iteration_config" \
         --output-dir "$OUTPUT_BASE/rep-$iteration" \
         --skip-plots
     echo "[sweep] END_ITER_${iteration}"
     if [[ "$iteration" == "1" ]]; then
-        for arm in vllm-serve vllm-serve-native-ec vllm-serve-dynamo-ec; do
-            artifact="$OUTPUT_BASE/rep-1/30u_8t_5w_8000word_base64_uuid_seed42/$arm/concurrency30"
-            python -m benchmarks.multimodal.jsonl.validate_uuid_transport \
+        for arm in "${arms[@]}"; do
+            artifact="$OUTPUT_BASE/rep-1/$dataset_tag/$arm/$sweep_tag"
+            "$PYTHON_BIN" -m benchmarks.multimodal.jsonl.validate_uuid_transport \
                 "$artifact/inputs.json" \
-                --expect-content 360 \
-                --expect-stripped 840 \
+                --expect-content "$expected_content" \
+                --expect-stripped "$expected_stripped" \
                 --output "$artifact/uuid_transport_summary.json"
         done
         echo "[sweep] UUID_TRANSPORT_VALIDATED"
diff --git a/benchmarks/multimodal/sweep/workflows/vllm_serve.sh b/benchmarks/multimodal/sweep/workflows/vllm_serve.sh
index 313c9108fcfc..4d9cf2524b9d 100755
--- a/benchmarks/multimodal/sweep/workflows/vllm_serve.sh
+++ b/benchmarks/multimodal/sweep/workflows/vllm_serve.sh
@@ -2,12 +2,19 @@
 # SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 #
-# Minimal vllm serve wrapper for benchmark sweeps.
-# Launched by the sweep orchestrator via: bash vllm_serve.sh --model <model> [extra_args...]
+# vLLM serve wrapper for benchmark sweeps with opt-in Nsight Systems capture.
+
+set -euo pipefail
+
+SCRIPT_DIR="$(dirname "$(readlink -f "$0")")"
+SWEEP_REPO_ROOT="$(readlink -f "$SCRIPT_DIR/../../../..")"
+source "$SWEEP_REPO_ROOT/examples/common/gpu_utils.sh"
+source "$SWEEP_REPO_ROOT/examples/common/launch_utils.sh"
 
 MODEL=""
 CAPACITY_GB=0
 EXTRA_ARGS=()
+MAX_MODEL_LEN="${MAX_MODEL_LEN:-16384}"
 
 while [[ $# -gt 0 ]]; do
     case "$1" in
@@ -20,6 +27,18 @@ while [[ $# -gt 0 ]]; do
     esac
 done
 
+has_gpu_mem_override=0
+for ((i = 0; i < ${#EXTRA_ARGS[@]}; i++)); do
+    case "${EXTRA_ARGS[$i]}" in
+        --max-model-len)
+            MAX_MODEL_LEN="${EXTRA_ARGS[$((i + 1))]}"
+            ;;
+        --gpu-memory-utilization|--kv-cache-memory-bytes)
+            has_gpu_mem_override=1
+            ;;
+    esac
+done
+
 if [[ -z "$MODEL" ]]; then
     echo "ERROR: --model is required" >&2
     exit 1
@@ -35,17 +54,101 @@ if [[ "$CAPACITY_GB" != "0" ]]; then
     }")
 fi
 
-GPU_MEM_UTIL=".9"
-KV_BYTES="${_PROFILE_OVERRIDE_VLLM_KV_CACHE_BYTES:-}"
-if [[ -n "$KV_BYTES" ]]; then
-    GPU_MEM_ARGS="--kv-cache-memory-bytes $KV_BYTES --gpu-memory-utilization 0.01"
-else
-    GPU_MEM_ARGS="--gpu-memory-utilization $GPU_MEM_UTIL"
+GPU_MEM_ARGS=""
+if [[ "$has_gpu_mem_override" == "0" ]]; then
+    GPU_MEM_ARGS="$(build_vllm_gpu_mem_args)"
+    if [[ -z "$GPU_MEM_ARGS" ]]; then
+        GPU_MEM_ARGS="--gpu-memory-utilization .9"
+    fi
+fi
+GPU_MEM_ARGV=()
+if [[ -n "$GPU_MEM_ARGS" ]]; then
+    read -r -a GPU_MEM_ARGV <<< "$GPU_MEM_ARGS"
+fi
+
+HTTP_PORT="${DYN_HTTP_PORT:-8000}"
+print_launch_banner --multimodal "Launching standalone vLLM" "$MODEL" "$HTTP_PORT"
+
+VLLM_CMD=(
+    vllm serve "$MODEL"
+    --port "$HTTP_PORT"
+    --enable-log-requests
+)
+if [[ ! " ${EXTRA_ARGS[*]} " =~ [[:space:]]--max-model-len[[:space:]] ]]; then
+    VLLM_CMD+=(--max-model-len "$MAX_MODEL_LEN")
 fi
+if [[ ${#GPU_MEM_ARGV[@]} -gt 0 ]]; then
+    VLLM_CMD+=("${GPU_MEM_ARGV[@]}")
+fi
+if [[ ${#EC_ARGS[@]} -gt 0 ]]; then
+    VLLM_CMD+=("${EC_ARGS[@]}")
+fi
+if [[ ${#EXTRA_ARGS[@]} -gt 0 ]]; then
+    VLLM_CMD+=("${EXTRA_ARGS[@]}")
+fi
+
+LAUNCH_PREFIX=()
+if [[ "${DYN_DISABLE_NSYS:-1}" != "1" ]]; then
+    NSYS_BIN="${DYN_NSYS_BIN:-/opt/nvidia/nsight-systems-cli/2026.2.1/bin/nsys}"
+    if [[ ! -x "$NSYS_BIN" ]]; then
+        echo "ERROR: nsys is not executable at $NSYS_BIN" >&2
+        exit 1
+    fi
+
+    NSYS_DIR="${DYN_NSYS_DIR:-/dynamo-tmp/nsys}"
+    NSYS_TMPDIR="${DYN_NSYS_TMPDIR:-/dynamo-tmp/nsys-staging}"
+    NSYS_PREFIX="${DYN_NSYS_OUTPUT_PREFIX:-vllm}-${DYN_BENCHMARK_ARM:-standalone}"
+    mkdir -p "$NSYS_DIR" "$NSYS_TMPDIR"
+    export TMPDIR="$NSYS_TMPDIR"
 
-vllm serve "$MODEL" \
-    --enable-log-requests \
-    --max-model-len 16384 \
-    $GPU_MEM_ARGS \
-    "${EC_ARGS[@]}" \
-    "${EXTRA_ARGS[@]}"
+    timestamp="$(date +%Y%m%d_%H%M%S)"
+    nsys_output="$NSYS_DIR/${NSYS_PREFIX}_${timestamp}.nsys-rep"
+    LAUNCH_PREFIX=(
+        "$NSYS_BIN" profile
+        --trace="${DYN_NSYS_TRACE:-cuda,nvtx}"
+        --sample=none
+        --cpuctxsw=none
+        --kill=sigterm
+        --force-overwrite=true
+        -o "$nsys_output"
+    )
+    echo "[nsys] vllm-serve -> $nsys_output" >&2
+fi
+
+server_pid=0
+cleanup() {
+    local exit_code="${1:-0}"
+    trap - EXIT INT TERM
+    if [[ "$server_pid" -gt 0 ]] && kill -0 "$server_pid" 2>/dev/null; then
+        kill -INT "$server_pid" 2>/dev/null || true
+        shutdown_grace="${DYN_SERVER_SHUTDOWN_GRACE_SECONDS:-}"
+        if [[ -z "$shutdown_grace" ]]; then
+            if [[ "${DYN_DISABLE_NSYS:-1}" == "1" ]]; then
+                shutdown_grace=10
+            else
+                shutdown_grace=150
+            fi
+        fi
+        for _ in $(seq 1 "$shutdown_grace"); do
+            kill -0 "$server_pid" 2>/dev/null || break
+            sleep 1
+        done
+        if kill -0 "$server_pid" 2>/dev/null; then
+            kill -KILL "$server_pid" 2>/dev/null || true
+        fi
+        wait "$server_pid" 2>/dev/null || true
+    fi
+    exit "$exit_code"
+}
+trap 'cleanup 0' INT TERM
+trap 'cleanup $?' EXIT
+
+if [[ ${#LAUNCH_PREFIX[@]} -gt 0 ]]; then
+    "${LAUNCH_PREFIX[@]}" "${VLLM_CMD[@]}" &
+else
+    "${VLLM_CMD[@]}" &
+fi
+server_pid=$!
+wait "$server_pid"
+server_pid=0
+trap - EXIT
diff --git a/tests/benchmarks/multimodal/sweep/test_orchestrator.py b/tests/benchmarks/multimodal/sweep/test_orchestrator.py
index c1356a68aa6b..b487b96c38bb 100644
--- a/tests/benchmarks/multimodal/sweep/test_orchestrator.py
+++ b/tests/benchmarks/multimodal/sweep/test_orchestrator.py
@@ -211,3 +211,9 @@ def test_uuid_and_strip_propagates_to_aiperf(
     run_sweep(config, repo_root=tmp_path)
 
     assert mock_aiperf.call_args.kwargs["uuid_and_strip"] is True
+    assert (
+        mock_server_cls.return_value.start.call_args.kwargs["env_overrides"][
+            "DYN_BENCHMARK_ARM"
+        ]
+        == "cfg-0"
+    )
diff --git a/tests/benchmarks/multimodal/sweep/test_repetition_plan.py b/tests/benchmarks/multimodal/sweep/test_repetition_plan.py
new file mode 100644
index 000000000000..4679ca528ec5
--- /dev/null
+++ b/tests/benchmarks/multimodal/sweep/test_repetition_plan.py
@@ -0,0 +1,53 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Any
+
+import pytest
+
+from benchmarks.multimodal.sweep.repetition_plan import balanced_config_orders
+
+pytestmark = [pytest.mark.unit, pytest.mark.pre_merge, pytest.mark.gpu_0]
+
+
+def _labels(orders: list[list[dict[str, Any]]]) -> list[list[str]]:
+    return [[config["label"] for config in order] for order in orders]
+
+
+def test_two_arm_four_repetition_order_is_balanced() -> None:
+    configs = [{"label": "a"}, {"label": "b"}]
+
+    orders = balanced_config_orders(configs, repetitions=4)
+
+    assert _labels(orders) == [["a", "b"], ["b", "a"], ["b", "a"], ["a", "b"]]
+
+
+def test_three_arm_orders_cover_rotations_reverses_and_wrap() -> None:
+    configs = [{"label": "a"}, {"label": "b"}, {"label": "c"}]
+
+    orders = _labels(balanced_config_orders(configs, repetitions=8))
+
+    expected_cycle = [
+        ["a", "b", "c"],
+        ["b", "c", "a"],
+        ["c", "a", "b"],
+        ["c", "b", "a"],
+        ["a", "c", "b"],
+        ["b", "a", "c"],
+    ]
+    assert orders == expected_cycle + expected_cycle[:2]
+
+
+@pytest.mark.parametrize(
+    ("configs", "repetitions", "error"),
+    [
+        ([], 1, "At least one"),
+        ([{"label": "a"}], 0, "positive"),
+        ([{"label": "a"}, {"label": "a"}], 1, "unique"),
+    ],
+)
+def test_invalid_repetition_plan_is_rejected(
+    configs: list[dict[str, Any]], repetitions: int, error: str
+) -> None:
+    with pytest.raises(ValueError, match=error):
+        balanced_config_orders(configs, repetitions)
diff --git a/tests/benchmarks/multimodal/sweep/test_server.py b/tests/benchmarks/multimodal/sweep/test_server.py
new file mode 100644
index 000000000000..746bc888d6ac
--- /dev/null
+++ b/tests/benchmarks/multimodal/sweep/test_server.py
@@ -0,0 +1,52 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+from benchmarks.multimodal.sweep.server import ServerManager
+
+pytestmark = [pytest.mark.unit, pytest.mark.pre_merge, pytest.mark.gpu_0]
+
+
+@pytest.mark.parametrize(
+    ("env", "expected"),
+    [
+        ({"DYN_DISABLE_NSYS": "1"}, 15.0),
+        ({"DYN_DISABLE_NSYS": "0"}, 300.0),
+        ({"DYN_SERVER_TERMINATE_TIMEOUT": "42"}, 42.0),
+    ],
+)
+def test_start_resolves_termination_timeout_and_port(
+    tmp_path, env: dict[str, str], expected: float
+) -> None:
+    workflow = tmp_path / "workflow.sh"
+    workflow.write_text("#!/bin/bash\n")
+    process = MagicMock()
+
+    with (
+        patch(
+            "benchmarks.multimodal.sweep.server.subprocess.Popen",
+            return_value=process,
+        ) as popen,
+        patch.object(ServerManager, "wait_for_ready"),
+    ):
+        manager = ServerManager(port=8123)
+        manager.start(str(workflow), "model", env_overrides=env)
+
+    assert manager.terminate_timeout == expected
+    assert popen.call_args.kwargs["env"]["DYN_HTTP_PORT"] == "8123"
+
+
+def test_start_rejects_non_positive_termination_timeout(tmp_path) -> None:
+    workflow = tmp_path / "workflow.sh"
+    workflow.write_text("#!/bin/bash\n")
+    manager = ServerManager()
+
+    with pytest.raises(ValueError, match="must be positive"):
+        manager.start(
+            str(workflow),
+            "model",
+            env_overrides={"DYN_SERVER_TERMINATE_TIMEOUT": "0"},
+        )
diff --git a/tests/benchmarks/multimodal/sweep/test_vllm_serve_workflow.py b/tests/benchmarks/multimodal/sweep/test_vllm_serve_workflow.py
new file mode 100644
index 000000000000..e223d1bc88c6
--- /dev/null
+++ b/tests/benchmarks/multimodal/sweep/test_vllm_serve_workflow.py
@@ -0,0 +1,107 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+import os
+import subprocess
+import sys
+from pathlib import Path
+
+import pytest
+import yaml
+
+pytestmark = [pytest.mark.unit, pytest.mark.pre_merge, pytest.mark.gpu_0]
+
+REPO_ROOT = Path(__file__).parents[4]
+WORKFLOW = REPO_ROOT / "benchmarks/multimodal/sweep/workflows/vllm_serve.sh"
+CONFIG = REPO_ROOT / (
+    "benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml"
+)
+
+
+def test_embedding_cache_sweep_selects_only_requested_arms() -> None:
+    config = yaml.safe_load(CONFIG.read_text())
+
+    assert [item["label"] for item in config["configs"]] == [
+        "vllm-serve",
+        "vllm-serve-native-ec",
+    ]
+    assert config["env"]["DYN_DISABLE_NSYS"] == "1"
+
+
+def _run_workflow(tmp_path: Path, *args: str, enable_nsys: bool = False):
+    bin_dir = tmp_path / "bin"
+    bin_dir.mkdir()
+    vllm = bin_dir / "vllm"
+    vllm.write_text("#!/bin/bash\nprintf '[fake-vllm] %s\\n' \"$*\"\n")
+    vllm.chmod(0o755)
+    nsys_bin = bin_dir / "nsys"
+    nsys_bin.write_text(
+        "#!/bin/bash\n"
+        "printf '[fake-nsys] %s\\n' \"$*\"\n"
+        "while [[ $# -gt 0 ]]; do\n"
+        '  if [[ $1 == vllm ]]; then exec "$@"; fi\n'
+        "  shift\n"
+        "done\n"
+        "exit 2\n"
+    )
+    nsys_bin.chmod(0o755)
+
+    env = os.environ.copy()
+    env.update(
+        {
+            "DYNAMO_HOME": str(REPO_ROOT),
+            "DYN_DISABLE_NSYS": "0" if enable_nsys else "1",
+            "DYN_NSYS_BIN": str(nsys_bin),
+            "DYN_NSYS_DIR": str(tmp_path / "nsys"),
+            "DYN_NSYS_TMPDIR": str(tmp_path / "staging"),
+            "DYN_NSYS_TRACE": "cuda,nvtx",
+            "PATH": f"{bin_dir}:{env['PATH']}",
+        }
+    )
+
+    result = subprocess.run(
+        ["bash", str(WORKFLOW), "--model", "test-model", *args],
+        check=True,
+        capture_output=True,
+        env=env,
+        text=True,
+    )
+
+    return result
+
+
+@pytest.mark.skipif(sys.platform == "darwin", reason="workflow requires GNU readlink")
+def test_vllm_workflow_profiles_only_when_enabled(tmp_path: Path) -> None:
+    result = _run_workflow(tmp_path, "--max-num-seqs", "2", enable_nsys=True)
+
+    assert "[fake-nsys] profile --trace=cuda,nvtx" in result.stdout
+    assert "[fake-vllm] serve test-model" in result.stdout
+    assert "--max-num-seqs 2" in result.stdout
+
+
+@pytest.mark.skipif(sys.platform == "darwin", reason="workflow requires GNU readlink")
+def test_vllm_workflow_launches_without_profiler(tmp_path: Path) -> None:
+    result = _run_workflow(tmp_path, "--max-num-seqs", "2")
+
+    assert "[fake-nsys]" not in result.stdout
+    assert "[fake-vllm] serve test-model" in result.stdout
+    assert "--port 8000" in result.stdout
+
+
+@pytest.mark.skipif(sys.platform == "darwin", reason="workflow requires GNU readlink")
+def test_vllm_workflow_builds_dynamo_ec_config(tmp_path: Path) -> None:
+    result = _run_workflow(tmp_path, "--multimodal-embedding-cache-capacity-gb", "4")
+
+    assert "[fake-vllm] serve test-model" in result.stdout
+    assert "DynamoMultimodalEmbeddingCacheConnector" in result.stdout
+    assert '"multimodal_embedding_cache_capacity_gb": 4' in result.stdout
+
+
+@pytest.mark.skipif(sys.platform == "darwin", reason="workflow requires bash 4.3")
+def test_vllm_workflow_requires_model() -> None:
+    result = subprocess.run(["bash", str(WORKFLOW)], capture_output=True, text=True)
+
+    assert result.returncode != 0
+    assert "--model is required" in result.stderr

