diff --git a/components/backends/trtllm/launch/agg.sh b/components/backends/trtllm/launch/agg.sh
index 5c7021c59c..5070b10f45 100755
--- a/components/backends/trtllm/launch/agg.sh
+++ b/components/backends/trtllm/launch/agg.sh
@@ -9,6 +9,7 @@ export AGG_ENGINE_ARGS=${AGG_ENGINE_ARGS:-"engine_configs/agg.yaml"}
 export MODALITY=${MODALITY:-"text"}
 # If you want to use multimodal, set MODALITY to "multimodal"
 #export MODALITY=${MODALITY:-"multimodal"}
+export MPI_CMD=${MPI_CMD:-""}
 
 # Setup cleanup trap
 cleanup() {
@@ -21,11 +22,11 @@ trap cleanup EXIT INT TERM
 
 
 # run frontend
-python3 -m dynamo.frontend --http-port 8000 &
+$MPI_CMD python3 -m dynamo.frontend --http-port 8000 &
 DYNAMO_PID=$!
 
 # run worker
-python3 -m dynamo.trtllm \
+$MPI_CMD python3 -m dynamo.trtllm \
   --model-path "$MODEL_PATH" \
   --served-model-name "$SERVED_MODEL_NAME" \
   --modality "$MODALITY" \
diff --git a/components/backends/trtllm/launch/agg_metrics.sh b/components/backends/trtllm/launch/agg_metrics.sh
index 3232576d76..212a012c41 100755
--- a/components/backends/trtllm/launch/agg_metrics.sh
+++ b/components/backends/trtllm/launch/agg_metrics.sh
@@ -7,6 +7,7 @@ export MODEL_PATH=${MODEL_PATH:-"Qwen/Qwen3-0.6B"}
 export SERVED_MODEL_NAME=${SERVED_MODEL_NAME:-"Qwen/Qwen3-0.6B"}
 export AGG_ENGINE_ARGS=${AGG_ENGINE_ARGS:-"engine_configs/agg.yaml"}
 export MODALITY=${MODALITY:-"text"}
+export MPI_CMD=${MPI_CMD:-""}
 
 # Setup cleanup trap
 cleanup() {
@@ -18,12 +19,12 @@ cleanup() {
 trap cleanup EXIT INT TERM
 
 # Run frontend
-python3 -m dynamo.frontend --http-port 8000 &
+$MPI_CMD python3 -m dynamo.frontend --http-port 8000 &
 DYNAMO_PID=$!
 
 # Run worker
 DYN_SYSTEM_ENABLED=true DYN_SYSTEM_PORT=8081 \
-python3 -m dynamo.trtllm \
+$MPI_CMD python3 -m dynamo.trtllm \
   --model-path "$MODEL_PATH" \
   --served-model-name "$SERVED_MODEL_NAME" \
   --modality "$MODALITY" \
diff --git a/components/backends/trtllm/launch/agg_router.sh b/components/backends/trtllm/launch/agg_router.sh
index ca6d439e63..907e6d4a8f 100755
--- a/components/backends/trtllm/launch/agg_router.sh
+++ b/components/backends/trtllm/launch/agg_router.sh
@@ -6,6 +6,7 @@
 export MODEL_PATH=${MODEL_PATH:-"Qwen/Qwen3-0.6B"}
 export SERVED_MODEL_NAME=${SERVED_MODEL_NAME:-"Qwen/Qwen3-0.6B"}
 export AGG_ENGINE_ARGS=${AGG_ENGINE_ARGS:-"engine_configs/agg.yaml"}
+export MPI_CMD=${MPI_CMD:-""}
 
 # Setup cleanup trap
 cleanup() {
@@ -18,11 +19,11 @@ trap cleanup EXIT INT TERM
 
 
 # run frontend
-python3 -m dynamo.frontend --router-mode kv --http-port 8000 &
+$MPI_CMD python3 -m dynamo.frontend --router-mode kv --http-port 8000 &
 DYNAMO_PID=$!
 
 # run worker
-python3 -m dynamo.trtllm \
+$MPI_CMD python3 -m dynamo.trtllm \
   --model-path "$MODEL_PATH" \
   --served-model-name "$SERVED_MODEL_NAME" \
   --extra-engine-args "$AGG_ENGINE_ARGS" \
diff --git a/components/backends/trtllm/launch/disagg.sh b/components/backends/trtllm/launch/disagg.sh
index f89eba5c9e..bedb685ffe 100755
--- a/components/backends/trtllm/launch/disagg.sh
+++ b/components/backends/trtllm/launch/disagg.sh
@@ -13,6 +13,7 @@ export DECODE_CUDA_VISIBLE_DEVICES=${DECODE_CUDA_VISIBLE_DEVICES:-"1"}
 export MODALITY=${MODALITY:-"text"}
 # If you want to use multimodal, set MODALITY to "multimodal"
 #export MODALITY=${MODALITY:-"multimodal"}
+export MPI_CMD=${MPI_CMD:-""}
 
 # Setup cleanup trap
 cleanup() {
@@ -25,11 +26,11 @@ trap cleanup EXIT INT TERM
 
 
 # run frontend
-python3 -m dynamo.frontend --http-port 8000 &
+$MPI_CMD python3 -m dynamo.frontend --http-port 8000 &
 DYNAMO_PID=$!
 
