diff --git a/docs/src/content/docs/en/cli_reference.md b/docs/src/content/docs/en/cli_reference.md
index 66d8c75bdf..ad2ac5d08a 100644
--- a/docs/src/content/docs/en/cli_reference.md
+++ b/docs/src/content/docs/en/cli_reference.md
@@ -6,6 +6,8 @@ sidebar:
 
 xLLM uses gflags to manage service startup parameters. `--model <PATH>` is the only required flag. When `--config_json_file` is used, values in the JSON file override command-line flag values. The tables below are grouped by the Config classes in `/xllm/core/framework/config`, with one Config per section. The `ConfigJsonUtils` section contains the common JSON config-file flags.
 
+> **Device selection**: xLLM no longer provides `--devices` / `--device_id` / `--draft_devices`. The devices to run on are determined by the visible-device mask environment variables (`ASCEND_RT_VISIBLE_DEVICES` for NPU, `CUDA_VISIBLE_DEVICES` for NVIDIA, `MLU_VISIBLE_DEVICES` for Cambricon, `HIP_VISIBLE_DEVICES` for DCU, `MUSA_VISIBLE_DEVICES` for Moore Threads). Each process uses all of its visible devices: a single-card process exposes one card, while single-process multi-card exposes several. The draft model always uses the same devices as the target model.
+
 ## ConfigJsonUtils
 
 | Parameter | Type | Default | Description |
@@ -37,8 +39,6 @@ xLLM uses gflags to manage service startup parameters. `--model <PATH>` is the o
 | `model` | `string` | `""` | Hugging Face model name or model path. |
 | `backend` | `string` | `""` | Backend model type: `llm` for text-only models, `vlm` for multimodal models, or `dit` for diffusion models. |
 | `task` | `string` | `"generate"` | Model task, for example `generate`, `embed`, or `mm_embed`. |
-| `devices` | `string` | `""` | Deprecated; use `device_id` instead. Devices used by the current process, for example `npu:0` or `npu:0,npu:1`. |
-| `device_id` | `int32` | `-1` | Device id to run the model on, for example `0`. |
 | `limit_image_per_prompt` | `int32` | `8` | Maximum number of images per prompt. Only applies to multimodal models. |
 | `max_encoder_cache_size` | `int64` | `0` | Maximum GPU/NPU memory size in MB for the encoder cache per worker. `0` disables the encoder cache. |
 | `reasoning_parser` | `string` | `""` | Reasoning parser, for example `auto`, `glm45`, `glm47`, `glm5`, `qwen3`, `qwen35`, or `deepseek-r1`. |
@@ -181,7 +181,6 @@ xLLM uses gflags to manage service startup parameters. `--model <PATH>` is the o
 | Parameter | Type | Default | Description |
 |:----------|:-----|:--------|:------------|
 | `draft_model` | `string` | `""` | Draft model path. See [MTP](/en/features/mtp/) for MTP usage. |
-| `draft_devices` | `string` | `""` | Devices used by the draft model, for example `npu:0` or `npu:0,npu:1`. If omitted, uses the target model devices when speculative decoding is enabled. |
 | `num_speculative_tokens` | `int32` | `0` | Number of speculative tokens generated per speculative decoding step. |
 | `speculative_algorithm` | `string` | `"MTP"` | Speculative decoding algorithm. Supported values: `MTP`, `Eagle3`, `Suffix`, `DFlash`. |
 | `speculative_suffix_cache_max_depth` | `int32` | `64` | Maximum suffix-tree depth for suffix speculative decoding. |
diff --git a/docs/src/content/docs/en/cookbook/autoregressive_models/deepseek/deepseek_v4.md b/docs/src/content/docs/en/cookbook/autoregressive_models/deepseek/deepseek_v4.md
index f68b06b4af..509e2685d5 100644
--- a/docs/src/content/docs/en/cookbook/autoregressive_models/deepseek/deepseek_v4.md
+++ b/docs/src/content/docs/en/cookbook/autoregressive_models/deepseek/deepseek_v4.md
@@ -131,20 +131,17 @@ MASTER_NODE_ADDR="11.87.49.110:10015"
 LOCAL_HOST="11.87.49.110"
 # Service Port
 START_PORT=18994
-START_DEVICE=0
 LOG_DIR="logs"
 NNODES=8
 