From 9032ce63033ab4fc11590aebac8341cbb8147d18 Mon Sep 17 00:00:00 2001
From: furionw <qiwa@nvidia.com>
Date: Tue, 15 Sep 2026 22:05:26 -0700
Subject: [PATCH 2/7] test(bench): harden EC sweep validation

Signed-off-by: furionw <qiwa@nvidia.com>
---
 benchmarks/multimodal/sweep/README.md         |  4 +
 benchmarks/multimodal/sweep/dataset_shape.py  | 50 ++++++++++
 benchmarks/multimodal/sweep/orchestrator.py   |  6 +-
 benchmarks/multimodal/sweep/server.py         | 18 +++-
 .../workflows/run_vllm_ec_uuid_repetitions.sh | 96 ++++++++++---------
 .../multimodal/sweep/workflows/vllm_serve.sh  | 30 ++++--
 .../multimodal/sweep/test_dataset_shape.py    | 58 +++++++++++
 .../multimodal/sweep/test_server.py           | 46 ++++++++-
 .../sweep/test_vllm_serve_workflow.py         | 58 +++++++++++
 9 files changed, 305 insertions(+), 61 deletions(-)
 create mode 100644 tests/benchmarks/multimodal/sweep/test_dataset_shape.py

diff --git a/benchmarks/multimodal/sweep/README.md b/benchmarks/multimodal/sweep/README.md
index 8eafc1f52555..666125c9c826 100644
--- a/benchmarks/multimodal/sweep/README.md
+++ b/benchmarks/multimodal/sweep/README.md
@@ -70,6 +70,10 @@ configs:
 | `DYN_SERVER_SHUTDOWN_GRACE_SECONDS` | Wrapper grace period before SIGKILL; defaults to 150 seconds with profiling and 10 otherwise. |
 | `DYN_PYTHON` | Python executable used by the repetition wrapper. |
 | `VLLM_SOURCE_REVISION` | Required tested-vLLM revision recorded by the repetition wrapper. |
+| `CONTAINER_IMAGE` | Required runtime image reference recorded by the repetition wrapper. |
+| `CONTAINER_IMAGE_DIGEST` | Required immutable runtime image digest. |
+| `CONTAINER_IMAGE_FILE` | Required imported image/squashfs path used by the GPU run. |
+| `HARNESS_REVISION` | Required Dynamo benchmark-harness commit. |
 
 
 ## CLI Overrides
diff --git a/benchmarks/multimodal/sweep/dataset_shape.py b/benchmarks/multimodal/sweep/dataset_shape.py
index 5360ff2a7428..bb4bced8337d 100644
--- a/benchmarks/multimodal/sweep/dataset_shape.py
+++ b/benchmarks/multimodal/sweep/dataset_shape.py
@@ -4,7 +4,9 @@
 from __future__ import annotations
 
 import json
+from collections import defaultdict
 from pathlib import Path
+from typing import Hashable
 
 
 def count_session_ids(jsonl_path: str | Path) -> int:
@@ -31,3 +33,51 @@ def count_session_ids(jsonl_path: str | Path) -> int:
             else:
                 sessions.add(str(sid))
     return len(sessions) + anon_count
+
+
+def count_uuid_expectations(
+    jsonl_path: str | Path,
+    conversation_num: int | None = None,
+) -> tuple[int, int]:
+    """Count image payloads expected to be present and stripped by AIPerf.
+
+    UUIDs are deduplicated independently within each session. Only the first
+    ``conversation_num`` sessions are included, matching the sequential
+    sampler used by the sweep. Missing or null session IDs make a row its own
+    anonymous session, using the same policy as :func:`count_session_ids`.
+    """
+    if conversation_num is not None and conversation_num < 0:
+        raise ValueError("conversation_num must be non-negative")
+
+    seen: dict[Hashable, set[Hashable]] = defaultdict(set)
+    included_sessions: set[Hashable] = set()
+    content = 0
+    stripped = 0
+
+    with open(jsonl_path) as f:
+        for line_index, line in enumerate(f):
+            line = line.strip()
+            if not line:
+                continue
+            row = json.loads(line)
+            sid = row.get("session_id")
+            session_key: Hashable = (
+                ("anonymous", line_index) if sid is None else ("session", str(sid))
+            )
+            if session_key not in included_sessions:
+                if (
+                    conversation_num is not None
+                    and len(included_sessions) >= conversation_num
+                ):
+                    continue
+                included_sessions.add(session_key)
+
+            session_seen = seen[session_key]
+            for image_uuid in row.get("image_uuids", []):
+                if image_uuid in session_seen:
+                    stripped += 1
+                else:
+                    session_seen.add(image_uuid)
+                    content += 1
+
+    return content, stripped
diff --git a/benchmarks/multimodal/sweep/orchestrator.py b/benchmarks/multimodal/sweep/orchestrator.py
index 5514f450a265..2490a3bd2f08 100644
--- a/benchmarks/multimodal/sweep/orchestrator.py
+++ b/benchmarks/multimodal/sweep/orchestrator.py
@@ -121,6 +121,7 @@ def _run_config(
     arm_env_overrides = {
         **env_overrides,
         "DYN_BENCHMARK_ARM": bench_cfg.label,
+        "DYN_BENCHMARK_SWEEP": "multi",
     }
     _print_banner(f"Config: {bench_cfg.label}", char="#")
 
@@ -170,7 +171,10 @@ def _run_config(
                     workflow_script=workflow_abs,
                     model=config.model,
                     extra_args=bench_cfg.extra_args,
-                    env_overrides=arm_env_overrides,
+                    env_overrides={
+                        **arm_env_overrides,
+                        "DYN_BENCHMARK_SWEEP": f"{sweep_mode}{value}",
+                    },
                 )
 
             try:
diff --git a/benchmarks/multimodal/sweep/server.py b/benchmarks/multimodal/sweep/server.py
index 2036df7ae1d6..24390d9b104c 100644
--- a/benchmarks/multimodal/sweep/server.py
+++ b/benchmarks/multimodal/sweep/server.py
@@ -52,14 +52,24 @@ def start(
         if env_overrides:
             env.update(env_overrides)
         env["DYN_HTTP_PORT"] = str(self.port)
-        default_terminate_timeout = (
-            300.0 if env.get("DYN_DISABLE_NSYS", "1") != "1" else 15.0
-        )
+        profiling = env.get("DYN_DISABLE_NSYS", "1") != "1"
+        default_terminate_timeout = 300.0 if profiling else 15.0
+        default_shutdown_grace = 150.0 if profiling else 10.0
+        raw_terminate_timeout = env.get("DYN_SERVER_TERMINATE_TIMEOUT")
+        raw_shutdown_grace = env.get("DYN_SERVER_SHUTDOWN_GRACE_SECONDS")
         self.terminate_timeout = float(
-            env.get("DYN_SERVER_TERMINATE_TIMEOUT", default_terminate_timeout)
+            raw_terminate_timeout or default_terminate_timeout
         )
+        shutdown_grace = float(raw_shutdown_grace or default_shutdown_grace)
         if self.terminate_timeout <= 0:
             raise ValueError("DYN_SERVER_TERMINATE_TIMEOUT must be positive")
+        if shutdown_grace <= 0:
+            raise ValueError("DYN_SERVER_SHUTDOWN_GRACE_SECONDS must be positive")
+        if self.terminate_timeout <= shutdown_grace:
+            raise ValueError(
+                "DYN_SERVER_TERMINATE_TIMEOUT must exceed "
+                "DYN_SERVER_SHUTDOWN_GRACE_SECONDS"
+            )
 
         print(f"Launching: {' '.join(cmd)}", flush=True)
         self._process = subprocess.Popen(
diff --git a/benchmarks/multimodal/sweep/workflows/run_vllm_ec_uuid_repetitions.sh b/benchmarks/multimodal/sweep/workflows/run_vllm_ec_uuid_repetitions.sh
index f1b6a43a1eda..a3a4d9058207 100755
--- a/benchmarks/multimodal/sweep/workflows/run_vllm_ec_uuid_repetitions.sh
+++ b/benchmarks/multimodal/sweep/workflows/run_vllm_ec_uuid_repetitions.sh
@@ -10,6 +10,10 @@ REPETITIONS="${3:-4}"
 PYTHON_BIN="${DYN_PYTHON:-python}"
 
 : "${VLLM_SOURCE_REVISION:?VLLM_SOURCE_REVISION must identify the tested vLLM commit}"
+: "${CONTAINER_IMAGE:?CONTAINER_IMAGE must identify the tested runtime image}"
+: "${CONTAINER_IMAGE_DIGEST:?CONTAINER_IMAGE_DIGEST must identify the tested image digest}"
+: "${CONTAINER_IMAGE_FILE:?CONTAINER_IMAGE_FILE must identify the imported image file}"
+: "${HARNESS_REVISION:?HARNESS_REVISION must identify the benchmark harness commit}"
 
 if [[ ! "$REPETITIONS" =~ ^[1-9][0-9]*$ ]]; then
     echo "REPETITIONS must be a positive integer, got: $REPETITIONS" >&2
@@ -41,6 +45,9 @@ config = yaml.safe_load(config_path.read_text())
 configs = config["configs"]
 if not configs:
     raise ValueError(f"No configs found in {config_path}")
+concurrencies = config.get("concurrencies")
+sweep_mode = "concurrency" if concurrencies else "request_rate"
+sweep_values = concurrencies or config.get("request_rates") or [4, 8, 16, 32, 64]
 
 arm_orders = []
 for iteration, ordered_configs in enumerate(
@@ -60,7 +67,10 @@ metadata = {
     "arm_orders": arm_orders,
     "tensor_parallel_sizes": {},
     "ec_cpu_capacity_bytes": {},
-    "nvtx": config.get("env", {}).get("DYN_DISABLE_NSYS", "1") != "1",
+    "nvtx": config.get("env", {}).get(
+        "DYN_DISABLE_NSYS", os.environ.get("DYN_DISABLE_NSYS", "1")
+    )
+    != "1",
     "uuid_and_strip": config.get("uuid_and_strip", False),
     "aiperf_version": subprocess.check_output(
         ["aiperf", "--version"], text=True
@@ -77,8 +87,9 @@ metadata = {
     "container_image_digest": os.environ["CONTAINER_IMAGE_DIGEST"],
     "container_image_file": os.environ["CONTAINER_IMAGE_FILE"],
     "harness_revision": os.environ["HARNESS_REVISION"],
-    "dataset": config["input_files"][0],
-    "concurrency": config["concurrencies"][0],
+    "datasets": config["input_files"],
+    "sweep_mode": sweep_mode,
+    "sweep_values": sweep_values,
     "utc_start_time": datetime.datetime.now(datetime.timezone.utc).isoformat(),
 }
 for arm in configs:
@@ -114,50 +125,38 @@ if [[ ${#arms[@]} -eq 0 || -z "${arms[0]}" ]]; then
 fi
 
 shape_raw="$("$PYTHON_BIN" - "$CONFIG" <<'PY'
-import json
 import pathlib
 import sys
-from collections import defaultdict
 
 import yaml
 
+from benchmarks.multimodal.sweep.config import input_file_tag
+from benchmarks.multimodal.sweep.dataset_shape import count_uuid_expectations
+
 config = yaml.safe_load(pathlib.Path(sys.argv[1]).read_text())
-dataset = pathlib.Path(config["input_files"][0])
-seen = defaultdict(set)
-included_sessions = set()
-conversation_limit = config.get("conversation_num")
-content = 0
-stripped = 0
-for line_index, line in enumerate(dataset.open()):
-    item = json.loads(line)
-    session_id = item.get("session_id", f"row-{line_index}")
-    if session_id not in included_sessions:
-        if (
-            conversation_limit is not None
-            and len(included_sessions) >= conversation_limit
-        ):
-            continue
-        included_sessions.add(session_id)
-    session_seen = seen[session_id]
-    for image_uuid in item.get("image_uuids", []):
-        if image_uuid in session_seen:
-            stripped += 1
-        else:
-            session_seen.add(image_uuid)
-            content += 1
-
-print(dataset.stem.replace(" ", "_"))
-print(f"concurrency{config['concurrencies'][0]}")
-print(content)
-print(stripped)
+concurrencies = config.get("concurrencies")
+sweep_mode = "concurrency" if concurrencies else "request_rate"
+sweep_values = concurrencies or config.get("request_rates") or [4, 8, 16, 32, 64]
+for dataset in config["input_files"]:
+    content, stripped = count_uuid_expectations(
+        dataset, conversation_num=config.get("conversation_num")
+    )
+    for value in sorted(sweep_values):
+        print(
+            "\t".join(
+                (
+                    input_file_tag(dataset),
+                    f"{sweep_mode}{value}",
+                    str(content),
+                    str(stripped),
+                )
+            )
+        )
 PY
 )"
-mapfile -t sweep_shape <<< "$shape_raw"
-dataset_tag="${sweep_shape[0]}"
-sweep_tag="${sweep_shape[1]}"
-expected_content="${sweep_shape[2]}"
-expected_stripped="${sweep_shape[3]}"
+mapfile -t sweep_shapes <<< "$shape_raw"
 