 # run prefill worker
-CUDA_VISIBLE_DEVICES=$PREFILL_CUDA_VISIBLE_DEVICES python3 -m dynamo.trtllm \
+CUDA_VISIBLE_DEVICES=$PREFILL_CUDA_VISIBLE_DEVICES $MPI_CMD python3 -m dynamo.trtllm \
   --model-path "$MODEL_PATH" \
   --served-model-name "$SERVED_MODEL_NAME" \
   --extra-engine-args  "$PREFILL_ENGINE_ARGS" \
@@ -39,7 +40,7 @@ CUDA_VISIBLE_DEVICES=$PREFILL_CUDA_VISIBLE_DEVICES python3 -m dynamo.trtllm \
 PREFILL_PID=$!
 
 # run decode worker
-CUDA_VISIBLE_DEVICES=$DECODE_CUDA_VISIBLE_DEVICES python3 -m dynamo.trtllm \
+CUDA_VISIBLE_DEVICES=$DECODE_CUDA_VISIBLE_DEVICES $MPI_CMD python3 -m dynamo.trtllm \
   --model-path "$MODEL_PATH" \
   --served-model-name "$SERVED_MODEL_NAME" \
   --extra-engine-args  "$DECODE_ENGINE_ARGS" \
diff --git a/components/backends/trtllm/launch/disagg_router.sh b/components/backends/trtllm/launch/disagg_router.sh
index e29c851a56..bef13e8ab3 100755
--- a/components/backends/trtllm/launch/disagg_router.sh
+++ b/components/backends/trtllm/launch/disagg_router.sh
@@ -10,6 +10,7 @@ export PREFILL_ENGINE_ARGS=${PREFILL_ENGINE_ARGS:-"engine_configs/prefill.yaml"}
 export DECODE_ENGINE_ARGS=${DECODE_ENGINE_ARGS:-"engine_configs/decode.yaml"}
 export PREFILL_CUDA_VISIBLE_DEVICES=${PREFILL_CUDA_VISIBLE_DEVICES:-"0"}
 export DECODE_CUDA_VISIBLE_DEVICES=${DECODE_CUDA_VISIBLE_DEVICES:-"1"}
+export MPI_CMD=${MPI_CMD:-""}
 
 # Setup cleanup trap
 cleanup() {
@@ -22,7 +23,7 @@ trap cleanup EXIT INT TERM
 
 
 # run frontend
-python3 -m dynamo.frontend --router-mode kv --http-port 8000 &
+$MPI_CMD python3 -m dynamo.frontend --router-mode kv --http-port 8000 &
 DYNAMO_PID=$!
 
 
@@ -35,7 +36,7 @@ else
 fi
 
 # run prefill worker
-CUDA_VISIBLE_DEVICES=$PREFILL_CUDA_VISIBLE_DEVICES python3 -m dynamo.trtllm \
+CUDA_VISIBLE_DEVICES=$PREFILL_CUDA_VISIBLE_DEVICES $MPI_CMD python3 -m dynamo.trtllm \
   --model-path "$MODEL_PATH" \
   --served-model-name "$SERVED_MODEL_NAME" \
   --extra-engine-args "$PREFILL_ENGINE_ARGS" \
@@ -45,7 +46,7 @@ CUDA_VISIBLE_DEVICES=$PREFILL_CUDA_VISIBLE_DEVICES python3 -m dynamo.trtllm \
 PREFILL_PID=$!
 
 # run decode worker
-CUDA_VISIBLE_DEVICES=$DECODE_CUDA_VISIBLE_DEVICES python3 -m dynamo.trtllm \
+CUDA_VISIBLE_DEVICES=$DECODE_CUDA_VISIBLE_DEVICES $MPI_CMD python3 -m dynamo.trtllm \
   --model-path "$MODEL_PATH" \
   --served-model-name "$SERVED_MODEL_NAME" \
   --extra-engine-args "$DECODE_ENGINE_ARGS" \
diff --git a/components/backends/trtllm/launch/epd_disagg.sh b/components/backends/trtllm/launch/epd_disagg.sh
index 60cfa1c249..40805c9413 100755
--- a/components/backends/trtllm/launch/epd_disagg.sh
+++ b/components/backends/trtllm/launch/epd_disagg.sh
@@ -16,6 +16,7 @@ export ENCODE_ENDPOINT=${ENCODE_ENDPOINT:-"dyn://dynamo.tensorrt_llm_encode.gene
 export MODALITY=${MODALITY:-"multimodal"}
 export ALLOWED_LOCAL_MEDIA_PATH=${ALLOWED_LOCAL_MEDIA_PATH:-"/tmp"}
 export MAX_FILE_SIZE_MB=${MAX_FILE_SIZE_MB:-50}
+export MPI_CMD=${MPI_CMD:-""}
 
 # Setup cleanup trap
 cleanup() {
@@ -28,11 +29,11 @@ trap cleanup EXIT INT TERM
 
 
 # run frontend
-python3 -m dynamo.frontend --http-port 8000 &
+$MPI_CMD python3 -m dynamo.frontend --http-port 8000 &
 DYNAMO_PID=$!
 
