chore: migrate vllm e/p/d test from gpu_2 -> gpu_1 (#6638)

Signed-off-by: Krishnan Prashanth <kprashanth@nvidia.com>

examples/backends/vllm/launch/disagg_multimodal_epd.sh

@@ -7,6 +7,16 @@ trap 'echo Cleaning up...; kill 0' EXIT
 # Default values
 MODEL_NAME="llava-hf/llava-1.5-7b-hf"
 
+# --single-gpu: Packs all 3 workers (encode, prefill, decode) onto a single GPU.
+# This is intended for functional testing with small models (e.g. 2B) where CI
+# only has 1 GPU available. It reduces performance by:
+#   - 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)
+#   - Using lower gpu-memory-utilization fractions to share the GPU
+SINGLE_GPU=false
+
 # Parse command line arguments
 while [[ $# -gt 0 ]]; do
     case $1 in
@@ -14,6 +24,10 @@ while [[ $# -gt 0 ]]; do
             MODEL_NAME=$2
             shift 2
             ;;
+        --single-gpu)
+            SINGLE_GPU=true
+            shift
+            ;;
         -h|--help)
             echo "Usage: $0 [OPTIONS]"
             echo ""
@@ -22,12 +36,14 @@ while [[ $# -gt 0 ]]; do
             echo "Options:"
             echo "  --model <model_name>          Specify the VLM model to use (default: $MODEL_NAME)"
             echo "                                LLaVA 1.5 7B, Qwen2.5-VL, and Phi3V models have predefined templates"
+            echo "  --single-gpu                  Pack all 3 workers on 1 GPU (for small models, e.g. 2B)"
             echo "  -h, --help                    Show this help message"
             echo ""
             echo "Examples:"
             echo "  $0 --model llava-hf/llava-1.5-7b-hf"
             echo "  $0 --model microsoft/Phi-3.5-vision-instruct"
             echo "  $0 --model Qwen/Qwen2.5-VL-7B-Instruct"
+            echo "  $0 --model Qwen/Qwen3-VL-2B-Instruct --single-gpu"
             echo ""
             exit 0
             ;;
@@ -41,7 +57,7 @@ done
 
 
 echo "=================================================="
-echo "Disaggregated Multimodal Serving"
+echo "Disaggregated Multimodal Serving (E + P + D)"
 echo "=================================================="
 echo "Model: $MODEL_NAME"
 echo "=================================================="
@@ -53,6 +69,7 @@ echo "Starting frontend..."
 python -m dynamo.frontend &
 
 EXTRA_ARGS=""
+PD_EXTRA_ARGS=""
 
 # GPU assignments (override via environment variables)
 DYN_ENCODE_WORKER_GPU=${DYN_ENCODE_WORKER_GPU:-0}
@@ -64,6 +81,17 @@ DYN_ENCODE_GPU_MEM=${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}
 
+# 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
+# OOM during profiling when 3 workers share a GPU. 512 MB covers the
+# minimum vLLM requires for max_model_len=4096 on Qwen3-VL-2B.
+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}"
+fi
+
 # Start encode worker
 echo "Starting encode worker on GPU $DYN_ENCODE_WORKER_GPU (GPU mem: $DYN_ENCODE_GPU_MEM)..."
 VLLM_NIXL_SIDE_CHANNEL_PORT=20097 CUDA_VISIBLE_DEVICES=$DYN_ENCODE_WORKER_GPU python -m dynamo.vllm --multimodal-encode-worker --enable-multimodal --model $MODEL_NAME --gpu-memory-utilization $DYN_ENCODE_GPU_MEM $EXTRA_ARGS --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both"}' --kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:20080"}' &
@@ -71,12 +99,12 @@ VLLM_NIXL_SIDE_CHANNEL_PORT=20097 CUDA_VISIBLE_DEVICES=$DYN_ENCODE_WORKER_GPU py
 # Start prefill worker (also handles encode routing via --route-to-encoder)
 echo "Starting prefill worker on GPU $DYN_PREFILL_WORKER_GPU (GPU mem: $DYN_PREFILL_GPU_MEM)..."
 VLLM_NIXL_SIDE_CHANNEL_PORT=20098 \
-CUDA_VISIBLE_DEVICES=$DYN_PREFILL_WORKER_GPU python -m dynamo.vllm --multimodal-worker --route-to-encoder --disaggregation-mode prefill --enable-multimodal --enable-mm-embeds --model $MODEL_NAME --gpu-memory-utilization $DYN_PREFILL_GPU_MEM $EXTRA_ARGS --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both"}' --kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:20081"}' &
+CUDA_VISIBLE_DEVICES=$DYN_PREFILL_WORKER_GPU python -m dynamo.vllm --multimodal-worker --route-to-encoder --disaggregation-mode prefill --enable-multimodal --enable-mm-embeds --model $MODEL_NAME --gpu-memory-utilization $DYN_PREFILL_GPU_MEM $EXTRA_ARGS $PD_EXTRA_ARGS --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both"}' --kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:20081"}' &
 