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   nohup $XLLM_PATH -model-id ds \
     --model $MODEL_PATH \
     --host $LOCAL_HOST \
     --port $PORT \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --node_rank=$i \
@@ -166,7 +163,6 @@ done
 
     # Variables required when MTP is enabled
     # --draft_model=$DRAFT_MODEL_PATH \
-    # --draft_devices="npu:$DEVICE" \
     # --num_speculative_tokens=1 \
 
 # numactl -C xxxxx          NUMA core binding (query with: npu-smi info -t topo)
@@ -179,7 +175,6 @@ done
 #--enable_chunked_prefill   Enable chunked prefill
 #--enable_graph             Enable aclgraph
 #--draft_model              MTP - MTP weights path
-#--draft_devices            MTP - MTP inference device (same as main model)
 #--num_speculative_tokens   MTP - Number of speculative tokens
 ```
 
@@ -206,7 +201,6 @@ export ATB_MATMUL_SHUFFLE_K_ENABLE=0
 MASTER_NODE_ADDR="11.87.49.110:19990"
 LOCAL_HOST="11.87.49.110"
 START_PORT=15890
-START_DEVICE=0
 LOG_DIR="logs"
 NNODES=32
 LOCAL_NODES=16
@@ -215,12 +209,11 @@ unset HCCL_OP_EXPANSION_MODE
 
 for (( i=0; i<$LOCAL_NODES; i++ )); do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i)); LOG_FILE="$LOG_DIR/node_$i.log"
+  LOG_FILE="$LOG_DIR/node_$i.log"
   nohup $XLLM_PATH \
     --model $MODEL_PATH \
     --host $LOCAL_HOST \
     --port $PORT \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --node_rank=$i \
@@ -236,7 +229,6 @@ done
 MASTER_NODE_ADDR="11.87.49.110:19990"
 LOCAL_HOST="11.87.49.111"
 START_PORT=15890
-START_DEVICE=0
 LOG_DIR="logs"
 NNODES=32
 LOCAL_NODES=16
@@ -245,12 +237,11 @@ unset HCCL_OP_EXPANSION_MODE
 
 for (( i=0; i<$LOCAL_NODES; i++ )); do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i)); LOG_FILE="$LOG_DIR/node_$i.log"
+  LOG_FILE="$LOG_DIR/node_$i.log"
   nohup  $XLLM_PATH \
     --model $MODEL_PATH \
     --host $LOCAL_HOST \
     --port $PORT \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --node_rank=$((i + LOCAL_NODES)) \
diff --git a/docs/src/content/docs/en/cookbook/autoregressive_models/glm/glm_5.md b/docs/src/content/docs/en/cookbook/autoregressive_models/glm/glm_5.md
index fac8101a9d..9a57acda10 100644
--- a/docs/src/content/docs/en/cookbook/autoregressive_models/glm/glm_5.md
+++ b/docs/src/content/docs/en/cookbook/autoregressive_models/glm/glm_5.md
@@ -135,7 +135,6 @@ do
   nohup $XLLM_PATH \
     --model $MODEL_PATH \
     --port $PORT \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --node_rank=$i \
@@ -150,7 +149,6 @@ do
     --communication_backend="hccl" \
     --graph_decode_batch_size_limit=2 \
     --draft_model=$DRAFT_MODEL_PATH \
-    --draft_devices="npu:$DEVICE" \
     --num_speculative_tokens=3 \
     --ep_size=16 \
     --dp_size=2 \
@@ -168,7 +166,6 @@ done
 # --enable_graph             Enable aclgraph. It requires extra memory.
 # --acl_graph_decode_batch_size_limit    Maximum decode batch size for graph capture. It must be <= 32 / (number of speculative tokens + 1).
 # --draft_model              MTP draft-model weight path.
-# --draft_devices            MTP inference device, the same as the main model.
 # --num_speculative_tokens   Number of speculative tokens predicted by MTP.
 ```
 
