Fix port allocation issues and rename to DYN_HTTP_PORT

keivenchang · keivenchang · commit 4463ec7c92d0 · 2025-11-04T21:38:46.000-08:00
- Rename DYN_FRONTEND_PORT to DYN_HTTP_PORT to match actual frontend env var
- Export DYN_HTTP_PORT in all cancellation tests so workers know frontend port
- Fix TensorRT-LLM disagg_same_gpu.sh port conflict (prefill used PORT2, now uses PORT1)

Signed-off-by: Keiven Chang &lt;keivenchang@users.noreply.github.com&gt;
diff --git a/examples/backends/sglang/launch/agg.sh b/examples/backends/sglang/launch/agg.sh
@@ -13,7 +13,7 @@ trap cleanup EXIT INT TERM
 
 
 # run ingress
-python3 -m dynamo.frontend --http-port=${DYN_FRONTEND_PORT:-8000} &
+python3 -m dynamo.frontend --http-port=${DYN_HTTP_PORT:-8000} &
 DYNAMO_PID=$!
 
 # run worker with metrics enabled
diff --git a/examples/backends/sglang/launch/agg_embed.sh b/examples/backends/sglang/launch/agg_embed.sh
@@ -13,7 +13,7 @@ trap cleanup EXIT INT TERM
 
 
 # run ingress
-python3 -m dynamo.frontend --http-port=${DYN_FRONTEND_PORT:-8000} &
+python3 -m dynamo.frontend --http-port=${DYN_HTTP_PORT:-8000} &
 DYNAMO_PID=$!
 
 # run worker
diff --git a/examples/backends/sglang/launch/agg_router.sh b/examples/backends/sglang/launch/agg_router.sh
@@ -13,7 +13,7 @@ trap cleanup EXIT INT TERM
 
 
 # run ingress
-python -m dynamo.frontend --router-mode kv --http-port=${DYN_FRONTEND_PORT:-8000} &
+python -m dynamo.frontend --router-mode kv --http-port=${DYN_HTTP_PORT:-8000} &
 DYNAMO_PID=$!
 
 # run worker
diff --git a/examples/backends/sglang/launch/disagg.sh b/examples/backends/sglang/launch/disagg.sh
@@ -13,7 +13,7 @@ trap cleanup EXIT INT TERM
 
 
 # run ingress
-python3 -m dynamo.frontend --http-port=${DYN_FRONTEND_PORT:-8000} &
+python3 -m dynamo.frontend --http-port=${DYN_HTTP_PORT:-8000} &
 DYNAMO_PID=$!
 
 # run prefill worker
diff --git a/examples/backends/sglang/launch/disagg_dp_attn.sh b/examples/backends/sglang/launch/disagg_dp_attn.sh
@@ -13,7 +13,7 @@ trap cleanup EXIT INT TERM
 
 
 # run ingress
-python3 -m dynamo.frontend --http-port=${DYN_FRONTEND_PORT:-8000} &
+python3 -m dynamo.frontend --http-port=${DYN_HTTP_PORT:-8000} &
 DYNAMO_PID=$!
 
 # run prefill worker
diff --git a/examples/backends/sglang/launch/disagg_router.sh b/examples/backends/sglang/launch/disagg_router.sh
@@ -13,7 +13,7 @@ trap cleanup EXIT INT TERM
 
 # run ingress
 python3 -m dynamo.frontend \
- --http-port=${DYN_FRONTEND_PORT:-8000} \
+ --http-port=${DYN_HTTP_PORT:-8000} \
  --router-mode kv \
  --kv-overlap-score-weight 0 \
  --router-reset-states &
diff --git a/examples/backends/sglang/launch/disagg_same_gpu.sh b/examples/backends/sglang/launch/disagg_same_gpu.sh
@@ -37,7 +37,7 @@ trap cleanup EXIT INT TERM
 
 
 # run ingress with KV router mode for disaggregated setup
-python3 -m dynamo.frontend --router-mode kv --http-port=${DYN_FRONTEND_PORT:-8000} &
+python3 -m dynamo.frontend --router-mode kv --http-port=${DYN_HTTP_PORT:-8000} &
 DYNAMO_PID=$!
 
