chore(multimodal): Cleanup multimodal docs and consolidate launch scripts (#7845)

examples/backends/vllm/launch/agg_multimodal.sh

@@ -2,11 +2,11 @@
 # SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 #
-# Aggregated multimodal serving with standard Dynamo preprocessing
+# Aggregated multimodal image/video serving with standard Dynamo preprocessing
 #
 # Architecture: Single-worker PD (Prefill-Decode)
-# - Frontend: Rust OpenAIPreprocessor handles image URLs (HTTP and data:// base64)
-# - Worker: Standard vLLM worker with vision model support
+# - Frontend: Rust OpenAIPreprocessor forwards multimodal requests
+# - Worker: Standard vLLM worker with multimodal model support
 #
 # For EPD (Encode-Prefill-Decode) architecture with dedicated encoding worker,
 # see agg_multimodal_epd.sh
@@ -19,7 +19,7 @@ source "$SCRIPT_DIR/../../../common/gpu_utils.sh"
 source "$SCRIPT_DIR/../../../common/launch_utils.sh"
 
 # Default values
-MODEL_NAME="Qwen/Qwen3-VL-30B-A3B-Instruct-FP8"
+MODEL_NAME="${DYN_MODEL_NAME:-Qwen/Qwen3-VL-30B-A3B-Instruct-FP8}"
 
 # Parse command line arguments
 # Extra arguments are passed through to the vLLM worker
@@ -48,13 +48,41 @@ while [[ $# -gt 0 ]]; do
 done
 
 HTTP_PORT="${DYN_HTTP_PORT:-8000}"
-print_launch_banner --multimodal "Launching Aggregated Multimodal Serving" "$MODEL_NAME" "$HTTP_PORT"
 
 # Use TCP transport (instead of default NATS)
 # TCP is preferred for multimodal workloads because it overcomes:
 # - NATS default 1MB max payload limit (multimodal base64 images can exceed this)
 export DYN_REQUEST_PLANE=tcp
 
+print_launch_banner --no-curl "Launching Aggregated Multimodal Serving" "$MODEL_NAME" "$HTTP_PORT" \
+    "Backend:     dynamo.vllm --enable-multimodal" \
+    "Media:       image_url and video_url (model support dependent)"
+
+print_curl_footer <<CURL
+  curl http://localhost:${HTTP_PORT}/v1/chat/completions \\
+    -H 'Content-Type: application/json' \\
+    -d '{
+      "model": "${MODEL_NAME}",
+      "messages": [{"role": "user", "content": [
+        {"type": "text", "text": "Describe the image"},
+        {"type": "image_url", "image_url": {"url": "https://upload.wikimedia.org/wikipedia/commons/thumb/4/47/PNG_transparency_demonstration_1.png/300px-PNG_transparency_demonstration_1.png"}}
+      ]}],
+      "max_tokens": 50
+    }'
+
+  # For video-capable models such as Qwen/Qwen3-VL-2B-Instruct:
+  curl http://localhost:${HTTP_PORT}/v1/chat/completions \\
+    -H 'Content-Type: application/json' \\
+    -d '{
+      "model": "Qwen/Qwen3-VL-2B-Instruct",
+      "messages": [{"role": "user", "content": [
+        {"type": "text", "text": "Describe the video in detail"},
+        {"type": "video_url", "video_url": {"url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/draw.mp4"}}
+      ]}],
+      "max_tokens": 128
+    }'
+CURL
+
 # Start frontend with Rust OpenAIPreprocessor
 # dynamo.frontend accepts either --http-port flag or DYN_HTTP_PORT env var (defaults to 8000)
 python -m dynamo.frontend &
@@ -65,7 +93,7 @@ MAX_CONCURRENT_SEQS="${MAX_CONCURRENT_SEQS:-2}"
 MODEL_EXTRA_ARGS=""
 case "$MODEL_NAME" in
     meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8)
-        MAX_MODEL_LEN="${MAX_MODEL_LEN:-108960}"
+        MAX_MODEL_LEN="108960"
         MODEL_EXTRA_ARGS="--tensor-parallel-size=8" ;;
 esac
 

examples/backends/vllm/launch/disagg_multimodal_e_pd.sh

@@ -7,6 +7,9 @@ trap 'echo Cleaning up...; kill 0' EXIT
 SCRIPT_DIR="$(dirname "$(readlink -f "$0")")"
 source "$SCRIPT_DIR/../../../common/launch_utils.sh"
 