@@ -211,7 +208,6 @@ do
     --model $MODEL_PATH \
     --host $LOCAL_HOST \
     --port $PORT \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --node_rank=$i \
@@ -226,7 +222,6 @@ do
     --enable_graph=true \
     --acl_graph_decode_batch_size_limit=4 \
     --draft_model=$DRAFT_MODEL_PATH \
-    --draft_devices="npu:$DEVICE" \
     --num_speculative_tokens=3 \
     --ep_size=32 \
     --dp_size=4 \
@@ -258,7 +253,6 @@ do
     --model $MODEL_PATH \
     --host $LOCAL_HOST \
     --port $PORT \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --node_rank=$((i + LOCAL_NODES)) \
@@ -273,7 +267,6 @@ do
     --enable_graph=true \
     --acl_graph_decode_batch_size_limit=4 \
     --draft_model=$DRAFT_MODEL_PATH \
-    --draft_devices="npu:$DEVICE" \
     --num_speculative_tokens=3 \
     --ep_size=32 \
     --dp_size=4 \
@@ -442,7 +435,6 @@ ENABLE_DECODE_RESPONSE_TO_SERVICE=true ../xllm-service/build/xllm_service/xllm_m
       --model $MODEL_PATH  --model_id glmmoe \
       --host $LOCAL_HOST \
       --port $PORT \
-      --devices="npu:$DEVICE" \
       --master_node_addr=$MASTER_NODE_ADDR \
       --nnodes=$NNODES \
       --node_rank=$i \
@@ -455,7 +447,6 @@ ENABLE_DECODE_RESPONSE_TO_SERVICE=true ../xllm-service/build/xllm_service/xllm_m
       --enable_chunked_prefill=false \
       --enable_graph=true \
       --draft_model $DRAFT_MODEL_PATH \
-      --draft_devices="npu:$DEVICE" \
       --num_speculative_tokens 1 \
       --tool_call_parser=auto \
       --enable_disagg_pd=true \
@@ -498,7 +489,6 @@ ENABLE_DECODE_RESPONSE_TO_SERVICE=true ../xllm-service/build/xllm_service/xllm_m
       --model $MODEL_PATH  --model_id glmmoe \
       --host $LOCAL_HOST \
       --port $PORT \
-      --devices="npu:$DEVICE" \
       --master_node_addr=$MASTER_NODE_ADDR \
       --nnodes=$NNODES \
       --node_rank=$i \
@@ -511,7 +501,6 @@ ENABLE_DECODE_RESPONSE_TO_SERVICE=true ../xllm-service/build/xllm_service/xllm_m
       --enable_chunked_prefill=false \
       --enable_graph=true \
       --draft_model $DRAFT_MODEL_PATH \
-      --draft_devices="npu:$DEVICE" \
       --num_speculative_tokens 1 \
       --tool_call_parser=auto \
       --enable_disagg_pd=true \
diff --git a/docs/src/content/docs/en/cookbook/autoregressive_models/kimi/kimi2_5.md b/docs/src/content/docs/en/cookbook/autoregressive_models/kimi/kimi2_5.md
index 00c2156c32..c6f4eeea4e 100644
--- a/docs/src/content/docs/en/cookbook/autoregressive_models/kimi/kimi2_5.md
+++ b/docs/src/content/docs/en/cookbook/autoregressive_models/kimi/kimi2_5.md
@@ -153,10 +153,10 @@ unset HCCL_OP_EXPANSION_MODE
 for (( i=0; i<$LOCAL_NODES; i++ ))do
   PORT=$((START_PORT + i))
   DEVICE=$((START_DEVICE + i));  LOG_FILE="$LOG_DIR/node_$i.log"
-  nohup numactl -C $((DEVICE*40))-$((DEVICE*40+39)) $XLLM_PATH \    --model $MODEL_PATH \
+  nohup numactl -C $((DEVICE*40))-$((DEVICE*40+39)) $XLLM_PATH \
+    --model $MODEL_PATH \
     --host $LOCAL_HOST \
     --port $PORT \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --node_rank=$i \
@@ -194,10 +194,10 @@ unset HCCL_OP_EXPANSION_MODE
 for (( i=0; i<$LOCAL_NODES; i++ ))do
   PORT=$((START_PORT + i))
   DEVICE=$((START_DEVICE + i));  LOG_FILE="$LOG_DIR/node_$i.log"
-  nohup numactl -C $((DEVICE*40))-$((DEVICE*40+39)) $XLLM_PATH \    --model $MODEL_PATH \
+  nohup numactl -C $((DEVICE*40))-$((DEVICE*40+39)) $XLLM_PATH \
+    --model $MODEL_PATH \
     --host $LOCAL_HOST \
     --port $PORT \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --node_rank=$((i + LOCAL_NODES)) \
diff --git a/docs/src/content/docs/en/cookbook/autoregressive_models/minmax/minmax_m2_7.md b/docs/src/content/docs/en/cookbook/autoregressive_models/minmax/minmax_m2_7.md
index 89bb876f5d..01e73b7118 100644
--- a/docs/src/content/docs/en/cookbook/autoregressive_models/minmax/minmax_m2_7.md
+++ b/docs/src/content/docs/en/cookbook/autoregressive_models/minmax/minmax_m2_7.md
@@ -167,20 +167,17 @@ MASTER_NODE_ADDR="10.143.3.204:10015"
 LOCAL_HOST="10.143.3.204"
 # Service port
 START_PORT=18994
-START_DEVICE=0
 LOG_DIR="logs"
 NNODES=16
 