 # run prefill worker with metrics on port 8081
diff --git a/examples/backends/sglang/launch/multimodal_agg.sh b/examples/backends/sglang/launch/multimodal_agg.sh
@@ -60,7 +60,7 @@ if [[ -n "$SERVED_MODEL_NAME" ]]; then
 fi
 
 # run ingress
-python3 -m dynamo.frontend --http-port=${DYN_FRONTEND_PORT:-8000} &
+python3 -m dynamo.frontend --http-port=${DYN_HTTP_PORT:-8000} &
 DYNAMO_PID=$!
 
 # run SGLang multimodal processor
diff --git a/examples/backends/sglang/launch/multimodal_disagg.sh b/examples/backends/sglang/launch/multimodal_disagg.sh
@@ -60,7 +60,7 @@ if [[ -n "$SERVED_MODEL_NAME" ]]; then
 fi
 
 # run ingress
-python3 -m dynamo.frontend --http-port=${DYN_FRONTEND_PORT:-8000} &
+python3 -m dynamo.frontend --http-port=${DYN_HTTP_PORT:-8000} &
 DYNAMO_PID=$!
 
 # run SGLang multimodal processor
diff --git a/examples/backends/trtllm/launch/agg.sh b/examples/backends/trtllm/launch/agg.sh
@@ -22,7 +22,7 @@ trap cleanup EXIT INT TERM
 
 
 # run frontend
-python3 -m dynamo.frontend --http-port ${DYN_FRONTEND_PORT:-8000} &
+python3 -m dynamo.frontend --http-port ${DYN_HTTP_PORT:-8000} &
 DYNAMO_PID=$!
 
 # run worker
diff --git a/examples/backends/trtllm/launch/agg_metrics.sh b/examples/backends/trtllm/launch/agg_metrics.sh
@@ -19,7 +19,7 @@ cleanup() {
 trap cleanup EXIT INT TERM
 
 # Run frontend
-python3 -m dynamo.frontend --http-port ${DYN_FRONTEND_PORT:-8000} &
+python3 -m dynamo.frontend --http-port ${DYN_HTTP_PORT:-8000} &
 DYNAMO_PID=$!
 
 # Run worker
diff --git a/examples/backends/trtllm/launch/agg_router.sh b/examples/backends/trtllm/launch/agg_router.sh
@@ -19,7 +19,7 @@ trap cleanup EXIT INT TERM
 
 
 # run frontend
-python3 -m dynamo.frontend --router-mode kv --http-port ${DYN_FRONTEND_PORT:-8000} &
+python3 -m dynamo.frontend --router-mode kv --http-port ${DYN_HTTP_PORT:-8000} &
 DYNAMO_PID=$!
 
 # run worker
diff --git a/examples/backends/trtllm/launch/disagg.sh b/examples/backends/trtllm/launch/disagg.sh
@@ -26,7 +26,7 @@ trap cleanup EXIT INT TERM
 
 
 # run frontend
-python3 -m dynamo.frontend --http-port ${DYN_FRONTEND_PORT:-8000} &
+python3 -m dynamo.frontend --http-port ${DYN_HTTP_PORT:-8000} &
 DYNAMO_PID=$!
 
 # run prefill worker
diff --git a/examples/backends/trtllm/launch/disagg_router.sh b/examples/backends/trtllm/launch/disagg_router.sh
@@ -23,7 +23,7 @@ trap cleanup EXIT INT TERM
 
 
 # run frontend
-python3 -m dynamo.frontend --router-mode kv --http-port ${DYN_FRONTEND_PORT:-8000} &
+python3 -m dynamo.frontend --router-mode kv --http-port ${DYN_HTTP_PORT:-8000} &
 DYNAMO_PID=$!
 
 
diff --git a/examples/backends/trtllm/launch/disagg_same_gpu.sh b/examples/backends/trtllm/launch/disagg_same_gpu.sh
@@ -49,12 +49,12 @@ trap cleanup EXIT INT TERM
 
 
 # run frontend
-python3 -m dynamo.frontend --http-port ${DYN_FRONTEND_PORT:-8000} &
+python3 -m dynamo.frontend --http-port ${DYN_HTTP_PORT:-8000} &
 DYNAMO_PID=$!
 