+# Use TCP transport for multimodal workloads (base64 images can exceed NATS 1MB limit)
+export DYN_REQUEST_PLANE=tcp
+
 # Default values
 MODEL_NAME="Qwen/Qwen3-VL-30B-A3B-Instruct-FP8"
 SINGLE_GPU=false

examples/backends/vllm/launch/disagg_multimodal_epd.sh

@@ -8,6 +8,9 @@ SCRIPT_DIR="$(dirname "$(readlink -f "$0")")"
 source "$SCRIPT_DIR/../../../common/gpu_utils.sh"
 source "$SCRIPT_DIR/../../../common/launch_utils.sh"
 
+# Use TCP transport for multimodal workloads (base64 images can exceed NATS 1MB limit)
+export DYN_REQUEST_PLANE=tcp
+
 # Default values
 MODEL_NAME="llava-hf/llava-1.5-7b-hf"
 
@@ -17,7 +20,7 @@ MODEL_NAME="llava-hf/llava-1.5-7b-hf"
 #   - Enabling --enforce-eager (disables torch.compile and CUDA graph capture)
 #   - Hardcoding P/D KV cache to 512 MB (skips all memory profiling)
 #   - Limiting --max-model-len to 4096 tokens on P/D workers
-#   - Limiting P/D workers to image=1,video=0,audio=0 (--limit-mm-per-prompt)
+#   - Limiting P/D workers to image=3,video=3,audio=0 (--limit-mm-per-prompt)
 #   - Using lower gpu-memory-utilization fractions to share the GPU
 SINGLE_GPU=false
 
@@ -77,10 +80,17 @@ python -m dynamo.frontend &
 EXTRA_ARGS=""
 PD_EXTRA_ARGS=""
 
-# GPU assignments (override via environment variables)
-DYN_ENCODE_WORKER_GPU=${DYN_ENCODE_WORKER_GPU:-0}
-DYN_PREFILL_WORKER_GPU=${DYN_PREFILL_WORKER_GPU:-1}
-DYN_DECODE_WORKER_GPU=${DYN_DECODE_WORKER_GPU:-2}
+# GPU assignments (override via environment variables).
+# In single-GPU mode all 3 workers default to GPU 0.
+if [[ "$SINGLE_GPU" == "true" ]]; then
+    DYN_ENCODE_WORKER_GPU=${DYN_ENCODE_WORKER_GPU:-0}
+    DYN_PREFILL_WORKER_GPU=${DYN_PREFILL_WORKER_GPU:-0}
+    DYN_DECODE_WORKER_GPU=${DYN_DECODE_WORKER_GPU:-0}
+else
+    DYN_ENCODE_WORKER_GPU=${DYN_ENCODE_WORKER_GPU:-0}
+    DYN_PREFILL_WORKER_GPU=${DYN_PREFILL_WORKER_GPU:-1}
+    DYN_DECODE_WORKER_GPU=${DYN_DECODE_WORKER_GPU:-2}
+fi
 
 # GPU memory utilization for workers.
 # NOTE: --kv-cache-memory-bytes (set below for P/D workers) overrides
@@ -93,9 +103,15 @@ if [[ -n "${_PROFILE_PYTEST_VRAM_FRAC_OVERRIDE:-}" ]]; then
     echo "WARNING: _PROFILE_PYTEST_VRAM_FRAC_OVERRIDE is set but has no effect here because" >&2
     echo "  --kv-cache-memory-bytes overrides --gpu-memory-utilization in vLLM." >&2
 fi
-DYN_ENCODE_GPU_MEM=${DYN_ENCODE_GPU_MEM:-0.9}
-DYN_PREFILL_GPU_MEM=${DYN_PREFILL_GPU_MEM:-0.9}
-DYN_DECODE_GPU_MEM=${DYN_DECODE_GPU_MEM:-0.9}
+if [[ "$SINGLE_GPU" == "true" ]]; then
+    DYN_ENCODE_GPU_MEM=${DYN_ENCODE_GPU_MEM:-0.1}
+    DYN_PREFILL_GPU_MEM=${DYN_PREFILL_GPU_MEM:-0.4}
+    DYN_DECODE_GPU_MEM=${DYN_DECODE_GPU_MEM:-0.4}
+else
+    DYN_ENCODE_GPU_MEM=${DYN_ENCODE_GPU_MEM:-0.9}
+    DYN_PREFILL_GPU_MEM=${DYN_PREFILL_GPU_MEM:-0.9}
+    DYN_DECODE_GPU_MEM=${DYN_DECODE_GPU_MEM:-0.9}
+fi
 