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   nohup numactl -C $((i*40))-$((i*40+39)) $XLLM_PATH \
     --model $MODEL_PATH \
     --host $LOCAL_HOST \
     --port $PORT \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --node_rank=$i \
diff --git a/docs/src/content/docs/en/cookbook/autoregressive_models/qwen/qwen3_5.md b/docs/src/content/docs/en/cookbook/autoregressive_models/qwen/qwen3_5.md
index 0166b86be0..92d39b3895 100644
--- a/docs/src/content/docs/en/cookbook/autoregressive_models/qwen/qwen3_5.md
+++ b/docs/src/content/docs/en/cookbook/autoregressive_models/qwen/qwen3_5.md
@@ -91,7 +91,6 @@ export LD_LIBRARY_PATH=/usr/local/libtorch_npu/lib:$LD_LIBRARY_PATH
 source /usr/local/Ascend/ascend-toolkit/set_env.sh
 source /usr/local/Ascend/nnal/atb/set_env.sh
 
-export ASCEND_RT_VISIBLE_DEVICES=14,15
 export ASDOPS_LOG_TO_STDOUT=1
 export ASDOPS_LOG_LEVEL=0
 export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
@@ -120,19 +119,17 @@ DRAFT_MODEL_PATH="/path/to/Qwen3.5-27B-mtp"
 
 MASTER_NODE_ADDR="<master-host>:32764"
 START_PORT=18076
-START_DEVICE=0
 NNODES=2
 
+export ASCEND_RT_VISIBLE_DEVICES=14,15
 export HCCL_IF_BASE_PORT=53433
 
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   ./xllm/build/xllm/core/server/xllm \
     --model $MODEL_PATH \
-    --devices="npu:$DEVICE" \
     --port $PORT \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
@@ -151,7 +148,6 @@ do
     --max_concurrent_requests=8 \
     --backend llm \
     --draft_model $DRAFT_MODEL_PATH \
-    --draft_devices="npu:$DEVICE" \
     --num_speculative_tokens 3 \
     >> $LOG_FILE 2>&1 &
 done
diff --git a/docs/src/content/docs/en/cookbook/diffusion_models/flux/flux2.md b/docs/src/content/docs/en/cookbook/diffusion_models/flux/flux2.md
index 2d98560a70..a8e6cf1562 100644
--- a/docs/src/content/docs/en/cookbook/diffusion_models/flux/flux2.md
+++ b/docs/src/content/docs/en/cookbook/diffusion_models/flux/flux2.md
@@ -114,7 +114,6 @@ export HCCL_IF_BASE_PORT=43432  # HCCL communication base port
 MODEL_PATH="/path/to/flux2/text_encoder/"          # text_encoder path
 MASTER_NODE_ADDR="127.0.0.1:9748"                  # Master node address (must be consistent globally)
 START_PORT=18001                                   # Service starting port
-START_DEVICE=2                                     # Starting logical device ID
 LOG_DIR="log"                                      # Log directory
 NNODES=2                                           # Number of nodes (this script starts 1 process)
 