 # run prefill worker (shares GPU with decode)
 CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES \
-DYN_SYSTEM_ENABLED=true DYN_SYSTEM_PORT=${DYN_SYSTEM_PORT2:-8082} \
+DYN_SYSTEM_ENABLED=true DYN_SYSTEM_PORT=${DYN_SYSTEM_PORT1:-8081} \
 python3 -m dynamo.trtllm \
   --model-path "$MODEL_PATH" \
   --served-model-name "$SERVED_MODEL_NAME" \
diff --git a/examples/backends/trtllm/launch/epd_disagg.sh b/examples/backends/trtllm/launch/epd_disagg.sh
@@ -29,7 +29,7 @@ trap cleanup EXIT INT TERM
 
 
 # run frontend
-python3 -m dynamo.frontend --http-port ${DYN_FRONTEND_PORT:-8000} &
+python3 -m dynamo.frontend --http-port ${DYN_HTTP_PORT:-8000} &
 DYNAMO_PID=$!
 
 # run encode worker
diff --git a/examples/backends/trtllm/launch/gpt_oss_disagg.sh b/examples/backends/trtllm/launch/gpt_oss_disagg.sh
@@ -15,7 +15,7 @@ trap 'echo Cleaning up...; kill 0' EXIT
 
 
 # run frontend
-python3 -m dynamo.frontend --router-mode round-robin --http-port ${DYN_FRONTEND_PORT:-8000} &
+python3 -m dynamo.frontend --router-mode round-robin --http-port ${DYN_HTTP_PORT:-8000} &
 
 # With tensor_parallel_size=4, each worker needs 4 GPUs
 # run prefill worker
diff --git a/examples/backends/trtllm/performance_sweeps/scripts/start_frontend.sh b/examples/backends/trtllm/performance_sweeps/scripts/start_frontend.sh
@@ -21,5 +21,5 @@ sleep 2
 
 # Start OpenAI Frontend which will dynamically discover workers when they startup
 # NOTE: This is a blocking call.
-python3 -m dynamo.frontend --http-port ${DYN_FRONTEND_PORT:-8000}
+python3 -m dynamo.frontend --http-port ${DYN_HTTP_PORT:-8000}
 
diff --git a/examples/backends/vllm/launch/agg.sh b/examples/backends/vllm/launch/agg.sh
@@ -5,7 +5,7 @@ set -e
 trap 'echo Cleaning up...; kill 0' EXIT
 
 # run ingress
-python -m dynamo.frontend --http-port=${DYN_FRONTEND_PORT:-8000} &
+python -m dynamo.frontend --http-port=${DYN_HTTP_PORT:-8000} &
 
 # run worker
 # --enforce-eager is added for quick deployment. for production use, need to remove this flag
diff --git a/examples/backends/vllm/launch/agg_kvbm.sh b/examples/backends/vllm/launch/agg_kvbm.sh
@@ -5,7 +5,7 @@ set -e
 trap 'echo Cleaning up...; kill 0' EXIT
 
 # run ingress
-python -m dynamo.frontend --http-port=${DYN_FRONTEND_PORT:-8000} &
+python -m dynamo.frontend --http-port=${DYN_HTTP_PORT:-8000} &
 
 # run worker with KVBM enabled
 # NOTE: remove --enforce-eager for production use
diff --git a/examples/backends/vllm/launch/agg_kvbm_router.sh b/examples/backends/vllm/launch/agg_kvbm_router.sh
@@ -13,7 +13,7 @@ MODEL="Qwen/Qwen3-0.6B"
 # run frontend + KV router
 python -m dynamo.frontend \
     --router-mode kv \
-    --http-port ${DYN_FRONTEND_PORT:-8000} \
+    --http-port ${DYN_HTTP_PORT:-8000} \
     --router-reset-states &
 