 # 512 MB KV cache per P/D worker. Setting --kv-cache-memory-bytes bypasses vLLM's
 # memory profiling entirely (both language model and multimodal encoder), which avoids
@@ -105,7 +121,7 @@ PD_KV_CACHE_BYTES=$((512 * 1024 * 1024))
 
 if [[ "$SINGLE_GPU" == "true" ]]; then
     EXTRA_ARGS="--enforce-eager"
-    PD_EXTRA_ARGS="--max-model-len 4096 --kv-cache-memory-bytes $PD_KV_CACHE_BYTES --limit-mm-per-prompt {\"image\":1,\"video\":0,\"audio\":0}"
+    PD_EXTRA_ARGS="--max-model-len 4096 --kv-cache-memory-bytes $PD_KV_CACHE_BYTES --limit-mm-per-prompt {\"image\":3,\"video\":3,\"audio\":0}"
 fi
 
 # Start encode worker

examples/backends/vllm/launch/disagg_multimodal_llama.sh

-#!/bin/bash
-# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
-# SPDX-License-Identifier: Apache-2.0
-set -ex
-
-SCRIPT_DIR="$(dirname "$(readlink -f "$0")")"
-source "$SCRIPT_DIR/../../../common/launch_utils.sh"
-
-# Default values
-HEAD_NODE=0
-MODEL_NAME="meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8"
-EXTRA_ARGS=()
-
-# Parse command line arguments
-while [[ $# -gt 0 ]]; do
-    case $1 in
-        --head-node)
-            HEAD_NODE=1
-            shift 1
-            ;;
-        --model)
-            MODEL_NAME=$2
-            shift 2
-            ;;
-        -h|--help)
-            echo "Usage: $0 [OPTIONS]"
-            echo ""
-            echo "Disaggregated multimodal serving with separate Prefill/Decode workers for Llama 4"
-            echo ""
-            echo "Options:"
-            echo "  --head-node          Run as head node. Head node will run the HTTP server, processor and prefill worker."
-            echo "  --model <model_name> Specify the VLM model to use (default: $MODEL_NAME)"
-            echo "  -h, --help           Show this help message"
-            echo ""
-            echo "Examples:"
-            echo "  # On head node:"
-            echo "  $0 --head-node"
-            echo ""
-            echo "  # On worker node (requires NATS_SERVER and ETCD_ENDPOINTS pointing to head node):"
-            echo "  $0"
-            echo ""
-            exit 0
-            ;;
-        *)
-            EXTRA_ARGS+=("$1")
-            shift
-            ;;
-    esac
-done
-
-trap 'echo Cleaning up...; kill 0' EXIT
-
-HTTP_PORT="${DYN_HTTP_PORT:-8000}"
-if [[ $HEAD_NODE -eq 1 ]]; then
-    print_launch_banner --multimodal "Launching Disaggregated Multimodal Llama 4 (Multi-Node)" "$MODEL_NAME" "$HTTP_PORT"
-else
-    print_launch_banner --no-curl "Launching Disaggregated Multimodal Llama 4 (Multi-Node)" "$MODEL_NAME" "$HTTP_PORT"
-fi
-
-# Use TCP transport to avoid NATS payload limits for multimodal
-export DYN_REQUEST_PLANE=tcp
-
-# Configure model-specific args
-GPU_MEM="0.80"
-KV_BYTES="${_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 $GPU_MEM"
-fi
-MODEL_SPECIFIC_ARGS=""
-if [[ "$MODEL_NAME" == "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8" ]]; then
-    MODEL_SPECIFIC_ARGS="--tensor-parallel-size=8 --max-model-len=208960 $GPU_MEM_ARGS"
-fi
-
-if [[ $HEAD_NODE -eq 1 ]]; then
-    # run ingress
-    # dynamo.frontend accepts either --http-port flag or DYN_HTTP_PORT env var (defaults to 8000)
-    python -m dynamo.frontend &
-
-    # run processor (CPU-only to avoid competing for GPU memory with workers)
-    CUDA_VISIBLE_DEVICES="" \
-    python -m dynamo.vllm --route-to-encoder --enable-multimodal --model $MODEL_NAME &
-
-    # Prefill worker handles prompt processing and image encoding
-    # Uses all 8 GPUs for tensor-parallel
-    CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
-    VLLM_NIXL_SIDE_CHANNEL_PORT=20097 \
-    python -m dynamo.vllm \
-        --enable-multimodal \
-        --model $MODEL_NAME \
-        --disaggregation-mode prefill \
-        --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both"}' \
-        $MODEL_SPECIFIC_ARGS \
-        --kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:20080"}' \
-        "${EXTRA_ARGS[@]}" &
-else
-    # run decode worker on non-head node
-    # Uses all 8 GPUs for tensor-parallel
-    CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
-    VLLM_NIXL_SIDE_CHANNEL_PORT=20098 \
-    python -m dynamo.vllm \
-        --enable-multimodal \
-        --model $MODEL_NAME \
-        --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both"}' \
-        $MODEL_SPECIFIC_ARGS \
-        --kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:20081"}' \
-        "${EXTRA_ARGS[@]}" &
-fi
-
-# Exit on first worker failure; kill 0 in the EXIT trap tears down the rest
-wait_any_exit