@@ -123,11 +122,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/mistral_node_$i.log"
   ./build/xllm/core/server/xllm \
     --model $MODEL_PATH \
-    --devices="npu:$DEVICE" \
     --port $PORT \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
@@ -185,21 +182,18 @@ export HCCL_IF_BASE_PORT=43432  # HCCL communication base port
 ```bash
 MASTER_NODE_ADDR="127.0.0.1:8999"                  # Master node address (must be consistent globally)
 START_PORT=18018                                   # Service starting port
-START_DEVICE=4                                     # Starting NPU logical device ID
 LOG_DIR="log"                                      # Log directory
 NNODES=2                                           # Number of nodes (this script starts 2 processes)
 
 for (( i=0; i<2; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/dit_node_$i.log"
   ./build/xllm/core/server/xllm \
     --model="/path/to/flux2/" \
     --max_memory_utilization=0.6 \
     --backend="dit" \
     --tp_size=2 \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --port $PORT \
diff --git a/docs/src/content/docs/en/features/disagg_pd.md b/docs/src/content/docs/en/features/disagg_pd.md
index 2a0edad184..f1a6832592 100644
--- a/docs/src/content/docs/en/features/disagg_pd.md
+++ b/docs/src/content/docs/en/features/disagg_pd.md
@@ -43,9 +43,8 @@ ENABLE_DECODE_RESPONSE_TO_SERVICE=true ./xllm_master_serving --etcd_addr="127.0.
 - Taking Qwen2-7B as an example  
     - Start Prefill Instance
         ```bash
-        /path/to/xllm --model=Qwen2-7B-Instruct \
+        ASCEND_RT_VISIBLE_DEVICES=0 /path/to/xllm --model=Qwen2-7B-Instruct \
                --port=8010 \
-               --devices="npu:0" \
                --master_node_addr="127.0.0.1:18888" \
                --enable_prefix_cache=false \
                --enable_chunked_prefill=false \
@@ -59,9 +58,8 @@ ENABLE_DECODE_RESPONSE_TO_SERVICE=true ./xllm_master_serving --etcd_addr="127.0.
         ```
     - Start Decode Instance 
         ```bash 
-        /path/to/xllm --model=Qwen2-7B-Instruct \
+        ASCEND_RT_VISIBLE_DEVICES=1 /path/to/xllm --model=Qwen2-7B-Instruct \
                --port=8020 \
-               --devices="npu:1" \
                --master_node_addr="127.0.0.1:18898" \
                --enable_prefix_cache=false \
                --enable_chunked_prefill=false \
diff --git a/docs/src/content/docs/en/features/mtp.md b/docs/src/content/docs/en/features/mtp.md
index f4d7ea5f17..91a458c0da 100644
--- a/docs/src/content/docs/en/features/mtp.md
+++ b/docs/src/content/docs/en/features/mtp.md
@@ -94,23 +94,19 @@ MODEL_PATH="/models/DeepSeek-V3"
 DRAFT_MODEL_PATH="/models/DeepSeek-V3-mtp"
 MASTER_NODE_ADDR="127.0.0.1:42123"
 START_PORT=13222
-START_DEVICE=0
 LOG_DIR="log"
 NNODES=16
 
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   nohup ./xllm \
     --model $MODEL_PATH \
-    --devices="npu:$DEVICE" \
     --port $PORT \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --draft_model $DRAFT_MODEL_PATH \
-    --draft_devices="npu:$DEVICE" \
     --num_speculative_tokens 1 \
     --max_memory_utilization=0.90 \
     --max_tokens_per_batch=10000 \
diff --git a/docs/src/content/docs/en/getting_started/launch_xllm.md b/docs/src/content/docs/en/getting_started/launch_xllm.md
index 028a6398fb..0c35d036d9 100644
--- a/docs/src/content/docs/en/getting_started/launch_xllm.md
+++ b/docs/src/content/docs/en/getting_started/launch_xllm.md
@@ -21,7 +21,6 @@ export HCCL_IF_BASE_PORT=43432  # HCCL communication base port
 MODEL_PATH="/path/to/model/Qwen3-8B"               # Model path
 MASTER_NODE_ADDR="127.0.0.1:9748"                  # Master node address (must be globally consistent)
 START_PORT=18000                                   # Service starting port
-START_DEVICE=0                                     # Starting logical device number
 LOG_DIR="log"                                      # Log directory
 NNODES=1                                           # Number of nodes (current script launches 1 process)
 
@@ -30,11 +29,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="npu:$DEVICE" \
     --port $PORT \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
@@ -64,7 +61,6 @@ export CUDA_VISIBLE_DEVICES=0
 MODEL_PATH="/path/to/model/Qwen3-8B"
 MASTER_NODE_ADDR="127.0.0.1:9748"
 START_PORT=18000
-START_DEVICE=0
 LOG_DIR="log"
 NNODES=1
 
@@ -73,11 +69,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="cuda:$DEVICE" \
     --port $PORT \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
@@ -103,7 +97,6 @@ export MLU_VISIBLE_DEVICES=0
 MODEL_PATH="/path/to/model/Qwen3-8B"
 MASTER_NODE_ADDR="127.0.0.1:9748"
 START_PORT=18000
-START_DEVICE=0
 LOG_DIR="log"
 NNODES=1
 
@@ -112,11 +105,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="mlu:$DEVICE" \
     --port $PORT \
     --nnodes=$NNODES \
     --master_node_addr=$MASTER_NODE_ADDR \
@@ -138,7 +129,6 @@ export HIP_VISIBLE_DEVICES=0
 MODEL_PATH="/path/to/model/Qwen3-8B"
 MASTER_NODE_ADDR="127.0.0.1:9748"
 START_PORT=18000
-START_DEVICE=0
 LOG_DIR="log"
 NNODES=1
 
@@ -147,11 +137,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="dcu:$DEVICE" \
     --port $PORT \
     --nnodes=$NNODES \
     --master_node_addr=$MASTER_NODE_ADDR \
@@ -174,7 +162,6 @@ export FLASHINFER_OPS_PATH=/opt/conda/lib/python3.10/site-packages/flashinfer/da
 MODEL_PATH="/path/to/model/Qwen3-8B"
 MASTER_NODE_ADDR="127.0.0.1:9748"
 START_PORT=18000
-START_DEVICE=0
 LOG_DIR="log"
 NNODES=1
 
@@ -183,11 +170,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="cuda:$DEVICE" \
     --port $PORT \
     --nnodes=$NNODES \
     --master_node_addr=$MASTER_NODE_ADDR \
@@ -213,7 +198,6 @@ export MUSA_VISIBLE_DEVICES=0
 MODEL_PATH="/path/to/model/Qwen3.5-27B"
 MASTER_NODE_ADDR="127.0.0.1:9748"
 START_PORT=18000
-START_DEVICE=0
 LOG_DIR="log"
 NNODES=1
 
@@ -222,11 +206,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="musa:$DEVICE" \
     --port $PORT \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
diff --git a/docs/src/content/docs/en/getting_started/multi_machine.md b/docs/src/content/docs/en/getting_started/multi_machine.md
index 199742acef..4b4c43c032 100644
--- a/docs/src/content/docs/en/getting_started/multi_machine.md
+++ b/docs/src/content/docs/en/getting_started/multi_machine.md
@@ -25,7 +25,6 @@ MODEL_PATH="/path/to/your/DeepSeek-R1"             # Model path
 MASTER_NODE_ADDR="123.123.123.123:9748"            # Master node address (must be globally consistent)
 LOCAL_HOST=123.123.123.123                         # Local IP for service launch
 START_PORT=18000                                   # Service starting port
-START_DEVICE=0                                     # Starting NPU logical device number
 LOCAL_NODES=16                                     # Number of local processes (this script launches 16 processes)
 LOG_DIR="log"                                      # Log directory
 NNODES=32                                          # Total number of GPUs/NPUs (32 in this 2-machine example)
@@ -35,13 +34,11 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$LOCAL_NODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   /path/to/xllm \
     --model $MODEL_PATH \
     --host $LOCAL_HOST \
     --port $PORT \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --max_memory_utilization=0.86 \
@@ -77,7 +74,6 @@ MODEL_PATH="/path/to/your/DeepSeek-R1"             # Model path
 MASTER_NODE_ADDR="123.123.123.123:9748"            # Master node address (must be globally consistent)
 LOCAL_HOST=456.456.456.456                         # Local IP for service launch
 START_PORT=18000                                   # Service starting port
-START_DEVICE=0                                     # Starting NPU logical device number
 LOCAL_NODES=16                                     # Number of local processes (this script launches 16 processes)
 LOG_DIR="log"                                      # Log directory
 NNODES=32                                          # Total number of GPUs/NPUs (32 in this 2-machine example)
@@ -87,13 +83,11 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$LOCAL_NODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   /path/to/xllm \
     --model $MODEL_PATH \
     --host $LOCAL_HOST \
     --port $PORT \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --max_memory_utilization=0.86 \
diff --git a/docs/src/content/docs/en/getting_started/offline_service.md b/docs/src/content/docs/en/getting_started/offline_service.md
index 5bc87cf513..703f3d162a 100644
--- a/docs/src/content/docs/en/getting_started/offline_service.md
+++ b/docs/src/content/docs/en/getting_started/offline_service.md
@@ -16,7 +16,9 @@ Use `BeamSearchParams` with `beam_width` greater than `1`, then call `llm.beam_s
 ```python
 from xllm import BeamSearchParams, LLM
 