 # run workers with KVBM enabled
diff --git a/examples/backends/vllm/launch/agg_lmcache.sh b/examples/backends/vllm/launch/agg_lmcache.sh
@@ -5,7 +5,7 @@ set -e
 trap 'echo Cleaning up...; kill 0' EXIT
 
 # run ingress
-python -m dynamo.frontend --http-port=${DYN_FRONTEND_PORT:-8000} &
+python -m dynamo.frontend --http-port=${DYN_HTTP_PORT:-8000} &
 
 # run worker with LMCache enabled
 ENABLE_LMCACHE=1 \
diff --git a/examples/backends/vllm/launch/agg_multimodal.sh b/examples/backends/vllm/launch/agg_multimodal.sh
@@ -53,7 +53,7 @@ else
 fi
 
 # run ingress
-python -m dynamo.frontend --http-port=${DYN_FRONTEND_PORT:-8000} &
+python -m dynamo.frontend --http-port=${DYN_HTTP_PORT:-8000} &
 
 # To make Qwen2.5-VL fit in A100 40GB, set the following extra arguments
 EXTRA_ARGS=""
diff --git a/examples/backends/vllm/launch/agg_multimodal_llama.sh b/examples/backends/vllm/launch/agg_multimodal_llama.sh
@@ -8,7 +8,7 @@ trap 'echo Cleaning up...; kill 0' EXIT
 MODEL_NAME="meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8"
 
 # run ingress
-python -m dynamo.frontend --http-port=${DYN_FRONTEND_PORT:-8000} &
+python -m dynamo.frontend --http-port=${DYN_HTTP_PORT:-8000} &
 
 # run processor
 python -m dynamo.vllm --multimodal-processor --model $MODEL_NAME --mm-prompt-template "<|image|>\n<prompt>" &
diff --git a/examples/backends/vllm/launch/agg_router.sh b/examples/backends/vllm/launch/agg_router.sh
@@ -14,7 +14,7 @@ BLOCK_SIZE=64
 # run frontend + KV router
 python -m dynamo.frontend \
     --router-mode kv \
-    --http-port ${DYN_FRONTEND_PORT:-8000} \
+    --http-port ${DYN_HTTP_PORT:-8000} \
     --router-reset-states &
 
 # run workers
diff --git a/examples/backends/vllm/launch/dep.sh b/examples/backends/vllm/launch/dep.sh
@@ -5,7 +5,7 @@ set -e
 trap 'echo Cleaning up...; kill 0' EXIT
 
 # run ingress
-python -m dynamo.frontend --router-mode kv --http-port=${DYN_FRONTEND_PORT:-8000} &
+python -m dynamo.frontend --router-mode kv --http-port=${DYN_HTTP_PORT:-8000} &
 
 # Data Parallel Attention / Expert Parallelism
 # Routing to DP workers managed by Dynamo
diff --git a/examples/backends/vllm/launch/disagg.sh b/examples/backends/vllm/launch/disagg.sh
@@ -5,7 +5,7 @@ set -e
 trap 'echo Cleaning up...; kill 0' EXIT
 
 # run ingress
-python -m dynamo.frontend --http-port=${DYN_FRONTEND_PORT:-8000} &
+python -m dynamo.frontend --http-port=${DYN_HTTP_PORT:-8000} &
 
 # --enforce-eager is added for quick deployment. for production use, need to remove this flag
 CUDA_VISIBLE_DEVICES=0 python3 -m dynamo.vllm --model Qwen/Qwen3-0.6B --enforce-eager &
diff --git a/examples/backends/vllm/launch/disagg_kvbm.sh b/examples/backends/vllm/launch/disagg_kvbm.sh
@@ -5,7 +5,7 @@ set -e
 trap 'echo Cleaning up...; kill 0' EXIT
 
 # run ingress
-python -m dynamo.frontend --http-port=${DYN_FRONTEND_PORT:-8000} &
+python -m dynamo.frontend --http-port=${DYN_HTTP_PORT:-8000} &
 
