diff --git a/examples/nixl/run_accuracy_test.sh b/examples/nixl/run_accuracy_test.sh
index ee999fd45..983be2b26 100755
--- a/examples/nixl/run_accuracy_test.sh
+++ b/examples/nixl/run_accuracy_test.sh
@@ -31,7 +31,7 @@ DECODER_TP_SIZE=${DECODER_TP_SIZE:-1}
 #GIT_ROOT=$(git rev-parse --show-toplevel)
 GIT_ROOT="/home/vllm-nixl/vllm"
 
-#SMI_BIN=$(which nvidia-smi || which rocm-smi)
+SMI_BIN=$(which nvidia-smi || which rocm-smi || which hl-smi)
 
 # Trap the SIGINT signal (triggered by Ctrl+C)
 trap 'kill $(jobs -pr)' SIGINT SIGTERM EXIT
@@ -67,6 +67,8 @@ get_model_args() {
 get_num_gpus() {
   if [[ "$SMI_BIN" == *"nvidia"* ]]; then
     echo "$($SMI_BIN --query-gpu=name --format=csv,noheader | wc -l)"
+  elif [[ "$SMI_BIN" == *"hl"* ]]; then
+    echo "$($SMI_BIN -Q index,name -f csv | tail -n +2 | wc -l)"
   else
     echo "$($SMI_BIN -l | grep GPU | wc -l)"
   fi
@@ -91,8 +93,7 @@ run_tests_for_model() {
   # Start prefill instances
   for i in $(seq 0 $((NUM_PREFILL_INSTANCES-1))); do
     # Calculate GPU ID - we'll distribute across available GPUs
-    #GPU_ID=$((i % $(get_num_gpus)))
-    GPU_ID=2
+    GPU_ID=$((i % $(get_num_gpus)))
 
     # Calculate port number (base port + instance number)
     PORT=$((8700 + i))
@@ -102,7 +103,7 @@ run_tests_for_model() {
     echo "Starting prefill instance $i on GPU $GPU_ID, port $PORT"
 
     # Build the command with or without model-specific args
-    BASE_CMD="RANK=0 UCX_TLS=tcp VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
+    BASE_CMD="HABANA_VISIBLE_MODULES='0,1,2,3' RANK=$GPU_ID UCX_TLS=tcp VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
     --port $PORT \
     --enforce-eager \
     --max_num_batched_tokens 8192 \
@@ -126,7 +127,7 @@ run_tests_for_model() {
   # Start decode instances
   for i in $(seq 0 $((NUM_DECODE_INSTANCES-1))); do
     # Calculate GPU ID - we'll distribute across available GPUs, starting from after prefill GPUs
-    #GPU_ID=$(((i + NUM_PREFILL_INSTANCES) % $(get_num_gpus)))
+    GPU_ID=$(((i + NUM_PREFILL_INSTANCES) % $(get_num_gpus)))
     # Calculate port number (base port + instance number)
     PORT=$((8800 + i))
     # Calculate side channel port
@@ -135,7 +136,7 @@ run_tests_for_model() {
     echo "Starting decode instance $i on GPU $GPU_ID, port $PORT"
 
     # Build the command with or without model-specific args
-    BASE_CMD="RANK=1 UCX_TLS=tcp VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
+    BASE_CMD="HABANA_VISIBLE_MODULES='4,5,6,7' RANK=$GPU_ID UCX_TLS=tcp VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
     --port $PORT \
     --enforce-eager \
     --max_num_batched_tokens 8192 \
diff --git a/examples/nixl/run_benchmark_profile.sh b/examples/nixl/run_benchmark_profile.sh
index 171d4b73b..8e79a1419 100644
--- a/examples/nixl/run_benchmark_profile.sh
+++ b/examples/nixl/run_benchmark_profile.sh
@@ -34,7 +34,7 @@ DECODER_TP_SIZE=${DECODER_TP_SIZE:-1}
 #GIT_ROOT=$(git rev-parse --show-toplevel)
 GIT_ROOT="/home/vllm-nixl/vllm"
 
-#SMI_BIN=$(which nvidia-smi || which rocm-smi)
+SMI_BIN=$(which nvidia-smi || which rocm-smi || which hl-smi)
 