+nsys_output_prefix_base="${DYN_NSYS_OUTPUT_PREFIX:-vllm}"
 for ((iteration = 1; iteration <= REPETITIONS; iteration++)); do
     iteration_config="$OUTPUT_BASE/config-rep-$iteration.yaml"
     iteration_order="$("$PYTHON_BIN" - "$iteration_config" <<'PY'
@@ -169,21 +168,26 @@ import yaml
 config = yaml.safe_load(pathlib.Path(sys.argv[1]).read_text())
 print(" -> ".join(item["label"] for item in config["configs"]))
 PY
-)"
+    )"
     echo "[sweep] ITERATION_ORDER_${iteration}=${iteration_order}"
+    export DYN_NSYS_OUTPUT_PREFIX="${nsys_output_prefix_base}-rep-${iteration}"
     "$PYTHON_BIN" -m benchmarks.multimodal.sweep \
         --config "$iteration_config" \
         --output-dir "$OUTPUT_BASE/rep-$iteration" \
         --skip-plots
     echo "[sweep] END_ITER_${iteration}"
     if [[ "$iteration" == "1" ]]; then
-        for arm in "${arms[@]}"; do
-            artifact="$OUTPUT_BASE/rep-1/$dataset_tag/$arm/$sweep_tag"
-            "$PYTHON_BIN" -m benchmarks.multimodal.jsonl.validate_uuid_transport \
-                "$artifact/inputs.json" \
-                --expect-content "$expected_content" \
-                --expect-stripped "$expected_stripped" \
-                --output "$artifact/uuid_transport_summary.json"
+        for shape in "${sweep_shapes[@]}"; do
+            IFS=$'\t' read -r dataset_tag sweep_tag expected_content expected_stripped \
+                <<< "$shape"
+            for arm in "${arms[@]}"; do
+                artifact="$OUTPUT_BASE/rep-1/$dataset_tag/$arm/$sweep_tag"
+                "$PYTHON_BIN" -m benchmarks.multimodal.jsonl.validate_uuid_transport \
+                    "$artifact/inputs.json" \
+                    --expect-content "$expected_content" \
+                    --expect-stripped "$expected_stripped" \
+                    --output "$artifact/uuid_transport_summary.json"
+            done
         done
         echo "[sweep] UUID_TRANSPORT_VALIDATED"
     fi
diff --git a/benchmarks/multimodal/sweep/workflows/vllm_serve.sh b/benchmarks/multimodal/sweep/workflows/vllm_serve.sh
index 4d9cf2524b9d..b4234db47597 100755
--- a/benchmarks/multimodal/sweep/workflows/vllm_serve.sh
+++ b/benchmarks/multimodal/sweep/workflows/vllm_serve.sh
@@ -28,12 +28,21 @@ while [[ $# -gt 0 ]]; do
 done
 
 has_gpu_mem_override=0
+has_max_model_len=0
 for ((i = 0; i < ${#EXTRA_ARGS[@]}; i++)); do
     case "${EXTRA_ARGS[$i]}" in
         --max-model-len)
+            if ((i + 1 >= ${#EXTRA_ARGS[@]})); then
+                echo "ERROR: --max-model-len requires a value" >&2
+                exit 2
+            fi
+            has_max_model_len=1
             MAX_MODEL_LEN="${EXTRA_ARGS[$((i + 1))]}"
             ;;
-        --gpu-memory-utilization|--kv-cache-memory-bytes)
+        --max-model-len=*)
+            has_max_model_len=1
+            ;;
+        --gpu-memory-utilization|--gpu-memory-utilization=*|--kv-cache-memory-bytes|--kv-cache-memory-bytes=*)
             has_gpu_mem_override=1
             ;;
     esac
@@ -74,7 +83,7 @@ VLLM_CMD=(
     --port "$HTTP_PORT"
     --enable-log-requests
 )
-if [[ ! " ${EXTRA_ARGS[*]} " =~ [[:space:]]--max-model-len[[:space:]] ]]; then
+if [[ "$has_max_model_len" == "0" ]]; then
     VLLM_CMD+=(--max-model-len "$MAX_MODEL_LEN")
 fi
 if [[ ${#GPU_MEM_ARGV[@]} -gt 0 ]]; then
@@ -98,6 +107,9 @@ if [[ "${DYN_DISABLE_NSYS:-1}" != "1" ]]; then
     NSYS_DIR="${DYN_NSYS_DIR:-/dynamo-tmp/nsys}"
     NSYS_TMPDIR="${DYN_NSYS_TMPDIR:-/dynamo-tmp/nsys-staging}"
     NSYS_PREFIX="${DYN_NSYS_OUTPUT_PREFIX:-vllm}-${DYN_BENCHMARK_ARM:-standalone}"
+    if [[ -n "${DYN_BENCHMARK_SWEEP:-}" ]]; then
+        NSYS_PREFIX="${NSYS_PREFIX}-${DYN_BENCHMARK_SWEEP}"
+    fi
     mkdir -p "$NSYS_DIR" "$NSYS_TMPDIR"
     export TMPDIR="$NSYS_TMPDIR"
 
@@ -119,8 +131,8 @@ server_pid=0
 cleanup() {
     local exit_code="${1:-0}"
     trap - EXIT INT TERM
-    if [[ "$server_pid" -gt 0 ]] && kill -0 "$server_pid" 2>/dev/null; then
-        kill -INT "$server_pid" 2>/dev/null || true
+    if [[ "$server_pid" -gt 0 ]] && kill -0 -- "-$server_pid" 2>/dev/null; then
+        kill -INT -- "-$server_pid" 2>/dev/null || true
         shutdown_grace="${DYN_SERVER_SHUTDOWN_GRACE_SECONDS:-}"
         if [[ -z "$shutdown_grace" ]]; then
             if [[ "${DYN_DISABLE_NSYS:-1}" == "1" ]]; then
@@ -130,11 +142,11 @@ cleanup() {
             fi
         fi
         for _ in $(seq 1 "$shutdown_grace"); do
-            kill -0 "$server_pid" 2>/dev/null || break
+            kill -0 -- "-$server_pid" 2>/dev/null || break
             sleep 1
         done
-        if kill -0 "$server_pid" 2>/dev/null; then
-            kill -KILL "$server_pid" 2>/dev/null || true
+        if kill -0 -- "-$server_pid" 2>/dev/null; then
+            kill -KILL -- "-$server_pid" 2>/dev/null || true
         fi
         wait "$server_pid" 2>/dev/null || true
     fi
@@ -144,9 +156,9 @@ trap 'cleanup 0' INT TERM
 trap 'cleanup $?' EXIT
 
 if [[ ${#LAUNCH_PREFIX[@]} -gt 0 ]]; then
-    "${LAUNCH_PREFIX[@]}" "${VLLM_CMD[@]}" &
+    setsid "${LAUNCH_PREFIX[@]}" "${VLLM_CMD[@]}" &
 else
-    "${VLLM_CMD[@]}" &
+    setsid "${VLLM_CMD[@]}" &
 fi
 server_pid=$!
 wait "$server_pid"
diff --git a/tests/benchmarks/multimodal/sweep/test_dataset_shape.py b/tests/benchmarks/multimodal/sweep/test_dataset_shape.py
new file mode 100644
index 000000000000..3d80d1db6262
--- /dev/null
+++ b/tests/benchmarks/multimodal/sweep/test_dataset_shape.py
@@ -0,0 +1,58 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+import json
+
+import pytest
+
+from benchmarks.multimodal.sweep.dataset_shape import count_uuid_expectations
+
+pytestmark = [pytest.mark.unit, pytest.mark.pre_merge, pytest.mark.gpu_0]
+
+
+def _write_jsonl(tmp_path, rows: list[dict]) -> str:
+    dataset = tmp_path / "dataset.jsonl"
+    dataset.write_text("".join(json.dumps(row) + "\n" for row in rows))
+    return str(dataset)
+
+
+def test_uuid_expectations_are_deduplicated_per_session(tmp_path) -> None:
+    dataset = _write_jsonl(
+        tmp_path,
+        [
+            {"session_id": "a", "image_uuids": ["x", "y"]},
+            {"session_id": "b", "image_uuids": ["x"]},
+            {"session_id": "a", "image_uuids": ["x", "z"]},
+            {"session_id": "b", "image_uuids": ["x", "y"]},
+        ],
+    )
+
+    assert count_uuid_expectations(dataset) == (5, 2)
+
+
+def test_uuid_expectations_respect_conversation_limit(tmp_path) -> None:
+    dataset = _write_jsonl(
+        tmp_path,
+        [
+            {"session_id": "a", "image_uuids": ["a0"]},
+            {"session_id": "b", "image_uuids": ["b0"]},
+            {"session_id": "a", "image_uuids": ["a0"]},
+            {"session_id": "c", "image_uuids": ["c0"]},
+        ],
+    )
+
+    assert count_uuid_expectations(dataset, conversation_num=2) == (2, 1)
+
+
+def test_uuid_expectations_treat_null_and_missing_sessions_as_rows(tmp_path) -> None:
+    dataset = _write_jsonl(
+        tmp_path,
+        [
+            {"session_id": None, "image_uuids": ["x", "x"]},
+            {"image_uuids": ["x"]},
+            {"session_id": 7, "image_uuids": ["z"]},
+            {"session_id": "7", "image_uuids": ["z"]},
+        ],
+    )
+
+    assert count_uuid_expectations(dataset) == (3, 2)
diff --git a/tests/benchmarks/multimodal/sweep/test_server.py b/tests/benchmarks/multimodal/sweep/test_server.py
index 746bc888d6ac..fb4b8fffe63a 100644
--- a/tests/benchmarks/multimodal/sweep/test_server.py
+++ b/tests/benchmarks/multimodal/sweep/test_server.py
@@ -26,6 +26,7 @@ def test_start_resolves_termination_timeout_and_port(
     process = MagicMock()
 
     with (
+        patch.dict("os.environ", {}, clear=True),
         patch(
             "benchmarks.multimodal.sweep.server.subprocess.Popen",
             return_value=process,
@@ -45,8 +46,51 @@ def test_start_rejects_non_positive_termination_timeout(tmp_path) -> None:
     manager = ServerManager()
 
     with pytest.raises(ValueError, match="must be positive"):
+        with patch.dict("os.environ", {}, clear=True):
+            manager.start(
+                str(workflow),
+                "model",
+                env_overrides={"DYN_SERVER_TERMINATE_TIMEOUT": "0"},
+            )
+
+
+def test_start_rejects_timeout_shorter_than_wrapper_grace(tmp_path) -> None:
+    workflow = tmp_path / "workflow.sh"
+    workflow.write_text("#!/bin/bash\n")
+    manager = ServerManager()
+
+    with (
+        patch.dict("os.environ", {}, clear=True),
+        pytest.raises(ValueError, match="must exceed"),
+    ):
         manager.start(
             str(workflow),
             "model",
-            env_overrides={"DYN_SERVER_TERMINATE_TIMEOUT": "0"},
+            env_overrides={
+                "DYN_SERVER_TERMINATE_TIMEOUT": "15",
+                "DYN_SERVER_SHUTDOWN_GRACE_SECONDS": "20",
+            },
         )
+
+
+def test_start_treats_empty_timeout_override_as_unset(tmp_path) -> None:
+    workflow = tmp_path / "workflow.sh"
+    workflow.write_text("#!/bin/bash\n")
+    process = MagicMock()
+
+    with (
+        patch.dict("os.environ", {}, clear=True),
+        patch(
+            "benchmarks.multimodal.sweep.server.subprocess.Popen",
+            return_value=process,
+        ),
+        patch.object(ServerManager, "wait_for_ready"),
+    ):
+        manager = ServerManager()
+        manager.start(
+            str(workflow),
+            "model",
+            env_overrides={"DYN_SERVER_TERMINATE_TIMEOUT": ""},
+        )
+
+    assert manager.terminate_timeout == 15.0
diff --git a/tests/benchmarks/multimodal/sweep/test_vllm_serve_workflow.py b/tests/benchmarks/multimodal/sweep/test_vllm_serve_workflow.py
index e223d1bc88c6..3cf11a29bd6e 100644
--- a/tests/benchmarks/multimodal/sweep/test_vllm_serve_workflow.py
+++ b/tests/benchmarks/multimodal/sweep/test_vllm_serve_workflow.py
@@ -4,8 +4,10 @@
 from __future__ import annotations
 
 import os
+import signal
 import subprocess
 import sys
+import time
 from pathlib import Path
 
 import pytest
@@ -105,3 +107,59 @@ def test_vllm_workflow_requires_model() -> None:
 
     assert result.returncode != 0
     assert "--model is required" in result.stderr
+
+
+@pytest.mark.skipif(sys.platform == "darwin", reason="workflow requires GNU setsid")
+def test_vllm_workflow_kills_server_group_after_grace(tmp_path: Path) -> None:
+    bin_dir = tmp_path / "bin"
+    bin_dir.mkdir()
+    child_pid_file = tmp_path / "child.pid"
+    vllm = bin_dir / "vllm"
+    vllm.write_text(
+        "#!/bin/bash\n"
+        'echo $$ > "$DYN_TEST_CHILD_PID_FILE"\n'
+        "trap '' INT TERM\n"
+        "while true; do sleep 60; done\n"
+    )
+    vllm.chmod(0o755)
+    env = os.environ.copy()
+    env.update(
+        {
+            "DYNAMO_HOME": str(REPO_ROOT),
+            "DYN_DISABLE_NSYS": "1",
+            "DYN_SERVER_SHUTDOWN_GRACE_SECONDS": "1",
+            "DYN_TEST_CHILD_PID_FILE": str(child_pid_file),
+            "PATH": f"{bin_dir}:{env['PATH']}",
+        }
+    )
+    process = subprocess.Popen(
+        ["bash", str(WORKFLOW), "--model", "test-model"],
+        env=env,
+        start_new_session=True,
+        stdout=subprocess.PIPE,
+        stderr=subprocess.PIPE,
+        text=True,
+    )
+    child_pid: int | None = None
+    try:
+        deadline = time.monotonic() + 5
+        while time.monotonic() < deadline and not child_pid_file.exists():
+            time.sleep(0.05)
+        assert child_pid_file.exists()
+        child_pid = int(child_pid_file.read_text())
+
+        os.killpg(process.pid, signal.SIGTERM)
+        process.communicate(timeout=5)
+
+        assert process.returncode == 0
+        with pytest.raises(ProcessLookupError):
+            os.kill(child_pid, 0)
+    finally:
+        if process.poll() is None:
+            os.killpg(process.pid, signal.SIGKILL)
+            process.wait(timeout=5)
+        if child_pid is not None:
+            try:
+                os.killpg(child_pid, signal.SIGKILL)
+            except ProcessLookupError:
+                pass

From 048a4da875304b25824b2c85d7ddc6e1e9cc57af Mon Sep 17 00:00:00 2001
From: furionw <qiwa@nvidia.com>
Date: Wed, 16 Sep 2026 00:17:25 -0700
Subject: [PATCH 3/7] perf(bench): compare baseline and overlapped EC