 # Start decode worker
 echo "Starting decode worker on GPU $DYN_DECODE_WORKER_GPU (GPU mem: $DYN_DECODE_GPU_MEM)..."
 VLLM_NIXL_SIDE_CHANNEL_PORT=20099 \
-CUDA_VISIBLE_DEVICES=$DYN_DECODE_WORKER_GPU python -m dynamo.vllm --multimodal-decode-worker --enable-multimodal --enable-mm-embeds --model $MODEL_NAME --gpu-memory-utilization $DYN_DECODE_GPU_MEM $EXTRA_ARGS --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both"}' --kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:20082"}' &
+CUDA_VISIBLE_DEVICES=$DYN_DECODE_WORKER_GPU python -m dynamo.vllm --multimodal-decode-worker --enable-multimodal --enable-mm-embeds --model $MODEL_NAME --gpu-memory-utilization $DYN_DECODE_GPU_MEM $EXTRA_ARGS $PD_EXTRA_ARGS --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both"}' --kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:20082"}' &
 
 echo "=================================================="
 echo "All components started. Waiting for initialization..."
@@ -84,4 +112,3 @@ echo "=================================================="
 
 # Wait for all background processes to complete
 wait
-

tests/serve/test_vllm.py

@@ -276,6 +276,7 @@ vllm_configs = {
             completion_payload_default(),
         ],
     ),
+    # NOTE: Pack all workers on 1 GPU for lower CI resource requirements
     "multimodal_disagg_qwen3vl_2b_e_pd": VLLMConfig(
         name="multimodal_disagg_qwen3vl_2b_e_pd",
         directory=vllm_dir,
@@ -335,20 +336,22 @@ vllm_configs = {
             )
         ],
     ),
+    # NOTE: Pack all workers on 1 GPU for lower CI resource requirements
     "multimodal_disagg_qwen3vl_2b_epd": VLLMConfig(
         name="multimodal_disagg_qwen3vl_2b_epd",
         directory=vllm_dir,
         script_name="disagg_multimodal_epd.sh",
-        marks=[pytest.mark.gpu_2, pytest.mark.pre_merge],
+        marks=[pytest.mark.gpu_1, pytest.mark.pre_merge],
         model="Qwen/Qwen3-VL-2B-Instruct",
-        script_args=["--model", "Qwen/Qwen3-VL-2B-Instruct"],
+        script_args=["--model", "Qwen/Qwen3-VL-2B-Instruct", "--single-gpu"],
+        timeout=360,
         env={
             "DYN_ENCODE_WORKER_GPU": "0",
             "DYN_PREFILL_WORKER_GPU": "0",
-            "DYN_DECODE_WORKER_GPU": "1",
-            "DYN_ENCODE_GPU_MEM": "0.4",
+            "DYN_DECODE_WORKER_GPU": "0",
+            "DYN_ENCODE_GPU_MEM": "0.1",
             "DYN_PREFILL_GPU_MEM": "0.4",
-            "DYN_DECODE_GPU_MEM": "0.85",
+            "DYN_DECODE_GPU_MEM": "0.4",
         },
         request_payloads=[
             chat_payload(