 # Trap the SIGINT signal (triggered by Ctrl+C)
 trap 'kill $(jobs -pr)' SIGINT SIGTERM EXIT
@@ -70,6 +70,8 @@ get_model_args() {
 get_num_gpus() {
   if [[ "$SMI_BIN" == *"nvidia"* ]]; then
     echo "$($SMI_BIN --query-gpu=name --format=csv,noheader | wc -l)"
+  elif [[ "$SMI_BIN" == *"hl"* ]]; then
+    echo "$($SMI_BIN -Q index,name -f csv | tail -n +2 | wc -l)"
   else
     echo "$($SMI_BIN -l | grep GPU | wc -l)"
   fi
@@ -94,8 +96,7 @@ run_tests_for_model() {
   # Start prefill instances
   for i in $(seq 0 $((NUM_PREFILL_INSTANCES-1))); do
     # Calculate GPU ID - we'll distribute across available GPUs
-    #GPU_ID=$((i % $(get_num_gpus)))
-    GPU_ID=2
+    GPU_ID=$((i % $(get_num_gpus)))
 
     # Calculate port number (base port + instance number)
     PORT=$((8300))
@@ -105,7 +106,7 @@ run_tests_for_model() {
     echo "Starting prefill instance $i on GPU $GPU_ID, port $PORT"
 
     # Build the command with or without model-specific args
-    BASE_CMD="RANK=0 UCX_TLS=rc,ud,ib VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
+    BASE_CMD="HABANA_VISIBLE_MODULES='0,1,2,3' RANK=$GPU_ID UCX_TLS=rc,ud,ib VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
     --port $PORT \
     --long_prefill_token_threshold 8192 \
     --max_num_batched_tokens 8192 \
@@ -130,7 +131,7 @@ run_tests_for_model() {
   # Start decode instances
   for i in $(seq 0 $((NUM_DECODE_INSTANCES-1))); do
     # Calculate GPU ID - we'll distribute across available GPUs, starting from after prefill GPUs
-    #GPU_ID=$(((i + NUM_PREFILL_INSTANCES) % $(get_num_gpus)))
+    GPU_ID=$(((i + NUM_PREFILL_INSTANCES) % $(get_num_gpus)))
     # Calculate port number (base port + instance number)
     PORT=$((8400))
     # Calculate side channel port
@@ -139,7 +140,7 @@ run_tests_for_model() {
     echo "Starting decode instance $i on GPU $GPU_ID, port $PORT"
 
     # Build the command with or without model-specific args
-    BASE_CMD="RANK=1 UCX_TLS=rc,ud,ib VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
+    BASE_CMD="HABANA_VISIBLE_MODULES='4,5,6,7' RANK=$GPU_ID UCX_TLS=rc,ud,ib VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
     --port $PORT \
     --gpu-memory-utilization 0.3 \
     --tensor-parallel-size $DECODER_TP_SIZE \
diff --git a/examples/nixl/run_benchmark_test.sh b/examples/nixl/run_benchmark_test.sh
index 766e09ae4..97b826998 100755
--- a/examples/nixl/run_benchmark_test.sh
+++ b/examples/nixl/run_benchmark_test.sh
@@ -31,7 +31,7 @@ DECODER_TP_SIZE=${DECODER_TP_SIZE:-1}
 #GIT_ROOT=$(git rev-parse --show-toplevel)
 GIT_ROOT="/home/vllm-nixl/vllm"
 
-#SMI_BIN=$(which nvidia-smi || which rocm-smi)
+SMI_BIN=$(which nvidia-smi || which rocm-smi || which hl-smi)
 
 # Trap the SIGINT signal (triggered by Ctrl+C)
 trap 'kill $(jobs -pr)' SIGINT SIGTERM EXIT
@@ -67,6 +67,8 @@ get_model_args() {
 get_num_gpus() {
   if [[ "$SMI_BIN" == *"nvidia"* ]]; then
     echo "$($SMI_BIN --query-gpu=name --format=csv,noheader | wc -l)"
+  elif [[ "$SMI_BIN" == *"hl"* ]]; then
+    echo "$($SMI_BIN -Q index,name -f csv | tail -n +2 | wc -l)"
   else
     echo "$($SMI_BIN -l | grep GPU | wc -l)"
   fi
@@ -91,8 +93,7 @@ run_tests_for_model() {
   # Start prefill instances
   for i in $(seq 0 $((NUM_PREFILL_INSTANCES-1))); do
     # Calculate GPU ID - we'll distribute across available GPUs
-    #GPU_ID=$((i % $(get_num_gpus)))
-    GPU_ID=2
+    GPU_ID=$((i % $(get_num_gpus)))
 