-llm = LLM(model="/path/models/Qwen2-7B-Instruct", devices="npu:0")
+# The devices to run on are selected via ASCEND_RT_VISIBLE_DEVICES
+# (or CUDA_VISIBLE_DEVICES / MLU_VISIBLE_DEVICES for other platforms).
+llm = LLM(model="/path/models/Qwen2-7B-Instruct")
 params = BeamSearchParams(
     beam_width=2,
     top_logprobs=4,
diff --git a/docs/src/content/docs/en/hardware/ascend_npu.md b/docs/src/content/docs/en/hardware/ascend_npu.md
index fbdb261286..1315370b6c 100644
--- a/docs/src/content/docs/en/hardware/ascend_npu.md
+++ b/docs/src/content/docs/en/hardware/ascend_npu.md
@@ -63,7 +63,6 @@ export HCCL_IF_BASE_PORT=43432  # HCCL communication base port
 MODEL_PATH="/path/to/model/Qwen3-8B"               # Model path
 MASTER_NODE_ADDR="127.0.0.1:9748"                  # Master node address (must be globally consistent)
 START_PORT=18000                                   # Service starting port
-START_DEVICE=0                                     # Starting logical device number
 LOG_DIR="log"                                      # Log directory
 NNODES=1                                           # Number of nodes (current script launches 1 process)
 