 # run decode worker on GPU 0, without enabling KVBM
 # NOTE: remove --enforce-eager for production use
diff --git a/examples/backends/vllm/launch/disagg_kvbm_2p2d.sh b/examples/backends/vllm/launch/disagg_kvbm_2p2d.sh
@@ -5,7 +5,7 @@ set -e
 trap 'echo Cleaning up...; kill 0' EXIT
 
 # run ingress with KV router
-python -m dynamo.frontend --router-mode kv --http-port=${DYN_FRONTEND_PORT:-8000} &
+python -m dynamo.frontend --router-mode kv --http-port=${DYN_HTTP_PORT:-8000} &
 
 # run decode workers on GPU 0 and 1, without enabling KVBM
 # NOTE: remove --enforce-eager for production use
diff --git a/examples/backends/vllm/launch/disagg_kvbm_router.sh b/examples/backends/vllm/launch/disagg_kvbm_router.sh
@@ -13,7 +13,7 @@ MODEL="Qwen/Qwen3-0.6B"
 # run decode router with kv-overlap-score-weight 0 for pure load balancing
 python -m dynamo.frontend \
     --router-mode kv \
-    --http-port ${DYN_FRONTEND_PORT:-8000} \
+    --http-port ${DYN_HTTP_PORT:-8000} \
     --kv-overlap-score-weight 0 \
     --router-reset-states &
 
diff --git a/examples/backends/vllm/launch/disagg_lmcache.sh b/examples/backends/vllm/launch/disagg_lmcache.sh
@@ -5,7 +5,7 @@ set -e
 trap 'echo Cleaning up...; kill 0' EXIT
 
 # run ingress with KV router
-python -m dynamo.frontend --router-mode kv --http-port=${DYN_FRONTEND_PORT:-8000} &
+python -m dynamo.frontend --router-mode kv --http-port=${DYN_HTTP_PORT:-8000} &
 
 # run decode worker on GPU 0, without enabling LMCache
 CUDA_VISIBLE_DEVICES=0 python3 -m dynamo.vllm --model Qwen/Qwen3-0.6B &
diff --git a/examples/backends/vllm/launch/disagg_router.sh b/examples/backends/vllm/launch/disagg_router.sh
@@ -15,7 +15,7 @@ BLOCK_SIZE=64
 # The frontend will automatically detect prefill workers and activate an internal prefill router
 python -m dynamo.frontend \
     --router-mode kv \
-    --http-port ${DYN_FRONTEND_PORT:-8000} \
+    --http-port ${DYN_HTTP_PORT:-8000} \
     --router-reset-states &
 
 # two decode workers
diff --git a/examples/backends/vllm/launch/dsr1_dep.sh b/examples/backends/vllm/launch/dsr1_dep.sh
@@ -83,7 +83,7 @@ trap 'echo Cleaning up...; kill 0' EXIT
 
 # run ingress if it's node 0
 if [ $NODE_RANK -eq 0 ]; then
-    DYN_LOG=debug python -m dynamo.frontend --router-mode kv --http-port=${DYN_FRONTEND_PORT:-8000} 2>&1 | tee $LOG_DIR/dsr1_dep_ingress.log &
+    DYN_LOG=debug python -m dynamo.frontend --router-mode kv --http-port=${DYN_HTTP_PORT:-8000} 2>&1 | tee $LOG_DIR/dsr1_dep_ingress.log &
 fi
 
 mkdir -p $LOG_DIR
diff --git a/tests/fault_tolerance/cancellation/test_sglang.py b/tests/fault_tolerance/cancellation/test_sglang.py
@@ -88,6 +88,7 @@ def __init__(
         env["DYN_SYSTEM_ENABLED"] = "true"
         env["DYN_SYSTEM_USE_ENDPOINT_HEALTH_STATUS"] = '["generate"]'
         env["DYN_SYSTEM_PORT"] = str(system_port)
+        env["DYN_HTTP_PORT"] = str(frontend_port)
 