     # Calculate port number (base port + instance number)
     PORT=$((8300 + i))
@@ -102,7 +103,7 @@ run_tests_for_model() {
     echo "Starting prefill instance $i on GPU $GPU_ID, port $PORT"
 
     # Build the command with or without model-specific args
-    BASE_CMD="RANK=0 UCX_TLS=rc,ud,ib VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
+    BASE_CMD="HABANA_VISIBLE_MODULES='0,1,2,3' RANK=$GPU_ID UCX_TLS=rc,ud,ib VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
     --port $PORT \
     --long_prefill_token_threshold 8192 \
     --max_num_batched_tokens 8192 \
@@ -127,7 +128,7 @@ run_tests_for_model() {
   # Start decode instances
   for i in $(seq 0 $((NUM_DECODE_INSTANCES-1))); do
     # Calculate GPU ID - we'll distribute across available GPUs, starting from after prefill GPUs
-    #GPU_ID=$(((i + NUM_PREFILL_INSTANCES) % $(get_num_gpus)))
+    GPU_ID=$(((i + NUM_PREFILL_INSTANCES) % $(get_num_gpus)))
     # Calculate port number (base port + instance number)
     PORT=$((8400 + i))
     # Calculate side channel port
@@ -136,7 +137,7 @@ run_tests_for_model() {
     echo "Starting decode instance $i on GPU $GPU_ID, port $PORT"
 
     # Build the command with or without model-specific args
-    BASE_CMD="RANK=1 UCX_TLS=rc,ud,ib VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
+    BASE_CMD="HABANA_VISIBLE_MODULES='4,5,6,7' RANK=$GPU_ID UCX_TLS=rc,ud,ib VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
     --port $PORT \
     --gpu-memory-utilization 0.3 \
     --tensor-parallel-size $DECODER_TP_SIZE \
diff --git a/examples/nixl/run_hpu_disagg_accuracy_test.sh b/examples/nixl/run_hpu_disagg_accuracy_test.sh
index cac822cdb..bde3295c4 100755
--- a/examples/nixl/run_hpu_disagg_accuracy_test.sh
+++ b/examples/nixl/run_hpu_disagg_accuracy_test.sh
@@ -22,7 +22,7 @@ DECODER_TP_SIZE=${DECODER_TP_SIZE:-1}
 #GIT_ROOT=$(git rev-parse --show-toplevel)
 GIT_ROOT="/home/vllm-nixl/vllm"
 
-#SMI_BIN=$(which nvidia-smi || which rocm-smi)
+SMI_BIN=$(which nvidia-smi || which rocm-smi || which hl-smi)
 
 # Trap the SIGINT signal (triggered by Ctrl+C)
 trap 'kill $(jobs -pr)' SIGINT SIGTERM EXIT
@@ -58,6 +58,8 @@ get_model_args() {
 get_num_gpus() {
   if [[ "$SMI_BIN" == *"nvidia"* ]]; then
     echo "$($SMI_BIN --query-gpu=name --format=csv,noheader | wc -l)"
+  elif [[ "$SMI_BIN" == *"hl"* ]]; then
+    echo "$($SMI_BIN -Q index,name -f csv | tail -n +2 | wc -l)"
   else
     echo "$($SMI_BIN -l | grep GPU | wc -l)"
   fi
@@ -82,8 +84,7 @@ run_tests_for_model() {
   # Start prefill instances
   for i in $(seq 0 $((NUM_PREFILL_INSTANCES-1))); do
     # Calculate GPU ID - we'll distribute across available GPUs
-    #GPU_ID=$((i % $(get_num_gpus)))
-    GPU_ID=2
+    GPU_ID=$((i % $(get_num_gpus)))
 
     # Calculate port number (base port + instance number)
     PORT=$((8300 + i))
@@ -93,7 +94,7 @@ run_tests_for_model() {
     echo "Starting prefill instance $i on GPU $GPU_ID, port $PORT"
 