@@ -72,11 +71,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="npu:$DEVICE" \
     --port $PORT \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
diff --git a/docs/src/content/docs/en/hardware/cambricon_mlu.md b/docs/src/content/docs/en/hardware/cambricon_mlu.md
index e1da637872..98b1f7461e 100644
--- a/docs/src/content/docs/en/hardware/cambricon_mlu.md
+++ b/docs/src/content/docs/en/hardware/cambricon_mlu.md
@@ -38,7 +38,6 @@ export MLU_VISIBLE_DEVICES=0
 MODEL_PATH="/path/to/model/Qwen3-8B"
 MASTER_NODE_ADDR="127.0.0.1:9748"
 START_PORT=18000
-START_DEVICE=0
 LOG_DIR="log"
 NNODES=1
 
@@ -47,11 +46,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="mlu:$DEVICE" \
     --port $PORT \
     --nnodes=$NNODES \
     --master_node_addr=$MASTER_NODE_ADDR \
diff --git a/docs/src/content/docs/en/hardware/dcu.md b/docs/src/content/docs/en/hardware/dcu.md
index 33ec7a6f6c..afcceb8df7 100644
--- a/docs/src/content/docs/en/hardware/dcu.md
+++ b/docs/src/content/docs/en/hardware/dcu.md
@@ -50,7 +50,6 @@ export HIP_VISIBLE_DEVICES=0
 MODEL_PATH="/path/to/model/Qwen3-8B"
 MASTER_NODE_ADDR="127.0.0.1:9748"
 START_PORT=18000
-START_DEVICE=0
 LOG_DIR="log"
 NNODES=1
 