         # Set GPU assignment for disaggregated mode (like disagg.sh)
         if mode == "decode":
diff --git a/tests/fault_tolerance/cancellation/test_trtllm.py b/tests/fault_tolerance/cancellation/test_trtllm.py
@@ -90,6 +90,7 @@ def __init__(
         env["DYN_SYSTEM_ENABLED"] = "true"
         env["DYN_SYSTEM_USE_ENDPOINT_HEALTH_STATUS"] = '["generate"]'
         env["DYN_SYSTEM_PORT"] = str(system_port)
+        env["DYN_HTTP_PORT"] = str(frontend_port)
 
         # Set log directory based on worker type
         log_dir = f"{request.node.name}_{mode}_worker"
diff --git a/tests/fault_tolerance/cancellation/test_vllm.py b/tests/fault_tolerance/cancellation/test_vllm.py
@@ -68,6 +68,7 @@ def __init__(
         env["DYN_SYSTEM_ENABLED"] = "true"
         env["DYN_SYSTEM_USE_ENDPOINT_HEALTH_STATUS"] = '["generate"]'
         env["DYN_SYSTEM_PORT"] = str(system_port)
+        env["DYN_HTTP_PORT"] = str(frontend_port)
 
         # Set log directory based on worker type
         worker_type = "prefill_worker" if is_prefill else "worker"
diff --git a/tests/serve/launch/template_verifier.sh b/tests/serve/launch/template_verifier.sh
@@ -17,7 +17,7 @@ cleanup() {
 trap cleanup EXIT INT TERM
 
 # run ingress
-python3 -m dynamo.frontend --http-port=${DYN_FRONTEND_PORT:-8000} &
+python3 -m dynamo.frontend --http-port=${DYN_HTTP_PORT:-8000} &
 FRONTEND_PID=$!
 
 # run the mock worker + template validation generate()
diff --git a/tests/utils/port_utils.py b/tests/utils/port_utils.py
@@ -14,7 +14,7 @@
 For new test code in tests/, use the functions from this module:
 - get_free_port(): Find a single free port
 - get_free_ports(count): Find multiple free ports
-- set_frontend_port_env(config): Find free frontend port and set DYN_FRONTEND_PORT
+- set_frontend_port_env(config): Find free frontend port and set DYN_HTTP_PORT
 - set_worker_ports_env(config): Find free worker ports and set DYN_SYSTEM_PORT*
 - set_dynamic_ports_env(config): Find all ports and set all env vars (recommended)
 """
@@ -79,7 +79,7 @@ def get_free_ports(count: int) -> list[int]:
 def set_frontend_port_env(
     config: "EngineConfig", logger: logging.Logger | None = None
 ) -> dict[str, str]:
-    """Find a free frontend port and set DYN_FRONTEND_PORT environment variable.
+    """Find a free frontend port and set DYN_HTTP_PORT environment variable.
 
     This helper allocates a free port for the frontend, updates config.models_port
     and all payloads that use the default frontend port (8000).
@@ -93,7 +93,7 @@ def set_frontend_port_env(
         logger: Optional logger for logging allocated port
 
     Returns:
-        Dictionary with DYN_FRONTEND_PORT environment variable
+        Dictionary with DYN_HTTP_PORT environment variable
     """
     frontend_port = get_free_port()
 
@@ -110,7 +110,7 @@ def set_frontend_port_env(
     if logger:
         logger.info(f"Allocated frontend port: {frontend_port}")
 
-    return {"DYN_FRONTEND_PORT": str(frontend_port)}
+    return {"DYN_HTTP_PORT": str(frontend_port)}
 
 
 def set_worker_ports_env(
@@ -184,8 +184,8 @@ def set_dynamic_ports_env(
         logger: Optional logger for logging allocated ports
 
     Returns:
-        Dictionary of environment variables (e.g., {"DYN_FRONTEND_PORT": "8000",
-        "DYN_SYSTEM_PORT": "8081"} or {"DYN_FRONTEND_PORT": "8000",
+        Dictionary of environment variables (e.g., {"DYN_HTTP_PORT": "8000",
+        "DYN_SYSTEM_PORT": "8081"} or {"DYN_HTTP_PORT": "8000",
         "DYN_SYSTEM_PORT1": "8081", "DYN_SYSTEM_PORT2": "8082"})
     """
     extra_env = {}