     # Build the command with or without model-specific args
-    BASE_CMD="CUDA_VISIBLE_DEVICES=$GPU_ID VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
+    BASE_CMD="HABANA_VISIBLE_MODULES='0,1,2,3' RANK=$GPU_ID VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
     --port $PORT \
     --enforce-eager \
     --disable-log-requests \
@@ -117,8 +118,7 @@ run_tests_for_model() {
   # Start decode instances
   for i in $(seq 0 $((NUM_DECODE_INSTANCES-1))); do
     # Calculate GPU ID - we'll distribute across available GPUs, starting from after prefill GPUs
-    #GPU_ID=$(((i + NUM_PREFILL_INSTANCES) % $(get_num_gpus)))
-    GPU_ID=6
+    GPU_ID=$(((i + NUM_PREFILL_INSTANCES) % $(get_num_gpus)))
     # Calculate port number (base port + instance number)
     PORT=$((8400 + i))
     # Calculate side channel port
@@ -127,7 +127,7 @@ run_tests_for_model() {
     echo "Starting decode instance $i on GPU $GPU_ID, port $PORT"
 
     # Build the command with or without model-specific args
-    BASE_CMD="CUDA_VISIBLE_DEVICES=$GPU_ID VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
+    BASE_CMD="HABANA_VISIBLE_MODULES='4,5,6,7' RANK=$GPU_ID VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
     --port $PORT \
     --enforce-eager \
     --disable-log-requests \
diff --git a/tests/unit_tests/run_accuracy_test.sh b/tests/unit_tests/run_accuracy_test.sh
index 7ab10554b..76220ffb4 100755
--- a/tests/unit_tests/run_accuracy_test.sh
+++ b/tests/unit_tests/run_accuracy_test.sh
@@ -49,7 +49,7 @@ DECODER_TP_SIZE=${DECODER_TP_SIZE:-2}
 # Find the git repository root directory
 GIT_ROOT=$(git rev-parse --show-toplevel)
 
-#SMI_BIN=$(which nvidia-smi || which rocm-smi)
+SMI_BIN=$(which nvidia-smi || which rocm-smi || which hl-smi)
 
 # Trap the SIGINT signal (triggered by Ctrl+C)
 trap 'kill $(jobs -pr)' SIGINT SIGTERM EXIT
@@ -85,6 +85,8 @@ get_model_args() {
 get_num_gpus() {
   if [[ "$SMI_BIN" == *"nvidia"* ]]; then
     echo "$($SMI_BIN --query-gpu=name --format=csv,noheader | wc -l)"
+  elif [[ "$SMI_BIN" == *"hl"* ]]; then
+    echo "$($SMI_BIN -Q index,name -f csv | tail -n +2 | wc -l)"
   else
     echo "$($SMI_BIN -l | grep GPU | wc -l)"
   fi
@@ -109,8 +111,7 @@ run_tests_for_model() {
   # Start prefill instances
   for i in $(seq 0 $((NUM_PREFILL_INSTANCES-1))); do
     # Calculate GPU ID - we'll distribute across available GPUs
-    #GPU_ID=$((i % $(get_num_gpus)))
-    GPU_ID=2
+    GPU_ID=$((i % $(get_num_gpus)))
 
     # Calculate port number (base port + instance number)
     PORT=$((8700 + i))
@@ -120,7 +121,7 @@ run_tests_for_model() {
     echo "Starting prefill instance $i on GPU $GPU_ID, port $PORT"
 
     # Build the command with or without model-specific args
-    BASE_CMD="RANK=0 UCX_TLS=$UCX_TLS VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
+    BASE_CMD="HABANA_VISIBLE_MODULES='0,1,2,3' RANK=$GPU_ID UCX_TLS=$UCX_TLS VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
     --port $PORT \
     --enforce-eager \
     --max_num_batched_tokens 8192 \
@@ -144,7 +145,7 @@ run_tests_for_model() {
   # Start decode instances
   for i in $(seq 0 $((NUM_DECODE_INSTANCES-1))); do
     # Calculate GPU ID - we'll distribute across available GPUs, starting from after prefill GPUs
-    #GPU_ID=$(((i + NUM_PREFILL_INSTANCES) % $(get_num_gpus)))
+    GPU_ID=$(((i + NUM_PREFILL_INSTANCES) % $(get_num_gpus)))
     # Calculate port number (base port + instance number)
     PORT=$((8800 + i))
     # Calculate side channel port
@@ -153,7 +154,7 @@ run_tests_for_model() {
     echo "Starting decode instance $i on GPU $GPU_ID, port $PORT"
 