examples/backends/vllm/launch/video_agg.sh

-#!/bin/bash
-# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
-# SPDX-License-Identifier: Apache-2.0
-#
-# Aggregated video serving with standard Dynamo preprocessing and vLLM backend.
-
-set -euo pipefail
-
-cleanup() {
-    echo "Cleaning up..."
-    local pids
-    pids="$(jobs -pr)"
-    if [[ -n "$pids" ]]; then
-        kill $pids 2>/dev/null || true
-    fi
-}
-
-trap cleanup EXIT
-
-SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-REPO_ROOT="$(cd "$SCRIPT_DIR/../../../.." && pwd)"
-source "$SCRIPT_DIR/../../../common/gpu_utils.sh"
-source "$SCRIPT_DIR/../../../common/launch_utils.sh"
-
-export PYTHONPATH="${REPO_ROOT}/components/src:${REPO_ROOT}/lib/bindings/python/src${PYTHONPATH:+:${PYTHONPATH}}"
-
-MODEL_NAME="${DYN_MODEL_NAME:-Qwen/Qwen3-VL-2B-Instruct}"
-HTTP_PORT="${DYN_HTTP_PORT:-8000}"
-GPU_DEVICE="${CUDA_VISIBLE_DEVICES:-0}"
-MAX_MODEL_LEN="${MAX_MODEL_LEN:-8192}"
-MAX_NUM_SEQS="${MAX_NUM_SEQS:-2}"
-
-EXTRA_ARGS=()
-while [[ $# -gt 0 ]]; do
-    case $1 in
-        --model)
-            MODEL_NAME=$2
-            shift 2
-            ;;
-        -h|--help)
-            cat <<USAGE
-Usage: $0 [OPTIONS] [-- EXTRA_VLLM_ARGS]
-
-Options:
-  --model <model_name>   Video-capable VLM to serve (default: $MODEL_NAME)
-  -h, --help             Show this help message
-
-Any arguments after '--' are passed through to the vLLM worker.
-USAGE
-            exit 0
-            ;;
-        --)
-            shift
-            EXTRA_ARGS+=("$@")
-            break
-            ;;
-        *)
-            EXTRA_ARGS+=("$1")
-            shift
-            ;;
-    esac
-done
-
-export DYN_REQUEST_PLANE=tcp
-
-GPU_MEM_ARGS=$(build_gpu_mem_args vllm)
-
-print_launch_banner --no-curl "Launching Aggregated Video Serving" "$MODEL_NAME" "$HTTP_PORT" \
-    "Backend:     dynamo.vllm --enable-multimodal" \
-    "Video path:  Standard TokensPrompt multi_modal_data flow"
-
-print_curl_footer <<CURL
-  curl http://localhost:${HTTP_PORT}/v1/chat/completions \\
-    -H 'Content-Type: application/json' \\
-    -d '{
-      "model": "${MODEL_NAME}",
-      "messages": [{"role": "user", "content": [
-        {"type": "text", "text": "Describe the video in detail"},
-        {"type": "video_url", "video_url": {"url": "https://storage.googleapis.com/gtv-videos-bucket/sample/BigBuckBunny.mp4"}}
-      ]}],
-      "max_tokens": 128
-    }'
-CURL
-
-python -m dynamo.frontend &
-
-CUDA_VISIBLE_DEVICES="$GPU_DEVICE" \
-    python -m dynamo.vllm \
-        --enable-multimodal \
-        --model "$MODEL_NAME" \
-        --max-model-len "$MAX_MODEL_LEN" \
-        --max-num-seqs "$MAX_NUM_SEQS" \
-        $GPU_MEM_ARGS \
-        "${EXTRA_ARGS[@]}" &
-
-wait_any_exit