Signed-off-by: furionw <qiwa@nvidia.com>
---
 benchmarks/multimodal/sweep/README.md         |  8 +++-
 benchmarks/multimodal/sweep/config.py         |  2 +
 .../embedding_cache/vllm_serve.yaml           | 45 +++++++++++++++++--
 benchmarks/multimodal/sweep/orchestrator.py   | 17 +++++++
 .../multimodal/sweep/repetition_plan.py       | 26 +++++++++++
 .../workflows/run_vllm_ec_uuid_repetitions.sh | 32 +++++++++++--
 .../multimodal/sweep/test_orchestrator.py     | 18 +++++++-
 .../multimodal/sweep/test_repetition_plan.py  | 27 ++++++++++-
 .../sweep/test_vllm_serve_workflow.py         |  5 ++-
 9 files changed, 168 insertions(+), 12 deletions(-)

diff --git a/benchmarks/multimodal/sweep/README.md b/benchmarks/multimodal/sweep/README.md
index 666125c9c826..45e39b8304ff 100644
--- a/benchmarks/multimodal/sweep/README.md
+++ b/benchmarks/multimodal/sweep/README.md
@@ -46,7 +46,8 @@ input_files:
   - benchmarks/multimodal/jsonl/1000req_1img_200pool_400word_http.jsonl
   - benchmarks/multimodal/jsonl/1000req_4img_200pool_400word_http.jsonl
 
-# Each config launches the workflow with its own extra_args
+# Each config launches the workflow with its own extra_args and optional env.
+# Per-arm env values expand variables from the harness process.
 configs:
   - label: cache-off
     workflow: benchmarks/multimodal/sweep/workflows/vllm_serve.sh
@@ -54,6 +55,8 @@ configs:
 