 # run encode worker
-CUDA_VISIBLE_DEVICES=$ENCODE_CUDA_VISIBLE_DEVICES python3 -m dynamo.trtllm \
+CUDA_VISIBLE_DEVICES=$ENCODE_CUDA_VISIBLE_DEVICES $MPI_CMD python3 -m dynamo.trtllm \
   --model-path "$MODEL_PATH" \
   --served-model-name "$SERVED_MODEL_NAME" \
   --extra-engine-args "$ENCODE_ENGINE_ARGS" \
@@ -44,7 +45,7 @@ CUDA_VISIBLE_DEVICES=$ENCODE_CUDA_VISIBLE_DEVICES python3 -m dynamo.trtllm \
 ENCODE_PID=$!
 
 # run prefill worker
-CUDA_VISIBLE_DEVICES=$PREFILL_CUDA_VISIBLE_DEVICES python3 -m dynamo.trtllm \
+CUDA_VISIBLE_DEVICES=$PREFILL_CUDA_VISIBLE_DEVICES $MPI_CMD python3 -m dynamo.trtllm \
   --model-path "$MODEL_PATH" \
   --served-model-name "$SERVED_MODEL_NAME" \
   --extra-engine-args "$PREFILL_ENGINE_ARGS" \
@@ -55,7 +56,7 @@ CUDA_VISIBLE_DEVICES=$PREFILL_CUDA_VISIBLE_DEVICES python3 -m dynamo.trtllm \
 PREFILL_PID=$!
 
 # run decode worker
-CUDA_VISIBLE_DEVICES=$DECODE_CUDA_VISIBLE_DEVICES python3 -m dynamo.trtllm \
+CUDA_VISIBLE_DEVICES=$DECODE_CUDA_VISIBLE_DEVICES $MPI_CMD python3 -m dynamo.trtllm \
   --model-path "$MODEL_PATH" \
   --served-model-name "$SERVED_MODEL_NAME" \
   --extra-engine-args "$DECODE_ENGINE_ARGS" \
diff --git a/components/backends/trtllm/launch/gpt_oss_disagg.sh b/components/backends/trtllm/launch/gpt_oss_disagg.sh
index db42c01771..67acedd6a8 100755
--- a/components/backends/trtllm/launch/gpt_oss_disagg.sh
+++ b/components/backends/trtllm/launch/gpt_oss_disagg.sh
@@ -8,17 +8,18 @@ export SERVED_MODEL_NAME=${SERVED_MODEL_NAME:-"openai/gpt-oss-120b"}
 export DISAGGREGATION_STRATEGY=${DISAGGREGATION_STRATEGY:-"prefill_first"}
 export PREFILL_ENGINE_ARGS=${PREFILL_ENGINE_ARGS:-"engine_configs/gpt_oss/prefill.yaml"}
 export DECODE_ENGINE_ARGS=${DECODE_ENGINE_ARGS:-"engine_configs/gpt_oss/decode.yaml"}
+export MPI_CMD=${MPI_CMD:-""}
 
 set -e
 trap 'echo Cleaning up...; kill 0' EXIT
 
 
 # run frontend
-python3 -m dynamo.frontend --router-mode round-robin --http-port 8000 &
+$MPI_CMD python3 -m dynamo.frontend --router-mode round-robin --http-port 8000 &
 
 # With tensor_parallel_size=4, each worker needs 4 GPUs
 # run prefill worker
-CUDA_VISIBLE_DEVICES=0,1,2,3 python3 -m dynamo.trtllm \
+CUDA_VISIBLE_DEVICES=0,1,2,3 $MPI_CMD python3 -m dynamo.trtllm \
   --model-path "$MODEL_PATH" \
   --served-model-name "$SERVED_MODEL_NAME" \
   --extra-engine-args "$PREFILL_ENGINE_ARGS" \
@@ -33,7 +34,7 @@ CUDA_VISIBLE_DEVICES=0,1,2,3 python3 -m dynamo.trtllm \
   --expert-parallel-size 4 &
 
 # run decode worker
-CUDA_VISIBLE_DEVICES=4,5,6,7 python3 -m dynamo.trtllm \
+CUDA_VISIBLE_DEVICES=4,5,6,7 $MPI_CMD python3 -m dynamo.trtllm \
   --model-path "$MODEL_PATH" \
   --served-model-name "$SERVED_MODEL_NAME" \
   --extra-engine-args "$DECODE_ENGINE_ARGS" \