examples/backends/vllm/launch/video_disagg.sh

-#!/bin/bash
-# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
-# SPDX-License-Identifier: Apache-2.0
-#
-# Disaggregated video serving with standard Dynamo preprocessing and vLLM backend.
-
-set -euo pipefail
-
-cleanup() {
-    echo "Cleaning up..."
-    local pids
-    pids="$(jobs -pr)"
-    if [[ -n "$pids" ]]; then
-        kill $pids 2>/dev/null || true
-    fi
-}
-
-trap cleanup EXIT
-
-SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-REPO_ROOT="$(cd "$SCRIPT_DIR/../../../.." && pwd)"
-source "$SCRIPT_DIR/../../../common/gpu_utils.sh"
-source "$SCRIPT_DIR/../../../common/launch_utils.sh"
-
-export PYTHONPATH="${REPO_ROOT}/components/src:${REPO_ROOT}/lib/bindings/python/src${PYTHONPATH:+:${PYTHONPATH}}"
-
-MODEL_NAME="${DYN_MODEL_NAME:-Qwen/Qwen3-VL-2B-Instruct}"
-HTTP_PORT="${DYN_HTTP_PORT:-8000}"
-SINGLE_GPU=false
-EXTRA_ARGS=()
-
-while [[ $# -gt 0 ]]; do
-    case $1 in
-        --model)
-            MODEL_NAME=$2
-            shift 2
-            ;;
-        --single-gpu)
-            SINGLE_GPU=true
-            shift
-            ;;
-        -h|--help)
-            cat <<USAGE
-Usage: $0 [OPTIONS] [-- EXTRA_VLLM_ARGS]
-
-Options:
-  --model <model_name>   Video-capable VLM to serve (default: $MODEL_NAME)
-  --single-gpu           Run prefill and decode on one GPU for functional testing
-  -h, --help             Show this help message
-
-Any arguments after '--' are passed through to both vLLM workers.
-USAGE
-            exit 0
-            ;;
-        --)
-            shift
-            EXTRA_ARGS+=("$@")
-            break
-            ;;
-        *)
-            EXTRA_ARGS+=("$1")
-            shift
-            ;;
-    esac
-done
-
-export DYN_REQUEST_PLANE=tcp
-
-if [[ "$SINGLE_GPU" == "true" ]]; then
-    GPU_LABEL="1 GPU"
-    PREFILL_GPU="${DYN_PREFILL_WORKER_GPU:-${CUDA_VISIBLE_DEVICES:-0}}"
-    DECODE_GPU="${DYN_DECODE_WORKER_GPU:-${CUDA_VISIBLE_DEVICES:-0}}"
-    MAX_MODEL_LEN="${MAX_MODEL_LEN:-4096}"
-    PD_KV_CACHE_BYTES=$((512 * 1024 * 1024))
-    SHARED_GPU_FRACTION=$(build_gpu_mem_args vllm --workers-per-gpu 2)
-    PREFILL_GPU_MEM="${DYN_PREFILL_GPU_MEM:-${SHARED_GPU_FRACTION:-0.45}}"
-    DECODE_GPU_MEM="${DYN_DECODE_GPU_MEM:-${SHARED_GPU_FRACTION:-0.45}}"
-    SHARED_ARGS=(
-        --enforce-eager
-        --max-model-len "$MAX_MODEL_LEN"
-        --kv-cache-memory-bytes "$PD_KV_CACHE_BYTES"
-        --limit-mm-per-prompt '{"image":1,"video":1,"audio":0}'
-    )
-else
-    GPU_LABEL="2 GPUs"
-    PREFILL_GPU="${DYN_PREFILL_WORKER_GPU:-0}"
-    DECODE_GPU="${DYN_DECODE_WORKER_GPU:-1}"
-    MAX_MODEL_LEN="${MAX_MODEL_LEN:-8192}"
-    GPU_MEM_ARGS=$(build_gpu_mem_args vllm)
-    PREFILL_GPU_MEM="${DYN_PREFILL_GPU_MEM:-${GPU_MEM_ARGS:-0.9}}"
-    DECODE_GPU_MEM="${DYN_DECODE_GPU_MEM:-${GPU_MEM_ARGS:-0.9}}"
-    SHARED_ARGS=(--max-model-len "$MAX_MODEL_LEN")
-fi
-
-print_launch_banner --no-curl "Launching Disaggregated Video Serving ($GPU_LABEL)" "$MODEL_NAME" "$HTTP_PORT" \
-    "Backend:     Prefill + decode workers via dynamo.vllm" \
-    "Video path:  Standard TokensPrompt multi_modal_data flow"
-
-print_curl_footer <<CURL
-  curl http://localhost:${HTTP_PORT}/v1/chat/completions \\
-    -H 'Content-Type: application/json' \\
-    -d '{
-      "model": "${MODEL_NAME}",
-      "messages": [{"role": "user", "content": [
-        {"type": "text", "text": "Describe the video in detail"},
-        {"type": "video_url", "video_url": {"url": "https://storage.googleapis.com/gtv-videos-bucket/sample/BigBuckBunny.mp4"}}
-      ]}],
-      "max_tokens": 128
-    }'
-CURL
-
-python -m dynamo.frontend &
-
-VLLM_NIXL_SIDE_CHANNEL_PORT=20098 \
-CUDA_VISIBLE_DEVICES="$PREFILL_GPU" \
-    python -m dynamo.vllm \
-        --disaggregation-mode prefill \
-        --enable-multimodal \
-        --model "$MODEL_NAME" \
-        --gpu-memory-utilization "$PREFILL_GPU_MEM" \
-        "${SHARED_ARGS[@]}" \
-        --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both"}' \
-        --kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:20081"}' \
-        "${EXTRA_ARGS[@]}" &
-
-VLLM_NIXL_SIDE_CHANNEL_PORT=20099 \
-CUDA_VISIBLE_DEVICES="$DECODE_GPU" \
-    python -m dynamo.vllm \
-        --disaggregation-mode decode \
-        --enable-multimodal \
-        --model "$MODEL_NAME" \
-        --gpu-memory-utilization "$DECODE_GPU_MEM" \
-        "${SHARED_ARGS[@]}" \
-        --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both"}' \
-        --kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:20082"}' \
-        "${EXTRA_ARGS[@]}" &
-
-wait_any_exit