   - label: cache-on
     workflow: benchmarks/multimodal/sweep/workflows/vllm_serve.sh
+    env:
+      PYTHONPATH: "${VLLM_PATCHED_PYTHONPATH}"
     extra_args: [--no-enable-prefix-caching, --multimodal-embedding-cache-capacity-gb, "10"]
 ```
 
@@ -69,7 +72,10 @@ configs:
 | `DYN_SERVER_TERMINATE_TIMEOUT` | Orchestrator shutdown timeout; defaults to 300 seconds with profiling and 15 otherwise. |
 | `DYN_SERVER_SHUTDOWN_GRACE_SECONDS` | Wrapper grace period before SIGKILL; defaults to 150 seconds with profiling and 10 otherwise. |
 | `DYN_PYTHON` | Python executable used by the repetition wrapper. |
+| `DYN_BENCHMARK_ORDER_SEED` | Seed for randomized per-repetition arm order; defaults to 42. |
 | `VLLM_SOURCE_REVISION` | Required tested-vLLM revision recorded by the repetition wrapper. |
+| `VLLM_BASELINE_SOURCE_REVISION` | Required baseline vLLM revision for the native-EC comparison. |
+| `VLLM_BASELINE_PYTHONPATH` / `VLLM_PATCHED_PYTHONPATH` | Source trees selected by the baseline and overlap arms. |
 | `CONTAINER_IMAGE` | Required runtime image reference recorded by the repetition wrapper. |
 | `CONTAINER_IMAGE_DIGEST` | Required immutable runtime image digest. |
 | `CONTAINER_IMAGE_FILE` | Required imported image/squashfs path used by the GPU run. |
diff --git a/benchmarks/multimodal/sweep/config.py b/benchmarks/multimodal/sweep/config.py
index 23e5e83edfb8..b231b82fd59c 100644
--- a/benchmarks/multimodal/sweep/config.py
+++ b/benchmarks/multimodal/sweep/config.py
@@ -18,6 +18,7 @@ class BenchmarkConfig:
     label: str
     workflow: str
     extra_args: List[str] = field(default_factory=list)
+    env: Dict[str, str] = field(default_factory=dict)
 
 
 @dataclass
@@ -96,6 +97,7 @@ def _parse_benchmark_config(raw: Dict[str, Any]) -> BenchmarkConfig:
         label=raw["label"],
         workflow=raw["workflow"],
         extra_args=[str(a) for a in raw.get("extra_args", [])],
+        env={str(k): str(v) for k, v in raw.get("env", {}).items()},
     )
 
 
diff --git a/benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml b/benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml
index 2f1fa76c0c33..281adcd639f4 100644
--- a/benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml
+++ b/benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml
@@ -1,10 +1,11 @@
 # SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
-# Qwen3.5-122B two-arm native embedding-cache comparison on 2xH100.
+# Qwen3.5-122B three-arm native embedding-cache comparison on 2xH100.
 # Repeated image payloads carry stable UUIDs and are stripped after their first
-# use in each session. The mm processor cache is deliberately identical across
-# arms; only the post-encoder embedding-cache connector changes.
+# use in each session. The baseline source tree runs both the no-EC control and
+# synchronous native EC; the patched tree runs native EC with same-step H2D /
+# vision-encoder overlap. All other vLLM arguments are identical.
 
 model: Qwen/Qwen3.5-122B-A10B-FP8
 concurrencies: [30]
@@ -29,6 +30,9 @@ input_files:
 configs:
   - label: vllm-serve
     workflow: benchmarks/multimodal/sweep/workflows/vllm_serve.sh
+    env:
+      PYTHONPATH: "${VLLM_BASELINE_PYTHONPATH}"
+      DYN_VLLM_SOURCE_REVISION: "${VLLM_BASELINE_SOURCE_REVISION}"
     extra_args:
       - --tensor-parallel-size
       - "2"
@@ -52,8 +56,41 @@ configs:
       - --no-enable-prefix-caching
       - --no-enable-log-requests
 
-  - label: vllm-serve-native-ec
+  - label: vllm-serve-native-ec-baseline
     workflow: benchmarks/multimodal/sweep/workflows/vllm_serve.sh
+    env:
+      PYTHONPATH: "${VLLM_BASELINE_PYTHONPATH}"
+      DYN_VLLM_SOURCE_REVISION: "${VLLM_BASELINE_SOURCE_REVISION}"
+    extra_args:
+      - --tensor-parallel-size
+      - "2"
+      - --quantization
+      - fp8
+      - --enable-expert-parallel
+      - --gpu-memory-utilization
+      - "0.90"
+      - --max-model-len
+      - "32768"
+      - --max-num-batched-tokens
+      - "32768"
+      - --max-num-seqs
+      - "50"
+      - --mm-encoder-tp-mode
+      - data
+      - --mm-processor-cache-gb
+      - "30"
+      - --limit-mm-per-prompt
+      - '{"image":5}'
+      - --no-enable-prefix-caching
+      - --no-enable-log-requests
+      - --ec-transfer-config
+      - '{"ec_connector":"ECCPUConnector","ec_role":"ec_both","ec_connector_extra_config":{"ec_cpu_bytes":4294967296}}'
+
+  - label: vllm-serve-native-ec-overlap
+    workflow: benchmarks/multimodal/sweep/workflows/vllm_serve.sh
+    env:
+      PYTHONPATH: "${VLLM_PATCHED_PYTHONPATH}"
+      DYN_VLLM_SOURCE_REVISION: "${VLLM_PATCHED_SOURCE_REVISION}"
     extra_args:
       - --tensor-parallel-size
       - "2"
diff --git a/benchmarks/multimodal/sweep/orchestrator.py b/benchmarks/multimodal/sweep/orchestrator.py
index 2490a3bd2f08..0d67b137487e 100644
--- a/benchmarks/multimodal/sweep/orchestrator.py
+++ b/benchmarks/multimodal/sweep/orchestrator.py
@@ -3,6 +3,8 @@
 
 from __future__ import annotations
 
+import os
+import re
 from pathlib import Path
 from typing import List, Optional
 
@@ -45,6 +47,20 @@ def _print_banner(title: str, char: str = "=", width: int = 70) -> None:
     print(f"{char * width}", flush=True)
 
 
+def _expand_arm_env(env: dict[str, str]) -> dict[str, str]:
+    """Expand host variables in per-arm environment values."""
+    expanded = {key: os.path.expandvars(value) for key, value in env.items()}
+    unresolved = {
+        key: value
+        for key, value in expanded.items()
+        if re.search(r"\$(?:[A-Za-z_][A-Za-z0-9_]*|\{[^}]+\})", value)
+    }
+    if unresolved:
+        names = ", ".join(sorted(unresolved))
+        raise ValueError(f"Unresolved variables in benchmark arm env: {names}")
+    return expanded
+
+
 def run_sweep(
     config: SweepConfig,
     repo_root: Optional[Path] = None,
@@ -120,6 +136,7 @@ def _run_config(
     workflow_abs = _resolve_workflow(bench_cfg.workflow, repo_root)
     arm_env_overrides = {
         **env_overrides,
+        **_expand_arm_env(bench_cfg.env),
         "DYN_BENCHMARK_ARM": bench_cfg.label,
         "DYN_BENCHMARK_SWEEP": "multi",
     }
diff --git a/benchmarks/multimodal/sweep/repetition_plan.py b/benchmarks/multimodal/sweep/repetition_plan.py
index bdb6ef77a44a..25ca772cc987 100644
--- a/benchmarks/multimodal/sweep/repetition_plan.py
+++ b/benchmarks/multimodal/sweep/repetition_plan.py
@@ -3,6 +3,8 @@
 
 from __future__ import annotations
 
+import itertools
+import random
 from collections.abc import Sequence
 from typing import Any
 
@@ -25,3 +27,27 @@ def balanced_config_orders(
     ]
     balanced = rotations + [list(reversed(order)) for order in rotations]
     return [balanced[index % len(balanced)] for index in range(repetitions)]
+
+
+def randomized_config_orders(
+    configs: Sequence[dict[str, Any]], repetitions: int, seed: int
+) -> list[list[dict[str, Any]]]:
+    """Return seeded random orders, avoiding repeats until all orders are used."""
+    if not configs:
+        raise ValueError("At least one benchmark config is required")
+    if repetitions < 1:
+        raise ValueError("repetitions must be positive")
+
+    labels = [config["label"] for config in configs]
+    if len(labels) != len(set(labels)):
+        raise ValueError("Benchmark config labels must be unique")
+
+    rng = random.Random(seed)
+    # Sweep configs are intentionally small. Enumerating permutations gives a
+    # reproducible sample without replacement for each complete cycle.
+    permutations = list(itertools.permutations(configs))
+    orders: list[list[dict[str, Any]]] = []
+    while len(orders) < repetitions:
+        rng.shuffle(permutations)
+        orders.extend(list(order) for order in permutations)
+    return orders[:repetitions]
diff --git a/benchmarks/multimodal/sweep/workflows/run_vllm_ec_uuid_repetitions.sh b/benchmarks/multimodal/sweep/workflows/run_vllm_ec_uuid_repetitions.sh
index a3a4d9058207..b912c3dae3d6 100755
--- a/benchmarks/multimodal/sweep/workflows/run_vllm_ec_uuid_repetitions.sh
+++ b/benchmarks/multimodal/sweep/workflows/run_vllm_ec_uuid_repetitions.sh
@@ -6,10 +6,14 @@ set -euo pipefail
 
 CONFIG="${1:-benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml}"
 OUTPUT_BASE="${2:-/dynamo-tmp/logs/09-15/qwen35-122b-vllm-ec-h2d-overlap}"
-REPETITIONS="${3:-4}"
+REPETITIONS="${3:-5}"
 PYTHON_BIN="${DYN_PYTHON:-python}"
+ORDER_SEED="${DYN_BENCHMARK_ORDER_SEED:-42}"
 
 : "${VLLM_SOURCE_REVISION:?VLLM_SOURCE_REVISION must identify the tested vLLM commit}"
+: "${VLLM_BASELINE_SOURCE_REVISION:?VLLM_BASELINE_SOURCE_REVISION must identify the baseline vLLM commit}"
+: "${VLLM_BASELINE_PYTHONPATH:?VLLM_BASELINE_PYTHONPATH must select the baseline vLLM tree}"
+: "${VLLM_PATCHED_PYTHONPATH:?VLLM_PATCHED_PYTHONPATH must select the patched vLLM tree}"
 : "${CONTAINER_IMAGE:?CONTAINER_IMAGE must identify the tested runtime image}"
 : "${CONTAINER_IMAGE_DIGEST:?CONTAINER_IMAGE_DIGEST must identify the tested image digest}"
 : "${CONTAINER_IMAGE_FILE:?CONTAINER_IMAGE_FILE must identify the imported image file}"
@@ -19,9 +23,15 @@ if [[ ! "$REPETITIONS" =~ ^[1-9][0-9]*$ ]]; then
     echo "REPETITIONS must be a positive integer, got: $REPETITIONS" >&2
     exit 2
 fi
+if [[ ! "$ORDER_SEED" =~ ^[0-9]+$ ]]; then
+    echo "DYN_BENCHMARK_ORDER_SEED must be a non-negative integer, got: $ORDER_SEED" >&2
+    exit 2
+fi
+
+export VLLM_PATCHED_SOURCE_REVISION="${VLLM_PATCHED_SOURCE_REVISION:-$VLLM_SOURCE_REVISION}"
 
 mkdir -p "$OUTPUT_BASE"
-"$PYTHON_BIN" - "$OUTPUT_BASE/run_metadata.json" "$CONFIG" "$OUTPUT_BASE" "$REPETITIONS" <<'PY'
+"$PYTHON_BIN" - "$OUTPUT_BASE/run_metadata.json" "$CONFIG" "$OUTPUT_BASE" "$REPETITIONS" "$ORDER_SEED" <<'PY'
 import datetime
 import json
 import os
@@ -35,12 +45,13 @@ import dynamo._core
 import vllm
 import yaml
 
-from benchmarks.multimodal.sweep.repetition_plan import balanced_config_orders
+from benchmarks.multimodal.sweep.repetition_plan import randomized_config_orders
 
 output = pathlib.Path(sys.argv[1])
 config_path = pathlib.Path(sys.argv[2])
 output_base = pathlib.Path(sys.argv[3])
 repetitions = int(sys.argv[4])
+order_seed = int(sys.argv[5])
 config = yaml.safe_load(config_path.read_text())
 configs = config["configs"]
 if not configs:
@@ -51,7 +62,7 @@ sweep_values = concurrencies or config.get("request_rates") or [4, 8, 16, 32, 64
 
 arm_orders = []
 for iteration, ordered_configs in enumerate(
-    balanced_config_orders(configs, repetitions), start=1
+    randomized_config_orders(configs, repetitions, order_seed), start=1
 ):
     labels = [item["label"] for item in ordered_configs]
     arm_orders.append(labels)
@@ -65,6 +76,7 @@ metadata = {
     "model": config["model"],
     "arms": [item["label"] for item in configs],
     "arm_orders": arm_orders,
+    "arm_order_seed": order_seed,
     "tensor_parallel_sizes": {},
     "ec_cpu_capacity_bytes": {},
     "nvtx": config.get("env", {}).get(
@@ -79,6 +91,8 @@ metadata = {
     "vllm_version": vllm.__version__,
     "vllm_executable": shutil.which("vllm"),
     "vllm_source_revision": os.environ["VLLM_SOURCE_REVISION"],
+    "vllm_source_revisions": {},
+    "vllm_pythonpaths": {},
     "python_executable": sys.executable,
     "dynamo_core_file": dynamo._core.__file__,
     "dynamo_version": version("ai-dynamo"),
@@ -95,6 +109,16 @@ metadata = {
 for arm in configs:
     args = arm.get("extra_args", [])
     label = arm["label"]
+    arm_env = {
+        key: os.path.expandvars(str(value))
+        for key, value in arm.get("env", {}).items()
+    }
+    if "DYN_VLLM_SOURCE_REVISION" in arm_env:
+        metadata["vllm_source_revisions"][label] = arm_env[
+            "DYN_VLLM_SOURCE_REVISION"
+        ]
+    if "PYTHONPATH" in arm_env:
+        metadata["vllm_pythonpaths"][label] = arm_env["PYTHONPATH"]
     if "--tensor-parallel-size" in args:
         index = args.index("--tensor-parallel-size")
         metadata["tensor_parallel_sizes"][label] = int(args[index + 1])
diff --git a/tests/benchmarks/multimodal/sweep/test_orchestrator.py b/tests/benchmarks/multimodal/sweep/test_orchestrator.py
index b487b96c38bb..0dc99d96e6fc 100644
--- a/tests/benchmarks/multimodal/sweep/test_orchestrator.py
+++ b/tests/benchmarks/multimodal/sweep/test_orchestrator.py
@@ -10,7 +10,7 @@
 import pytest
 
 from benchmarks.multimodal.sweep.config import BenchmarkConfig, SweepConfig
-from benchmarks.multimodal.sweep.orchestrator import run_sweep
+from benchmarks.multimodal.sweep.orchestrator import _expand_arm_env, run_sweep
 
 pytestmark = [pytest.mark.unit, pytest.mark.pre_merge, pytest.mark.gpu_0]
 
@@ -217,3 +217,19 @@ def test_uuid_and_strip_propagates_to_aiperf(
         ]
         == "cfg-0"
     )
+
+
+def test_per_arm_env_expands_and_overrides_top_level(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setenv("VLLM_TREE", "/workspace/vllm-baseline")
+
+    assert _expand_arm_env({"PYTHONPATH": "${VLLM_TREE}", "DYN_SOURCE": "base"}) == {
+        "PYTHONPATH": "/workspace/vllm-baseline",
+        "DYN_SOURCE": "base",
+    }
+
+
+def test_per_arm_env_rejects_unresolved_variables() -> None:
+    with pytest.raises(ValueError, match="PYTHONPATH"):
+        _expand_arm_env({"PYTHONPATH": "${MISSING_VLLM_TREE}"})
diff --git a/tests/benchmarks/multimodal/sweep/test_repetition_plan.py b/tests/benchmarks/multimodal/sweep/test_repetition_plan.py
index 4679ca528ec5..926a408dfc59 100644
--- a/tests/benchmarks/multimodal/sweep/test_repetition_plan.py
+++ b/tests/benchmarks/multimodal/sweep/test_repetition_plan.py
@@ -5,7 +5,10 @@
 
 import pytest
 
-from benchmarks.multimodal.sweep.repetition_plan import balanced_config_orders
+from benchmarks.multimodal.sweep.repetition_plan import (
+    balanced_config_orders,
+    randomized_config_orders,
+)
 
 pytestmark = [pytest.mark.unit, pytest.mark.pre_merge, pytest.mark.gpu_0]
 
@@ -38,6 +41,26 @@ def test_three_arm_orders_cover_rotations_reverses_and_wrap() -> None:
     assert orders == expected_cycle + expected_cycle[:2]
 
 
+def test_three_arm_five_repetition_order_is_seeded_and_unique() -> None:
+    configs = [{"label": "a"}, {"label": "b"}, {"label": "c"}]
+
+    first = _labels(randomized_config_orders(configs, repetitions=5, seed=42))
+    second = _labels(randomized_config_orders(configs, repetitions=5, seed=42))
+
+    assert first == second
+    assert len({tuple(order) for order in first}) == 5
+    assert all(sorted(order) == ["a", "b", "c"] for order in first)
+
+
+def test_randomized_orders_wrap_after_all_permutations() -> None:
+    configs = [{"label": "a"}, {"label": "b"}]
+
+    orders = _labels(randomized_config_orders(configs, repetitions=5, seed=7))
+
+    assert all(sorted(order) == ["a", "b"] for order in orders)
+    assert len(orders) == 5
+
+
 @pytest.mark.parametrize(
     ("configs", "repetitions", "error"),
     [
@@ -51,3 +74,5 @@ def test_invalid_repetition_plan_is_rejected(
 ) -> None:
     with pytest.raises(ValueError, match=error):
         balanced_config_orders(configs, repetitions)
+    with pytest.raises(ValueError, match=error):
+        randomized_config_orders(configs, repetitions, seed=42)
diff --git a/tests/benchmarks/multimodal/sweep/test_vllm_serve_workflow.py b/tests/benchmarks/multimodal/sweep/test_vllm_serve_workflow.py
index 3cf11a29bd6e..97341f0cb72c 100644
--- a/tests/benchmarks/multimodal/sweep/test_vllm_serve_workflow.py
+++ b/tests/benchmarks/multimodal/sweep/test_vllm_serve_workflow.py
@@ -27,9 +27,12 @@ def test_embedding_cache_sweep_selects_only_requested_arms() -> None:
 
     assert [item["label"] for item in config["configs"]] == [
         "vllm-serve",
-        "vllm-serve-native-ec",
+        "vllm-serve-native-ec-baseline",
+        "vllm-serve-native-ec-overlap",
     ]
     assert config["env"]["DYN_DISABLE_NSYS"] == "1"
+    assert config["configs"][0]["env"]["PYTHONPATH"] == ("${VLLM_BASELINE_PYTHONPATH}")
+    assert config["configs"][2]["env"]["PYTHONPATH"] == ("${VLLM_PATCHED_PYTHONPATH}")
 
 
 def _run_workflow(tmp_path: Path, *args: str, enable_nsys: bool = False):

From ebf79e62e2fdd6db36305654add400ae1d9491d7 Mon Sep 17 00:00:00 2001
From: furionw <qiwa@nvidia.com>
Date: Wed, 16 Sep 2026 14:20:07 -0700
Subject: [PATCH 4/7] perf(bench): record GPU placement metadata

---
 .../embedding_cache/vllm_serve.yaml           |  2 +-
 .../workflows/run_vllm_ec_uuid_repetitions.sh | 26 +++++++++++++++++++
 2 files changed, 27 insertions(+), 1 deletion(-)

diff --git a/benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml b/benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml
index 281adcd639f4..d3de4729aa79 100644
--- a/benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml
+++ b/benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml
@@ -1,7 +1,7 @@
 # SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
-# Qwen3.5-122B three-arm native embedding-cache comparison on 2xH100.
+# Qwen3.5-122B three-arm native embedding-cache comparison on two GPUs.
 # Repeated image payloads carry stable UUIDs and are stripped after their first
 # use in each session. The baseline source tree runs both the no-EC control and
 # synchronous native EC; the patched tree runs native EC with same-step H2D /
diff --git a/benchmarks/multimodal/sweep/workflows/run_vllm_ec_uuid_repetitions.sh b/benchmarks/multimodal/sweep/workflows/run_vllm_ec_uuid_repetitions.sh
index b912c3dae3d6..1729cb87e7f7 100755
--- a/benchmarks/multimodal/sweep/workflows/run_vllm_ec_uuid_repetitions.sh
+++ b/benchmarks/multimodal/sweep/workflows/run_vllm_ec_uuid_repetitions.sh
@@ -36,6 +36,7 @@ import datetime
 import json
 import os
 import pathlib
+import platform
 import shutil
 import subprocess
 import sys
@@ -60,6 +61,23 @@ concurrencies = config.get("concurrencies")
 sweep_mode = "concurrency" if concurrencies else "request_rate"
 sweep_values = concurrencies or config.get("request_rates") or [4, 8, 16, 32, 64]
 
+
+def command_output(args: list[str]) -> str | None:
+    try:
+        return subprocess.check_output(args, text=True).strip()
+    except (OSError, subprocess.CalledProcessError):
+        return None
+
+
+gpu_inventory = command_output(
+    [
+        "nvidia-smi",
+        "--query-gpu=index,uuid,pci.bus_id,name,memory.total",
+        "--format=csv,noheader,nounits",
+    ]
+)
+gpu_topology = command_output(["nvidia-smi", "topo", "-m"])
+
 arm_orders = []
 for iteration, ordered_configs in enumerate(
     randomized_config_orders(configs, repetitions, order_seed), start=1
@@ -101,6 +119,13 @@ metadata = {
     "container_image_digest": os.environ["CONTAINER_IMAGE_DIGEST"],
     "container_image_file": os.environ["CONTAINER_IMAGE_FILE"],
     "harness_revision": os.environ["HARNESS_REVISION"],
+    "node": os.environ.get(
+        "SLURMD_NODENAME", os.environ.get("SLURM_NODELIST")
+    ),
+    "platform_machine": platform.machine(),
+    "cuda_visible_devices": os.environ.get("CUDA_VISIBLE_DEVICES"),
+    "gpu_inventory": gpu_inventory.splitlines() if gpu_inventory else [],
+    "gpu_topology": gpu_topology,
     "datasets": config["input_files"],
     "sweep_mode": sweep_mode,
     "sweep_values": sweep_values,
@@ -194,6 +219,7 @@ print(" -> ".join(item["label"] for item in config["configs"]))
 PY
     )"
     echo "[sweep] ITERATION_ORDER_${iteration}=${iteration_order}"
+    echo "[sweep] CUDA_VISIBLE_DEVICES_${iteration}=${CUDA_VISIBLE_DEVICES:-<unset>}"
     export DYN_NSYS_OUTPUT_PREFIX="${nsys_output_prefix_base}-rep-${iteration}"
     "$PYTHON_BIN" -m benchmarks.multimodal.sweep \
         --config "$iteration_config" \

From 5c5a32dcf2e88fd9a4ef0706cb994ff79d59178d Mon Sep 17 00:00:00 2001
From: furionw <qiwa@nvidia.com>
Date: Wed, 16 Sep 2026 23:06:38 -0700
Subject: [PATCH 5/7] perf(benchmarks): add shared-prefix EC overlap sweep

---
 .../multimodal/jsonl/generate_images.py       |   3 +-
 .../jsonl/prepare_ec_h2d_overlap.py           | 265 ++++++++++++++++++
 benchmarks/multimodal/sweep/config.py         |  16 ++
 .../embedding_cache/vllm_serve.yaml           | 101 ++++---
 benchmarks/multimodal/sweep/orchestrator.py   |  20 ++
 benchmarks/multimodal/sweep/server.py         |  64 ++++-
 .../workflows/run_vllm_ec_uuid_repetitions.sh |  96 ++++++-
 .../multimodal/sweep/workflows/vllm_serve.sh  |  14 +
 .../benchmarks/multimodal/jsonl/test_main.py  |  11 +
 .../jsonl/test_prepare_ec_h2d_overlap.py      |  75 +++++
 .../multimodal/sweep/test_orchestrator.py     |  27 ++
 .../multimodal/sweep/test_server.py           |  75 +++++
 .../sweep/test_vllm_serve_workflow.py         |  13 +-
 13 files changed, 728 insertions(+), 52 deletions(-)
 create mode 100755 benchmarks/multimodal/jsonl/prepare_ec_h2d_overlap.py
 create mode 100644 tests/benchmarks/multimodal/jsonl/test_prepare_ec_h2d_overlap.py

diff --git a/benchmarks/multimodal/jsonl/generate_images.py b/benchmarks/multimodal/jsonl/generate_images.py
index 6c63fdc171d6..056014f2bb4a 100644
--- a/benchmarks/multimodal/jsonl/generate_images.py
+++ b/benchmarks/multimodal/jsonl/generate_images.py
@@ -36,10 +36,11 @@ def generate_image_pool_base64(
 ) -> list[str]:
     """Generate pool_size random PNG files and return their paths."""
     image_dir.mkdir(parents=True, exist_ok=True)
+    width, height = image_size
     pool: list[str] = []
     for idx in range(pool_size):
         path = image_dir / f"img_{idx:04d}.png"
-        pixels = np_rng.integers(0, 256, (*image_size, 3), dtype=np.uint8)
+        pixels = np_rng.integers(0, 256, (height, width, 3), dtype=np.uint8)
         Image.fromarray(pixels).save(path)
         pool.append(str(path.resolve()))
     print(
diff --git a/benchmarks/multimodal/jsonl/prepare_ec_h2d_overlap.py b/benchmarks/multimodal/jsonl/prepare_ec_h2d_overlap.py
new file mode 100755
index 000000000000..413868a71f9a
--- /dev/null
+++ b/benchmarks/multimodal/jsonl/prepare_ec_h2d_overlap.py
@@ -0,0 +1,265 @@
+#!/usr/bin/env python3
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Prepare deterministic workloads for the native CPU EC overlap benchmark."""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+from collections.abc import Sequence
+from pathlib import Path
+from typing import Any, Protocol
+
+import numpy as np
+
+try:
+    from transformers import AutoTokenizer
+except ModuleNotFoundError:  # Optional for tokenizer-independent unit tests.
+    AutoTokenizer = None
+
+from benchmarks.multimodal.jsonl.generate_images import (
+    compute_image_uuid,
+    generate_image_pool_base64,
+)
+
+DEFAULT_MODEL = "Qwen/Qwen3.5-122B-A10B-FP8"
+DEFAULT_OUTPUT_DIR = Path("/dynamo-tmp/data")
+DEFAULT_IMAGE_DIR = DEFAULT_OUTPUT_DIR / "ec_h2d_overlap_images_2400x1080_seed42"
+SYSTEM_PROMPT_TOKENS = 8000
+USER_TEXT = "Describe the newest image and summarize only its visible content."
+SYSTEM_CONTEXT = (
+    " The attached images may contain objects, text, diagrams, tables, labels,"
+    " quantities, and annotations. Consider visual details in context, distinguish"
+    " similar elements carefully, and report only information supported by the images."
+)
+
+
+class Tokenizer(Protocol):
+    chat_template: str | dict[str, str] | None
+    init_kwargs: dict[str, Any]
+
+    def encode(self, text: str, *, add_special_tokens: bool = False) -> list[int]:
+        ...
+
+    def decode(
+        self,
+        ids: Sequence[int],
+        *,
+        skip_special_tokens: bool = True,
+        clean_up_tokenization_spaces: bool = False,
+    ) -> str:
+        ...
+
+    def apply_chat_template(self, conversation: list[dict[str, str]], **kwargs: Any):
+        ...
+
+
+def _sha256(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as source:
+        for chunk in iter(lambda: source.read(1024 * 1024), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def count_tokens(tokenizer: Tokenizer, text: str) -> int:
+    return len(tokenizer.encode(text, add_special_tokens=False))
+
+
+def make_exact_text(tokenizer: Tokenizer, target_tokens: int) -> str:
+    """Return deterministic text with exactly ``target_tokens`` tokenizer tokens."""
+    if target_tokens <= 0:
+        raise ValueError("target_tokens must be positive")
+
+    unit_tokens = count_tokens(tokenizer, SYSTEM_CONTEXT)
+    repeats = max(2, target_tokens // unit_tokens + 2)
+    keep = target_tokens
+    for _ in range(64):
+        ids = tokenizer.encode(SYSTEM_CONTEXT * repeats, add_special_tokens=False)
+        while len(ids) < keep:
+            repeats *= 2
+            ids = tokenizer.encode(SYSTEM_CONTEXT * repeats, add_special_tokens=False)
+        text = tokenizer.decode(
+            ids[:keep],
+            skip_special_tokens=True,
+            clean_up_tokenization_spaces=False,
+        )
+        observed = count_tokens(tokenizer, text)
+        if observed == target_tokens:
+            return text
+        keep += target_tokens - observed
+    raise RuntimeError(f"could not construct exact {target_tokens}-token text")
+
+
+def build_chat_template(base_template: str, system_prompt: str) -> str:
+    """Prepend the benchmark system message without changing client payloads."""
+    prompt_literal = json.dumps(system_prompt, ensure_ascii=False)
+    prefix = (
+        "{%- if not messages or messages[0]['role'] != 'system' -%}"
+        "{%- set messages = [{'role': 'system', 'content': "
+        f"{prompt_literal}"
+        "}] + messages -%}"
+        "{%- endif -%}\n"
+    )
+    return prefix + base_template
+
+
+def write_sliding_dataset(
+    path: Path,
+    image_pool: Sequence[str],
+    *,
+    num_users: int,
+    turns_per_user: int,
+    window_size: int,
+    images_per_user: int,
+    user_text: str,
+) -> dict[str, int]:
+    """Write one turn-major sliding-window dataset from a shared image pool."""
+    required_images = num_users * images_per_user
+    if len(image_pool) < required_images:
+        raise ValueError(
+            f"image pool has {len(image_pool)} entries; {required_images} required"
+        )
+    if window_size + turns_per_user - 1 > images_per_user:
+        raise ValueError("images_per_user cannot cover every sliding window")
+
+    path.parent.mkdir(parents=True, exist_ok=True)
+    unique_refs: set[str] = set()
+    rows = 0
+    with path.open("w", encoding="utf-8") as output:
+        for turn_idx in range(turns_per_user):
+            for user_idx in range(num_users):
+                offset = user_idx * images_per_user + turn_idx
+                images = list(image_pool[offset : offset + window_size])
+                unique_refs.update(images)
+                row = {
+                    "session_id": f"user_{user_idx}",
+                    "text": user_text,
+                    "images": images,
+                    "image_uuids": [compute_image_uuid(ref) for ref in images],
+                }
+                output.write(json.dumps(row, separators=(",", ":")) + "\n")
+                rows += 1
+
+    total_slots = rows * window_size
+    return {
+        "rows": rows,
+        "unique_images": len(unique_refs),
+        "content_images": len(unique_refs),
+        "stripped_images": total_slots - len(unique_refs),
+    }
+
+
+def _parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--model", default=DEFAULT_MODEL)
+    parser.add_argument("--output-dir", type=Path, default=DEFAULT_OUTPUT_DIR)
+    parser.add_argument("--image-dir", type=Path, default=DEFAULT_IMAGE_DIR)
+    parser.add_argument("--seed", type=int, default=42)
+    parser.add_argument("--force", action="store_true")
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = _parse_args()
+    args.output_dir.mkdir(parents=True, exist_ok=True)
+
+    prompt_path = args.output_dir / "qwen35_shared_system_8000.txt"
+    template_path = args.output_dir / "qwen35_shared_system_8000.jinja"
+    manifest_path = args.output_dir / "qwen35_ec_h2d_overlap_manifest.json"
+    dataset_paths = {
+        5: args.output_dir / "30u_8t_5w_shared8k_base64_uuid_seed42.jsonl",
+        10: args.output_dir / "30u_8t_10w_shared8k_base64_uuid_seed42.jsonl",
+    }
+    artifacts = [prompt_path, template_path, manifest_path, *dataset_paths.values()]
+    if not args.force and all(path.is_file() for path in artifacts):
+        print(f"All workload artifacts already exist; keeping {manifest_path}")
+        return
+
+    if AutoTokenizer is None:
+        raise RuntimeError("transformers is required to prepare EC workloads")
+    tokenizer = AutoTokenizer.from_pretrained(
+        args.model,
+        trust_remote_code=True,
+        local_files_only=True,
+    )
+    base_template = tokenizer.chat_template
+    if not isinstance(base_template, str):
+        raise TypeError("target tokenizer must expose one string chat_template")
+
+    system_prompt = make_exact_text(tokenizer, SYSTEM_PROMPT_TOKENS)
+    prompt_path.write_text(system_prompt, encoding="utf-8")
+    template = build_chat_template(base_template, system_prompt)
+    template_path.write_text(template, encoding="utf-8")
+
+    rendered_ids = tokenizer.apply_chat_template(
+        [{"role": "user", "content": USER_TEXT}],
+        chat_template=template,
+        tokenize=True,
+        add_generation_prompt=True,
+    )
+
+    num_users = 30
+    turns_per_user = 8
+    max_window_size = max(dataset_paths)
+    images_per_user = max_window_size + turns_per_user - 1
+    image_pool = generate_image_pool_base64(
+        np.random.default_rng(args.seed),
+        num_users * images_per_user,
+        args.image_dir,
+        (2400, 1080),
+    )
+
+    datasets: dict[str, dict[str, Any]] = {}
+    for window_size, path in dataset_paths.items():
+        counts = write_sliding_dataset(
+            path,
+            image_pool,
+            num_users=num_users,
+            turns_per_user=turns_per_user,
+            window_size=window_size,
+            images_per_user=images_per_user,
+            user_text=USER_TEXT,
+        )
+        datasets[str(window_size)] = {
+            "path": str(path),
+            "sha256": _sha256(path),
+            "window_size": window_size,
+            **counts,
+        }
+
+    manifest = {
+        "model": args.model,
+        "tokenizer_commit": tokenizer.init_kwargs.get("_commit_hash"),
+        "seed": args.seed,
+        "system_prompt": {
+            "path": str(prompt_path),
+            "tokens": count_tokens(tokenizer, system_prompt),
+            "sha256": _sha256(prompt_path),
+        },
+        "chat_template": {
+            "path": str(template_path),
+            "sha256": _sha256(template_path),
+        },
+        "user_text": USER_TEXT,
+        "user_text_tokens": count_tokens(tokenizer, USER_TEXT),
+        "rendered_text_prompt_tokens": len(rendered_ids),
+        "images": {
+            "directory": str(args.image_dir),
+            "count": len(image_pool),
+            "width": 2400,
+            "height": 1080,
+        },
+        "datasets": datasets,
+    }
+    manifest_path.write_text(
+        json.dumps(manifest, indent=2, sort_keys=True) + "\n", encoding="utf-8"
+    )
+    print(json.dumps(manifest, indent=2, sort_keys=True))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/benchmarks/multimodal/sweep/config.py b/benchmarks/multimodal/sweep/config.py
index b231b82fd59c..0e983281278c 100644
--- a/benchmarks/multimodal/sweep/config.py
+++ b/benchmarks/multimodal/sweep/config.py
@@ -43,6 +43,8 @@ class SweepConfig:
     skip_plots: bool = False
     restart_server_every_benchmark: bool = True
     uuid_and_strip: bool = False
+    prompt_manifest: Optional[str] = None
+    prefix_cache_probe_min_cached_tokens: Optional[int] = None
     env: Dict[str, str] = field(default_factory=dict)
 
     @property
@@ -70,6 +72,16 @@ def validate(self, repo_root: Optional[Path] = None) -> None:
             if not Path(f).is_file():
                 raise FileNotFoundError(f"Input file not found: {f}")
 
+        if self.prompt_manifest and not Path(self.prompt_manifest).is_file():
+            raise FileNotFoundError(
+                f"Prompt manifest not found: {self.prompt_manifest}"
+            )
+        if (
+            self.prefix_cache_probe_min_cached_tokens is not None
+            and self.prefix_cache_probe_min_cached_tokens <= 0
+        ):
+            raise ValueError("prefix_cache_probe_min_cached_tokens must be positive")
+
         for cfg in self.configs:
             script = Path(cfg.workflow)
             if repo_root and not script.is_absolute():
@@ -140,6 +152,10 @@ def load_config(
         skip_plots=raw.get("skip_plots", False),
         restart_server_every_benchmark=raw.get("restart_server_every_benchmark", True),
         uuid_and_strip=raw.get("uuid_and_strip", False),
+        prompt_manifest=raw.get("prompt_manifest"),
+        prefix_cache_probe_min_cached_tokens=raw.get(
+            "prefix_cache_probe_min_cached_tokens"
+        ),
         env=raw.get("env", {}),
     )
 
diff --git a/benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml b/benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml
index d3de4729aa79..e49622da2371 100644
--- a/benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml
+++ b/benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml
@@ -1,11 +1,11 @@
 # SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
-# Qwen3.5-122B three-arm native embedding-cache comparison on two GPUs.