     # Build the command with or without model-specific args
-    BASE_CMD="RANK=1 UCX_TLS=$UCX_TLS VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
+    BASE_CMD="HABANA_VISIBLE_MODULES='4,5,6,7' RANK=$GPU_ID UCX_TLS=$UCX_TLS VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
     --port $PORT \
     --enforce-eager \
     --max_num_batched_tokens 8192 \
diff --git a/vllm_gaudi/v1/worker/hpu_worker.py b/vllm_gaudi/v1/worker/hpu_worker.py
index ae5705e54..0810f2773 100644
--- a/vllm_gaudi/v1/worker/hpu_worker.py
+++ b/vllm_gaudi/v1/worker/hpu_worker.py
@@ -94,6 +94,60 @@ def __init__(
         self.kv_cache_sleeping = False
         self.kv_cache_config = None
 
+        # Select available Habana modules before initializing the device.
+        self._configure_habana_visible_modules()
+
+    def _configure_habana_visible_modules(self):
+        import pyhlml
+        pyhlml.hlmlInit()
+        try:
+            available_module_ids = []
+            device_count = torch.hpu.device_count()
+            if device_count < 1:
+                raise RuntimeError("No Habana devices found.")
+            for i in range(device_count):
+                try:
+                    device = pyhlml.hlmlDeviceGetHandleByIndex(i)
+                    utility = pyhlml.hlmlDeviceGetUtilizationRates(device)
+                    if utility.aip == 0 and utility.memory == 0:
+                        module_id = pyhlml.hlmlDeviceGetModuleID(device)
+                        available_module_ids.append(module_id)
+                except Exception:
+                    continue
+            if len(available_module_ids) < 1:
+                raise RuntimeError("No available Habana modules found. All modules are currently in use.")
+            env_visible_modules = os.getenv("HABANA_VISIBLE_MODULES")
+            if env_visible_modules is None:
+                if len(available_module_ids) < self.parallel_config.world_size:
+                    raise RuntimeError(
+                        f"Not enough available modules for world_size={self.parallel_config.world_size}.")
+                available_modules_str = ",".join(map(str, sorted(available_module_ids)))
+                logger.info("HABANA_VISIBLE_MODULES is not set, using all available modules: %s", available_modules_str)
+                os.environ["HABANA_VISIBLE_MODULES"] = available_modules_str
+            else:
+                if not all(c.isdigit() for c in env_visible_modules.split(",")):
+                    raise RuntimeError(f"Invalid HABANA_VISIBLE_MODULES={env_visible_modules}. "
+                                       "It should be a comma-separated list of integers.")
+                env_module_ids = list(map(int, env_visible_modules.split(",")))
+                if any(module_id < 0 or module_id >= device_count for module_id in env_module_ids):
+                    raise RuntimeError(f"Invalid HABANA_VISIBLE_MODULES={env_visible_modules}. "
+                                       f"Module IDs should be between 0 and {device_count - 1}.")
+                if any(env_module_id not in available_module_ids for env_module_id in env_module_ids):
+                    logger.warning("Some device for HABANA_VISIBLE_MODULES=%s are not available.", env_visible_modules)
+                    selected_modules = [x for x in env_module_ids if x in available_module_ids]
+                    if len(selected_modules) < self.parallel_config.world_size:
+                        raise RuntimeError(
+                            f"Not enough available modules for world_size={self.parallel_config.world_size}. "
+                            "Set HABANA_VISIBLE_MODULES to include more available modules and try again.")
+                    else:
+                        selected_modules_str = ",".join(map(str, sorted(selected_modules)))
+                        os.environ["HABANA_VISIBLE_MODULES"] = selected_modules_str
+                        logger.warning("Using selected available modules: %s", selected_modules_str)
+        except Exception as e:
+            raise e
+        finally:
+            pyhlml.hlmlShutdown()
+
     def init_profiler(self):
         """Initialize the profiler."""
         if envs.VLLM_TORCH_PROFILER_DIR:
@@ -134,7 +188,10 @@ def stop_profile(self):
         self.profiler.stop()
 
     def init_device(self):
+        # Set the device for this worker.
         self.device = torch.device("hpu")
+        torch.hpu.set_device(self.local_rank)
+
         # Initialize the distributed environment.
         init_worker_distributed_environment(self.vllm_config, self.rank, self.distributed_init_method, self.local_rank)
         # Set random seed.