tests/serve/test_vllm.py

@@ -428,14 +428,6 @@ vllm_configs = {
         model="Qwen/Qwen3-VL-2B-Instruct",
         script_args=["--model", "Qwen/Qwen3-VL-2B-Instruct", "--single-gpu"],
         timeout=300,
-        env={
-            "DYN_ENCODE_WORKER_GPU": "0",
-            "DYN_PREFILL_WORKER_GPU": "0",
-            "DYN_DECODE_WORKER_GPU": "0",
-            "DYN_ENCODE_GPU_MEM": "0.1",
-            "DYN_PREFILL_GPU_MEM": "0.4",
-            "DYN_DECODE_GPU_MEM": "0.4",
-        },
         request_payloads=[
             chat_payload(
                 [
@@ -536,11 +528,11 @@ vllm_configs = {
             ),
         ],
     ),
-    # Video multimodal tests for CI using the vLLM video launch scripts.
+    # Video multimodal tests for CI use the canonical aggregated multimodal launcher.
     "multimodal_video_agg": VLLMConfig(
         name="multimodal_video_agg",
         directory=vllm_dir,
-        script_name="video_agg.sh",
+        script_name="agg_multimodal.sh",
         marks=[
             pytest.mark.gpu_1,
             pytest.mark.pre_merge,
@@ -568,7 +560,7 @@ vllm_configs = {
     "multimodal_video_disagg": VLLMConfig(
         name="multimodal_video_disagg",
         directory=vllm_dir,
-        script_name="video_disagg.sh",
+        script_name="disagg_multimodal_epd.sh",
         marks=[
             pytest.mark.gpu_1,
             pytest.mark.pre_merge,