-# Repeated image payloads carry stable UUIDs and are stripped after their first
-# use in each session. The baseline source tree runs both the no-EC control and
-# synchronous native EC; the patched tree runs native EC with same-step H2D /
-# vision-encoder overlap. All other vLLM arguments are identical.
+# Qwen3.5-122B native CPU embedding-cache comparison on two GPUs. A custom
+# server-side chat template injects the same 8000-token system prompt into
+# every request while AIPerf keeps its stock input-file and UUID-strip paths.
+# The only measured difference is the vLLM source tree: synchronous native EC
+# baseline versus same-step H2D / vision-encoder overlap.
 
 model: Qwen/Qwen3.5-122B-A10B-FP8
 concurrencies: [30]
@@ -14,10 +14,12 @@ conversation_num: 30
 warmup_count: 2
 port: 8000
 timeout: 2400
-output_dir: /dynamo-tmp/logs/09-15/qwen35-122b-vllm-ec-h2d-overlap/default
+output_dir: /dynamo-tmp/logs/09-16/qwen35-122b-vllm-ec-h2d-overlap-shared8k/default
 skip_plots: true
 restart_server_every_benchmark: true
 uuid_and_strip: true
+prompt_manifest: /dynamo-tmp/data/qwen35_ec_h2d_overlap_manifest.json
+prefix_cache_probe_min_cached_tokens: 7936
 
 env:
   DYN_DISABLE_NSYS: "1"
@@ -25,42 +27,50 @@ env:
   VLLM_USE_V2_MODEL_RUNNER: "1"
 
 input_files:
-  - /dynamo-tmp/data/30u_8t_5w_8000word_base64_uuid_seed42.jsonl
+  - /dynamo-tmp/data/30u_8t_5w_shared8k_base64_uuid_seed42.jsonl
+  - /dynamo-tmp/data/30u_8t_10w_shared8k_base64_uuid_seed42.jsonl
 
 configs:
-  - label: vllm-serve
-    workflow: benchmarks/multimodal/sweep/workflows/vllm_serve.sh
-    env:
-      PYTHONPATH: "${VLLM_BASELINE_PYTHONPATH}"
-      DYN_VLLM_SOURCE_REVISION: "${VLLM_BASELINE_SOURCE_REVISION}"
-    extra_args:
-      - --tensor-parallel-size
-      - "2"
-      - --quantization
-      - fp8
-      - --enable-expert-parallel
-      - --gpu-memory-utilization
-      - "0.90"
-      - --max-model-len
-      - "32768"
-      - --max-num-batched-tokens
-      - "32768"
-      - --max-num-seqs
-      - "50"
-      - --mm-encoder-tp-mode
-      - data
-      - --mm-processor-cache-gb
-      - "30"
-      - --limit-mm-per-prompt
-      - '{"image":5}'
-      - --no-enable-prefix-caching
-      - --no-enable-log-requests
+  # The no-EC control is intentionally parked: this experiment isolates the
+  # scheduling difference between two otherwise identical native EC servers.
+  # - label: vllm-serve
+  #   workflow: benchmarks/multimodal/sweep/workflows/vllm_serve.sh
+  #   env:
+  #     PYTHONPATH: "${VLLM_BASELINE_PYTHONPATH}"
+  #     DYN_VLLM_SOURCE_REVISION: "${VLLM_BASELINE_SOURCE_REVISION}"
+  #     DYN_VLLM_EXPECTED_ROOT: "${VLLM_BASELINE_PYTHONPATH}"
+  #   extra_args:
+  #     - --tensor-parallel-size
+  #     - "2"
+  #     - --quantization
+  #     - fp8
+  #     - --enable-expert-parallel
+  #     - --gpu-memory-utilization
+  #     - "0.90"
+  #     - --max-model-len
+  #     - "65536"
+  #     - --max-num-batched-tokens
+  #     - "32768"
+  #     - --max-num-seqs
+  #     - "50"
+  #     - --mm-encoder-tp-mode
+  #     - data
+  #     - --mm-processor-cache-gb
+  #     - "30"
+  #     - --limit-mm-per-prompt
+  #     - '{"image":10}'
+  #     - --chat-template
+  #     - /dynamo-tmp/data/qwen35_shared_system_8000.jinja
+  #     - --enable-prefix-caching
+  #     - --enable-prompt-tokens-details
+  #     - --no-enable-log-requests
 
   - label: vllm-serve-native-ec-baseline
     workflow: benchmarks/multimodal/sweep/workflows/vllm_serve.sh
     env:
       PYTHONPATH: "${VLLM_BASELINE_PYTHONPATH}"
       DYN_VLLM_SOURCE_REVISION: "${VLLM_BASELINE_SOURCE_REVISION}"
+      DYN_VLLM_EXPECTED_ROOT: "${VLLM_BASELINE_PYTHONPATH}"
     extra_args:
       - --tensor-parallel-size
       - "2"
@@ -70,7 +80,7 @@ configs:
       - --gpu-memory-utilization
       - "0.90"
       - --max-model-len
-      - "32768"
+      - "65536"
       - --max-num-batched-tokens
       - "32768"
       - --max-num-seqs
@@ -80,17 +90,21 @@ configs:
       - --mm-processor-cache-gb
       - "30"
       - --limit-mm-per-prompt
-      - '{"image":5}'
-      - --no-enable-prefix-caching
+      - '{"image":10}'
+      - --chat-template
+      - /dynamo-tmp/data/qwen35_shared_system_8000.jinja
+      - --enable-prefix-caching
+      - --enable-prompt-tokens-details
       - --no-enable-log-requests
       - --ec-transfer-config
-      - '{"ec_connector":"ECCPUConnector","ec_role":"ec_both","ec_connector_extra_config":{"ec_cpu_bytes":4294967296}}'
+      - '{"ec_connector":"ECCPUConnector","ec_role":"ec_both","ec_connector_extra_config":{"ec_cpu_bytes":8589934592}}'
 
   - label: vllm-serve-native-ec-overlap
     workflow: benchmarks/multimodal/sweep/workflows/vllm_serve.sh
     env:
       PYTHONPATH: "${VLLM_PATCHED_PYTHONPATH}"
       DYN_VLLM_SOURCE_REVISION: "${VLLM_PATCHED_SOURCE_REVISION}"
+      DYN_VLLM_EXPECTED_ROOT: "${VLLM_PATCHED_PYTHONPATH}"
     extra_args:
       - --tensor-parallel-size
       - "2"
@@ -100,7 +114,7 @@ configs:
       - --gpu-memory-utilization
       - "0.90"
       - --max-model-len
-      - "32768"
+      - "65536"
       - --max-num-batched-tokens
       - "32768"
       - --max-num-seqs
@@ -110,11 +124,14 @@ configs:
       - --mm-processor-cache-gb
       - "30"
       - --limit-mm-per-prompt
-      - '{"image":5}'
-      - --no-enable-prefix-caching
+      - '{"image":10}'
+      - --chat-template
+      - /dynamo-tmp/data/qwen35_shared_system_8000.jinja
+      - --enable-prefix-caching
+      - --enable-prompt-tokens-details
       - --no-enable-log-requests
       - --ec-transfer-config
-      - '{"ec_connector":"ECCPUConnector","ec_role":"ec_both","ec_connector_extra_config":{"ec_cpu_bytes":4294967296}}'
+      - '{"ec_connector":"ECCPUConnector","ec_role":"ec_both","ec_connector_extra_config":{"ec_cpu_bytes":8589934592}}'
 
   # Dynamo EC is parked while this experiment isolates native vLLM EC.
   # - label: vllm-serve-dynamo-ec
diff --git a/benchmarks/multimodal/sweep/orchestrator.py b/benchmarks/multimodal/sweep/orchestrator.py
index 0d67b137487e..8a1dccaecac0 100644
--- a/benchmarks/multimodal/sweep/orchestrator.py
+++ b/benchmarks/multimodal/sweep/orchestrator.py
@@ -3,6 +3,7 @@
 
 from __future__ import annotations
 
+import json
 import os
 import re
 from pathlib import Path
@@ -47,6 +48,18 @@ def _print_banner(title: str, char: str = "=", width: int = 70) -> None:
     print(f"{char * width}", flush=True)
 
 
+def _first_user_text(input_file: str) -> str:
+    with open(input_file, encoding="utf-8") as source:
+        for line in source:
+            if not line.strip():
+                continue
+            value = json.loads(line).get("text")
+            if not isinstance(value, str):
+                raise ValueError(f"First row in {input_file} has no string text field")
+            return value
+    raise ValueError(f"Dataset is empty: {input_file}")
+
+
 def _expand_arm_env(env: dict[str, str]) -> dict[str, str]:
     """Expand host variables in per-arm environment values."""
     expanded = {key: os.path.expandvars(value) for key, value in env.items()}
@@ -195,6 +208,13 @@ def _run_config(
                 )
 
             try:
+                if config.prefix_cache_probe_min_cached_tokens is not None:
+                    server.validate_prefix_cache(
+                        model=config.model,
+                        user_text=_first_user_text(input_file),
+                        min_cached_tokens=(config.prefix_cache_probe_min_cached_tokens),
+                        output_path=artifact_dir / "prefix_cache_probe.json",
+                    )
                 run_aiperf_single(
                     model=config.model,
                     port=config.port,
diff --git a/benchmarks/multimodal/sweep/server.py b/benchmarks/multimodal/sweep/server.py
index 24390d9b104c..cc7bc0d78db3 100644
--- a/benchmarks/multimodal/sweep/server.py
+++ b/benchmarks/multimodal/sweep/server.py
@@ -3,12 +3,15 @@
 
 from __future__ import annotations
 
+import json
 import os
 import signal
 import subprocess
 import time
+import urllib.error
+import urllib.request
 from pathlib import Path
-from typing import List, Optional
+from typing import Any, List, Optional
 
 
 class ServerManager:
@@ -82,9 +85,6 @@ def start(
 
     def wait_for_ready(self, model: str) -> None:
         """Poll /v1/models until the expected model name appears."""
-        import urllib.error
-        import urllib.request
-
         url = f"http://localhost:{self.port}/v1/models"
         deadline = time.monotonic() + self.timeout
 
@@ -114,6 +114,62 @@ def wait_for_ready(self, model: str) -> None:
         self.stop()
         raise TimeoutError(f"Server did not become ready within {self.timeout}s")
 
+    def validate_prefix_cache(
+        self,
+        model: str,
+        user_text: str,
+        min_cached_tokens: int,
+        output_path: Path,
+    ) -> dict[str, Any]:
+        """Warm and verify the shared text prefix through chat completions."""
+        if min_cached_tokens <= 0:
+            raise ValueError("min_cached_tokens must be positive")
+
+        url = f"http://localhost:{self.port}/v1/chat/completions"
+        payload = {
+            "model": model,
+            "messages": [{"role": "user", "content": user_text}],
+            "max_tokens": 1,
+            "temperature": 0,
+            "stream": False,
+        }
+
+        def send() -> dict[str, Any]:
+            request = urllib.request.Request(
+                url,
+                data=json.dumps(payload).encode(),
+                headers={"Content-Type": "application/json"},
+            )
+            try:
+                with urllib.request.urlopen(request, timeout=120) as response:
+                    return json.loads(response.read())
+            except urllib.error.HTTPError as exc:
+                body = exc.read().decode(errors="replace")
+                raise RuntimeError(
+                    f"prefix-cache probe failed with HTTP {exc.code}: {body}"
+                ) from exc
+
+        responses = [send(), send()]
+        usages = [response.get("usage", {}) for response in responses]
+        cached_tokens = (usages[1].get("prompt_tokens_details") or {}).get(
+            "cached_tokens", 0
+        )
+        summary = {
+            "minimum_cached_tokens": min_cached_tokens,
+            "first_usage": usages[0],
+            "second_usage": usages[1],
+            "passed": cached_tokens >= min_cached_tokens,
+        }
+        output_path.parent.mkdir(parents=True, exist_ok=True)
+        output_path.write_text(json.dumps(summary, indent=2, sort_keys=True) + "\n")
+        if not summary["passed"]:
+            raise RuntimeError(
+                "prefix-cache probe cached "
+                f"{cached_tokens} tokens; expected at least {min_cached_tokens}"
+            )
+        print(f"Prefix-cache probe passed: cached_tokens={cached_tokens}", flush=True)
+        return summary
+
     def stop(self) -> None:
         """Stop the server by killing its process group."""
         if self._process is None:
diff --git a/benchmarks/multimodal/sweep/workflows/run_vllm_ec_uuid_repetitions.sh b/benchmarks/multimodal/sweep/workflows/run_vllm_ec_uuid_repetitions.sh
index 1729cb87e7f7..0c4a2d306beb 100755
--- a/benchmarks/multimodal/sweep/workflows/run_vllm_ec_uuid_repetitions.sh
+++ b/benchmarks/multimodal/sweep/workflows/run_vllm_ec_uuid_repetitions.sh
@@ -5,8 +5,10 @@
 set -euo pipefail
 
 CONFIG="${1:-benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml}"
-OUTPUT_BASE="${2:-/dynamo-tmp/logs/09-15/qwen35-122b-vllm-ec-h2d-overlap}"
+OUTPUT_BASE="${2:-/dynamo-tmp/logs/09-16/qwen35-122b-vllm-ec-h2d-overlap-shared8k}"
 REPETITIONS="${3:-5}"
+DATASET_FILTER="${4:-}"
+PROFILE="${5:-0}"
 PYTHON_BIN="${DYN_PYTHON:-python}"
 ORDER_SEED="${DYN_BENCHMARK_ORDER_SEED:-42}"
 
@@ -27,12 +29,17 @@ if [[ ! "$ORDER_SEED" =~ ^[0-9]+$ ]]; then
     echo "DYN_BENCHMARK_ORDER_SEED must be a non-negative integer, got: $ORDER_SEED" >&2
     exit 2
 fi
+if [[ "$PROFILE" != "0" && "$PROFILE" != "1" ]]; then
+    echo "PROFILE must be 0 or 1, got: $PROFILE" >&2
+    exit 2
+fi
 
 export VLLM_PATCHED_SOURCE_REVISION="${VLLM_PATCHED_SOURCE_REVISION:-$VLLM_SOURCE_REVISION}"
 
 mkdir -p "$OUTPUT_BASE"
-"$PYTHON_BIN" - "$OUTPUT_BASE/run_metadata.json" "$CONFIG" "$OUTPUT_BASE" "$REPETITIONS" "$ORDER_SEED" <<'PY'
+"$PYTHON_BIN" - "$OUTPUT_BASE/run_metadata.json" "$CONFIG" "$OUTPUT_BASE" "$REPETITIONS" "$ORDER_SEED" "$DATASET_FILTER" "$PROFILE" <<'PY'
 import datetime
+import hashlib
 import json
 import os
 import pathlib
@@ -53,7 +60,24 @@ config_path = pathlib.Path(sys.argv[2])
 output_base = pathlib.Path(sys.argv[3])
 repetitions = int(sys.argv[4])
 order_seed = int(sys.argv[5])
+dataset_filter = sys.argv[6]
+profile = sys.argv[7] == "1"
 config = yaml.safe_load(config_path.read_text())
+if dataset_filter:
+    selected = [
+        path
+        for path in config["input_files"]
+        if path == dataset_filter or pathlib.Path(path).name == dataset_filter
+    ]
+    if len(selected) != 1:
+        raise ValueError(
+            f"Dataset filter {dataset_filter!r} matched {len(selected)} inputs"
+        )
+    config["input_files"] = selected
+if profile:
+    config.setdefault("env", {})["DYN_DISABLE_NSYS"] = "0"
+    config["env"]["DYN_NSYS_TRACE"] = "cuda,nvtx"
+    config["env"]["DYN_NSYS_DIR"] = str(output_base / "nsys")
 configs = config["configs"]
 if not configs:
     raise ValueError(f"No configs found in {config_path}")
@@ -69,6 +93,14 @@ def command_output(args: list[str]) -> str | None:
         return None
 
 
+def sha256(path: pathlib.Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as source:
+        for chunk in iter(lambda: source.read(1024 * 1024), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
 gpu_inventory = command_output(
     [
         "nvidia-smi",
@@ -127,10 +159,21 @@ metadata = {
     "gpu_inventory": gpu_inventory.splitlines() if gpu_inventory else [],
     "gpu_topology": gpu_topology,
     "datasets": config["input_files"],
+    "dataset_sha256": {
+        path: sha256(pathlib.Path(path)) for path in config["input_files"]
+    },
     "sweep_mode": sweep_mode,
     "sweep_values": sweep_values,
     "utc_start_time": datetime.datetime.now(datetime.timezone.utc).isoformat(),
 }
+prompt_manifest = config.get("prompt_manifest")
+if prompt_manifest:
+    prompt_manifest_path = pathlib.Path(prompt_manifest)
+    metadata["prompt_manifest"] = prompt_manifest
+    metadata["prompt_manifest_sha256"] = sha256(prompt_manifest_path)
+    metadata["prompt_manifest_contents"] = json.loads(
+        prompt_manifest_path.read_text()
+    )
 for arm in configs:
     args = arm.get("extra_args", [])
     label = arm["label"]
@@ -173,7 +216,7 @@ if [[ ${#arms[@]} -eq 0 || -z "${arms[0]}" ]]; then
     exit 2
 fi
 
-shape_raw="$("$PYTHON_BIN" - "$CONFIG" <<'PY'
+shape_raw="$("$PYTHON_BIN" - "$OUTPUT_BASE/config-rep-1.yaml" <<'PY'
 import pathlib
 import sys
 
@@ -225,6 +268,53 @@ PY
         --config "$iteration_config" \
         --output-dir "$OUTPUT_BASE/rep-$iteration" \
         --skip-plots
+    "$PYTHON_BIN" - "$iteration_config" "$OUTPUT_BASE/rep-$iteration" <<'PY'
+import json
+import pathlib
+import sys
+
+import yaml
+
+from benchmarks.multimodal.sweep.config import input_file_tag
+
+config = yaml.safe_load(pathlib.Path(sys.argv[1]).read_text())
+output = pathlib.Path(sys.argv[2])
+concurrencies = config.get("concurrencies")
+sweep_mode = "concurrency" if concurrencies else "request_rate"
+sweep_values = concurrencies or config.get("request_rates") or [4, 8, 16, 32, 64]
+for dataset in config["input_files"]:
+    expected_requests = sum(
+        1 for line in pathlib.Path(dataset).read_text().splitlines() if line.strip()
+    )
+    for arm in config["configs"]:
+        for value in sweep_values:
+            artifact = (
+                output
+                / input_file_tag(dataset)
+                / arm["label"]
+                / f"{sweep_mode}{value}"
+            )
+            profile_path = artifact / "profile_export_aiperf.json"
+            profile = json.loads(profile_path.read_text())
+            completed = int(profile["request_count"]["avg"])
+            if completed != expected_requests:
+                raise ValueError(
+                    f"{profile_path}: completed {completed}, expected {expected_requests}"
+                )
+            if profile.get("was_cancelled"):
+                raise ValueError(f"{profile_path}: benchmark was cancelled")
+            if profile.get("error_summary"):
+                raise ValueError(
+                    f"{profile_path}: errors present: {profile['error_summary']}"
+                )
+            minimum = config.get("prefix_cache_probe_min_cached_tokens")
+            if minimum is not None:
+                probe_path = artifact / "prefix_cache_probe.json"
+                probe = json.loads(probe_path.read_text())
+                if not probe.get("passed"):
+                    raise ValueError(f"{probe_path}: prefix-cache probe failed")
+print("[sweep] ARTIFACTS_VALIDATED")
+PY
     echo "[sweep] END_ITER_${iteration}"
     if [[ "$iteration" == "1" ]]; then
         for shape in "${sweep_shapes[@]}"; do
diff --git a/benchmarks/multimodal/sweep/workflows/vllm_serve.sh b/benchmarks/multimodal/sweep/workflows/vllm_serve.sh
index b4234db47597..e7ce6f30bd76 100755
--- a/benchmarks/multimodal/sweep/workflows/vllm_serve.sh
+++ b/benchmarks/multimodal/sweep/workflows/vllm_serve.sh
@@ -53,6 +53,20 @@ if [[ -z "$MODEL" ]]; then
     exit 1
 fi
 
+if [[ -n "${DYN_VLLM_EXPECTED_ROOT:-}" ]]; then
+    expected_root="${DYN_VLLM_EXPECTED_ROOT%%:*}"
+    active_vllm="$(python -c 'import pathlib, vllm; print(pathlib.Path(vllm.__file__).resolve())')"
+    case "$active_vllm" in
+        "$expected_root"/*) ;;
+        *)
+            echo "ERROR: active vLLM $active_vllm is not under $expected_root" >&2
+            exit 2
+            ;;
+    esac
+    python -c 'import vllm._C_stable_libtorch, vllm._custom_ops'
+    echo "[vllm] source=${DYN_VLLM_SOURCE_REVISION:-unknown} file=$active_vllm"
+fi
+
 EC_ARGS=()
 if [[ "$CAPACITY_GB" != "0" ]]; then
     EC_ARGS=(--ec-transfer-config "{
diff --git a/tests/benchmarks/multimodal/jsonl/test_main.py b/tests/benchmarks/multimodal/jsonl/test_main.py
index 8c55afd312f3..0a694e8c7f5e 100644
--- a/tests/benchmarks/multimodal/jsonl/test_main.py
+++ b/tests/benchmarks/multimodal/jsonl/test_main.py
@@ -9,7 +9,9 @@
 from pathlib import Path
 from unittest.mock import patch
 
+import numpy as np
 import pytest
+from PIL import Image
 
 pytest.importorskip("PIL", reason="Pillow required for image generation benchmarks")
 
@@ -17,6 +19,7 @@
 JSONL_DIR = Path(__file__).resolve().parents[4] / "benchmarks" / "multimodal" / "jsonl"
 sys.path.insert(0, str(JSONL_DIR))
 
+from generate_images import generate_image_pool_base64  # noqa: E402
 from generate_videos import generate_synthetic_video_pool  # noqa: E402
 from main import main  # noqa: E402
 
@@ -78,6 +81,14 @@ def test_default_produces_independent_requests(self, tmp_path: Path) -> None:
             assert len(line["images"]) == 2
             assert "session_id" not in line
 
+    def test_non_square_image_size_is_width_then_height(self, tmp_path: Path) -> None:
+        pool = generate_image_pool_base64(
+            np.random.default_rng(1), 1, tmp_path / "imgs", (40, 24)
+        )
+
+        with Image.open(pool[0]) as image:
+            assert image.size == (40, 24)
+
 
 class TestSlidingWindow:
     """sliding-window produces causal sessions with image overlap."""
diff --git a/tests/benchmarks/multimodal/jsonl/test_prepare_ec_h2d_overlap.py b/tests/benchmarks/multimodal/jsonl/test_prepare_ec_h2d_overlap.py
new file mode 100644
index 000000000000..0600441677b3
--- /dev/null
+++ b/tests/benchmarks/multimodal/jsonl/test_prepare_ec_h2d_overlap.py
@@ -0,0 +1,75 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+import pytest
+
+from benchmarks.multimodal.jsonl.prepare_ec_h2d_overlap import (
+    build_chat_template,
+    make_exact_text,
+    write_sliding_dataset,
+)
+
+pytestmark = [pytest.mark.unit, pytest.mark.pre_merge, pytest.mark.gpu_0]
+
+
+class CharacterTokenizer:
+    def encode(self, text: str, *, add_special_tokens: bool = False) -> list[int]:
+        return [ord(character) for character in text]
+
+    def decode(
+        self,
+        ids: list[int],
+        *,
+        skip_special_tokens: bool = True,
+        clean_up_tokenization_spaces: bool = False,
+    ) -> str:
+        return "".join(chr(token_id) for token_id in ids)
+
+
+def test_exact_text_hits_requested_token_count() -> None:
+    tokenizer = CharacterTokenizer()
+
+    value = make_exact_text(tokenizer, 257)
+
+    assert len(tokenizer.encode(value)) == 257
+
+
+def test_chat_template_prepends_system_message_only_when_missing() -> None:
+    base = "{% for message in messages %}[{{ message['role'] }}]{{ message['content'] }}{% endfor %}"
+    template = build_chat_template(base, "shared prefix")
+
+    assert "messages[0]['role'] != 'system'" in template
+    assert "'role': 'system', 'content': \"shared prefix\"" in template
+    assert template.endswith(base)
+
+
+def test_sliding_datasets_share_fixed_text_and_stable_overlap(tmp_path: Path) -> None:
+    pool = [str(tmp_path / f"image-{index}.png") for index in range(8)]
+    output = tmp_path / "dataset.jsonl"
+
+    counts = write_sliding_dataset(
+        output,
+        pool,
+        num_users=2,
+        turns_per_user=2,
+        window_size=3,
+        images_per_user=4,
+        user_text="same text",
+    )
+    rows = [json.loads(line) for line in output.read_text().splitlines()]
+
+    assert counts == {
+        "rows": 4,
+        "unique_images": 8,
+        "content_images": 8,
+        "stripped_images": 4,
+    }
+    assert {row["text"] for row in rows} == {"same text"}
+    user_zero = [row for row in rows if row["session_id"] == "user_0"]
+    assert user_zero[0]["images"][1:] == user_zero[1]["images"][:-1]
+    assert all(len(row["image_uuids"]) == 3 for row in rows)
diff --git a/tests/benchmarks/multimodal/sweep/test_orchestrator.py b/tests/benchmarks/multimodal/sweep/test_orchestrator.py
index 0dc99d96e6fc..7683bf8e888e 100644
--- a/tests/benchmarks/multimodal/sweep/test_orchestrator.py
+++ b/tests/benchmarks/multimodal/sweep/test_orchestrator.py
@@ -219,6 +219,33 @@ def test_uuid_and_strip_propagates_to_aiperf(
     )
 
 
+@patch("benchmarks.multimodal.sweep.orchestrator.run_aiperf_single")
+@patch("benchmarks.multimodal.sweep.orchestrator.ServerManager")
+def test_prefix_cache_probe_runs_before_aiperf(
+    mock_server_cls: MagicMock,
+    mock_aiperf: MagicMock,
+    tmp_path: Path,
+) -> None:
+    config = _make_config(
+        tmp_path,
+        num_configs=1,
+        num_input_files=1,
+        request_rates=[4],
+    )
+    Path(config.input_files[0]).write_text('{"text":"same question"}\n')
+    config.prefix_cache_probe_min_cached_tokens = 7936
+    mock_server_cls.return_value.is_running = False
+
+    run_sweep(config, repo_root=tmp_path)
+
+    probe = mock_server_cls.return_value.validate_prefix_cache
+    probe.assert_called_once()
+    assert probe.call_args.kwargs["user_text"] == "same question"
+    assert probe.call_args.kwargs["min_cached_tokens"] == 7936
+    assert probe.call_args.kwargs["output_path"].name == "prefix_cache_probe.json"
+    assert mock_aiperf.call_count == 1
+
+
 def test_per_arm_env_expands_and_overrides_top_level(
     monkeypatch: pytest.MonkeyPatch,
 ) -> None:
diff --git a/tests/benchmarks/multimodal/sweep/test_server.py b/tests/benchmarks/multimodal/sweep/test_server.py
index fb4b8fffe63a..919170099dc3 100644
--- a/tests/benchmarks/multimodal/sweep/test_server.py
+++ b/tests/benchmarks/multimodal/sweep/test_server.py
@@ -1,6 +1,7 @@
 # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
+import json
 from unittest.mock import MagicMock, patch
 
 import pytest
@@ -94,3 +95,77 @@ def test_start_treats_empty_timeout_override_as_unset(tmp_path) -> None:
         )
 
     assert manager.terminate_timeout == 15.0
+
+
+def test_validate_prefix_cache_persists_successful_probe(tmp_path) -> None:
+    responses = [
+        {"usage": {"prompt_tokens": 8016, "prompt_tokens_details": {}}},
+        {
+            "usage": {
+                "prompt_tokens": 8016,
+                "prompt_tokens_details": {"cached_tokens": 8000},
+            }
+        },
+    ]
+
+    class FakeResponse:
+        def __init__(self, value: dict) -> None:
+            self.value = value
+
+        def __enter__(self):
+            return self
+
+        def __exit__(self, *_args) -> None:
+            return None
+
+        def read(self) -> bytes:
+            return json.dumps(self.value).encode()
+
+    output = tmp_path / "probe.json"
+    with patch(
+        "benchmarks.multimodal.sweep.server.urllib.request.urlopen",
+        side_effect=[FakeResponse(value) for value in responses],
+    ) as urlopen:
+        result = ServerManager(port=8123).validate_prefix_cache(
+            model="model",
+            user_text="question",
+            min_cached_tokens=7936,
+            output_path=output,
+        )
+
+    assert urlopen.call_count == 2
+    assert result["passed"] is True
+    assert json.loads(output.read_text()) == result
+
+
+def test_validate_prefix_cache_rejects_uncached_prompt(tmp_path) -> None:
+    response = {
+        "usage": {
+            "prompt_tokens": 8016,
+            "prompt_tokens_details": {"cached_tokens": 128},
+        }
+    }
+
+    class FakeResponse:
+        def __enter__(self):
+            return self
+
+        def __exit__(self, *_args) -> None:
+            return None
+
+        def read(self) -> bytes:
+            return json.dumps(response).encode()
+
+    with (
+        patch(
+            "benchmarks.multimodal.sweep.server.urllib.request.urlopen",
+            side_effect=[FakeResponse(), FakeResponse()],
+        ),
+        pytest.raises(RuntimeError, match="cached 128 tokens"),
+    ):
+        ServerManager().validate_prefix_cache(
+            model="model",
+            user_text="question",
+            min_cached_tokens=7936,
+            output_path=tmp_path / "probe.json",
+        )
diff --git a/tests/benchmarks/multimodal/sweep/test_vllm_serve_workflow.py b/tests/benchmarks/multimodal/sweep/test_vllm_serve_workflow.py
index 97341f0cb72c..8d4e5a3030f8 100644
--- a/tests/benchmarks/multimodal/sweep/test_vllm_serve_workflow.py
+++ b/tests/benchmarks/multimodal/sweep/test_vllm_serve_workflow.py
@@ -26,13 +26,22 @@ def test_embedding_cache_sweep_selects_only_requested_arms() -> None:
     config = yaml.safe_load(CONFIG.read_text())
 
     assert [item["label"] for item in config["configs"]] == [
-        "vllm-serve",
         "vllm-serve-native-ec-baseline",
         "vllm-serve-native-ec-overlap",
     ]
     assert config["env"]["DYN_DISABLE_NSYS"] == "1"
+    assert config["prefix_cache_probe_min_cached_tokens"] == 7936
+    assert len(config["input_files"]) == 2
     assert config["configs"][0]["env"]["PYTHONPATH"] == ("${VLLM_BASELINE_PYTHONPATH}")
-    assert config["configs"][2]["env"]["PYTHONPATH"] == ("${VLLM_PATCHED_PYTHONPATH}")
+    assert config["configs"][1]["env"]["PYTHONPATH"] == ("${VLLM_PATCHED_PYTHONPATH}")
+    for benchmark in config["configs"]:
+        assert "--enable-prefix-caching" in benchmark["extra_args"]
+        assert "--enable-prompt-tokens-details" in benchmark["extra_args"]
+        assert "--chat-template" in benchmark["extra_args"]
+        ec_config = benchmark["extra_args"][
+            benchmark["extra_args"].index("--ec-transfer-config") + 1
+        ]
+        assert '"ec_cpu_bytes":8589934592' in ec_config
 
 
 def _run_workflow(tmp_path: Path, *args: str, enable_nsys: bool = False):

From 681b4f7578eed28a5ecca291dfeaaee0520533c7 Mon Sep 17 00:00:00 2001
From: furionw <qiwa@nvidia.com>
Date: Thu, 17 Sep 2026 05:37:55 -0700
Subject: [PATCH 6/7] fix(benchmarks): count rendered prompt input IDs

---
 .../multimodal/jsonl/prepare_ec_h2d_overlap.py    | 15 +++++++++++++--
 .../jsonl/test_prepare_ec_h2d_overlap.py          | 11 +++++++++++
 2 files changed, 24 insertions(+), 2 deletions(-)

diff --git a/benchmarks/multimodal/jsonl/prepare_ec_h2d_overlap.py b/benchmarks/multimodal/jsonl/prepare_ec_h2d_overlap.py
index 413868a71f9a..31da3eacdba1 100755
--- a/benchmarks/multimodal/jsonl/prepare_ec_h2d_overlap.py
+++ b/benchmarks/multimodal/jsonl/prepare_ec_h2d_overlap.py
@@ -9,7 +9,7 @@
 import argparse
 import hashlib
 import json
-from collections.abc import Sequence
+from collections.abc import Mapping, Sequence
 from pathlib import Path
 from typing import Any, Protocol
 
@@ -69,6 +69,17 @@ def count_tokens(tokenizer: Tokenizer, text: str) -> int:
     return len(tokenizer.encode(text, add_special_tokens=False))
 
 
+def tokenized_length(tokenized: Any) -> int:
+    """Count input IDs returned as a sequence or tokenizer batch encoding."""
+    if isinstance(tokenized, Mapping):
+        tokenized = tokenized["input_ids"]
+    if tokenized and isinstance(tokenized[0], Sequence):
+        if len(tokenized) != 1:
+            raise ValueError("expected exactly one rendered conversation")
+        tokenized = tokenized[0]
+    return len(tokenized)
+
+
 def make_exact_text(tokenizer: Tokenizer, target_tokens: int) -> str:
     """Return deterministic text with exactly ``target_tokens`` tokenizer tokens."""
     if target_tokens <= 0:
@@ -246,7 +257,7 @@ def main() -> None:
         },
         "user_text": USER_TEXT,
         "user_text_tokens": count_tokens(tokenizer, USER_TEXT),
-        "rendered_text_prompt_tokens": len(rendered_ids),
+        "rendered_text_prompt_tokens": tokenized_length(rendered_ids),
         "images": {
             "directory": str(args.image_dir),
             "count": len(image_pool),
diff --git a/tests/benchmarks/multimodal/jsonl/test_prepare_ec_h2d_overlap.py b/tests/benchmarks/multimodal/jsonl/test_prepare_ec_h2d_overlap.py
index 0600441677b3..937f2fe57fde 100644
--- a/tests/benchmarks/multimodal/jsonl/test_prepare_ec_h2d_overlap.py
+++ b/tests/benchmarks/multimodal/jsonl/test_prepare_ec_h2d_overlap.py
@@ -11,6 +11,7 @@
 from benchmarks.multimodal.jsonl.prepare_ec_h2d_overlap import (
     build_chat_template,
     make_exact_text,
+    tokenized_length,
     write_sliding_dataset,
 )
 
@@ -39,6 +40,16 @@ def test_exact_text_hits_requested_token_count() -> None:
     assert len(tokenizer.encode(value)) == 257
 
 
+@pytest.mark.parametrize(
+    ("tokenized", "expected"),
+    [([1, 2, 3], 3), ({"input_ids": [1, 2, 3]}, 3), ({"input_ids": [[1, 2, 3]]}, 3)],
+)
+def test_tokenized_length_accepts_sequence_and_batch_encoding_shapes(
+    tokenized: object, expected: int
+) -> None:
+    assert tokenized_length(tokenized) == expected
+
+
 def test_chat_template_prepends_system_message_only_when_missing() -> None:
     base = "{% for message in messages %}[{{ message['role'] }}]{{ message['content'] }}{% endfor %}"
     template = build_chat_template(base, "shared prefix")

From 92c32c8c64d93287a296766f9d2f699ae119c228 Mon Sep 17 00:00:00 2001
From: furionw <qiwa@nvidia.com>
Date: Thu, 17 Sep 2026 05:53:06 -0700
Subject: [PATCH 7/7] fix(benchmarks): account for hybrid prefix alignment

---
 .../sweep/experiments/embedding_cache/vllm_serve.yaml        | 5 ++++-
 .../benchmarks/multimodal/sweep/test_vllm_serve_workflow.py  | 2 +-
 2 files changed, 5 insertions(+), 2 deletions(-)

diff --git a/benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml b/benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml
index e49622da2371..42bd15ed74b1 100644
--- a/benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml
+++ b/benchmarks/multimodal/sweep/experiments/embedding_cache/vllm_serve.yaml
@@ -19,7 +19,10 @@ skip_plots: true
 restart_server_every_benchmark: true
 uuid_and_strip: true
 prompt_manifest: /dynamo-tmp/data/qwen35_ec_h2d_overlap_manifest.json
-prefix_cache_probe_min_cached_tokens: 7936
+# Qwen3.5's hybrid Mamba/KV cache aligns this 8026-token rendered prompt to
+# 6288 reusable tokens. Require a stable majority hit without assuming pure-KV
+# block accounting.
+prefix_cache_probe_min_cached_tokens: 6000
 
 env:
   DYN_DISABLE_NSYS: "1"
diff --git a/tests/benchmarks/multimodal/sweep/test_vllm_serve_workflow.py b/tests/benchmarks/multimodal/sweep/test_vllm_serve_workflow.py
index 8d4e5a3030f8..19204127742a 100644
--- a/tests/benchmarks/multimodal/sweep/test_vllm_serve_workflow.py
+++ b/tests/benchmarks/multimodal/sweep/test_vllm_serve_workflow.py
@@ -30,7 +30,7 @@ def test_embedding_cache_sweep_selects_only_requested_arms() -> None:
         "vllm-serve-native-ec-overlap",
     ]
     assert config["env"]["DYN_DISABLE_NSYS"] == "1"
-    assert config["prefix_cache_probe_min_cached_tokens"] == 7936
+    assert config["prefix_cache_probe_min_cached_tokens"] == 6000
     assert len(config["input_files"]) == 2
     assert config["configs"][0]["env"]["PYTHONPATH"] == ("${VLLM_BASELINE_PYTHONPATH}")
     assert config["configs"][1]["env"]["PYTHONPATH"] == ("${VLLM_PATCHED_PYTHONPATH}")