@@ -59,11 +58,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="dcu:$DEVICE" \
     --port $PORT \
     --nnodes=$NNODES \
     --master_node_addr=$MASTER_NODE_ADDR \
diff --git a/docs/src/content/docs/en/hardware/metax_maca.md b/docs/src/content/docs/en/hardware/metax_maca.md
index 073e60c804..e661315b0a 100644
--- a/docs/src/content/docs/en/hardware/metax_maca.md
+++ b/docs/src/content/docs/en/hardware/metax_maca.md
@@ -53,7 +53,6 @@ export FLASHINFER_OPS_PATH=/opt/conda/lib/python3.10/site-packages/flashinfer/da
 MODEL_PATH="/path/to/model/Qwen3-8B"
 MASTER_NODE_ADDR="127.0.0.1:9748"
 START_PORT=18000
-START_DEVICE=0
 LOG_DIR="log"
 NNODES=1
 
@@ -62,11 +61,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="cuda:$DEVICE" \
     --port $PORT \
     --nnodes=$NNODES \
     --master_node_addr=$MASTER_NODE_ADDR \
diff --git a/docs/src/content/docs/en/hardware/musa.md b/docs/src/content/docs/en/hardware/musa.md
index 5471398d08..3625ba5764 100644
--- a/docs/src/content/docs/en/hardware/musa.md
+++ b/docs/src/content/docs/en/hardware/musa.md
@@ -5,7 +5,7 @@ sidebar:
   order: 7
 ---
 
-Use the MUSA device backend (`--devices=musa:<id>`) when running xLLM on Mthreads MUSA hardware.
+Use the MUSA device backend when running xLLM on Mthreads MUSA hardware; select the visible card(s) with `MUSA_VISIBLE_DEVICES`.
 
 ## Image and Container Startup
 
@@ -49,7 +49,6 @@ export MUSA_VISIBLE_DEVICES=0
 MODEL_PATH="/path/to/model/Qwen3.5-27B"
 MASTER_NODE_ADDR="127.0.0.1:9748"
 START_PORT=18000
-START_DEVICE=0
 LOG_DIR="log"
 NNODES=1
 
@@ -58,11 +57,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="musa:$DEVICE" \
     --port $PORT \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
diff --git a/docs/src/content/docs/en/hardware/nvidia_gpu.md b/docs/src/content/docs/en/hardware/nvidia_gpu.md
index be797ce95d..01fe120e06 100644
--- a/docs/src/content/docs/en/hardware/nvidia_gpu.md
+++ b/docs/src/content/docs/en/hardware/nvidia_gpu.md
@@ -48,7 +48,6 @@ export CUDA_VISIBLE_DEVICES=0
 MODEL_PATH="/path/to/model/Qwen3-8B"
 MASTER_NODE_ADDR="127.0.0.1:9748"
 START_PORT=18000
-START_DEVICE=0
 LOG_DIR="log"
 NNODES=1
 
@@ -57,11 +56,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="cuda:$DEVICE" \
     --port $PORT \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
diff --git a/docs/src/content/docs/en/hardware/overview.md b/docs/src/content/docs/en/hardware/overview.md
index 417011e7c2..e95127b400 100644
--- a/docs/src/content/docs/en/hardware/overview.md
+++ b/docs/src/content/docs/en/hardware/overview.md
@@ -14,7 +14,7 @@ xLLM supports multiple accelerator backends for large-scale model inference. Thi
 - [Cambricon MLU](/en/hardware/cambricon_mlu/) - MLU backend setup and launch entry points.
 - [Hygon DCU](/en/hardware/dcu/) - Hygon DCU backend setup and launch entry points.
 - [MetaX MACA](/en/hardware/metax_maca/) - MetaX MACA backend setup and launch entry points.
-- [Mthreads MUSA](/en/hardware/musa/) - Mthreads MUSA GPU image, build, and `--devices=musa` launch entry points.
+- [Mthreads MUSA](/en/hardware/musa/) - Mthreads MUSA GPU image, build, and launch entry points (device selection via `MUSA_VISIBLE_DEVICES`).
 
 ## Common Workflow
 
diff --git a/docs/src/content/docs/zh/cli_reference.md b/docs/src/content/docs/zh/cli_reference.md
index 368bf908e5..e22b1b4ea6 100644
--- a/docs/src/content/docs/zh/cli_reference.md
+++ b/docs/src/content/docs/zh/cli_reference.md
@@ -6,6 +6,8 @@ sidebar:
 
 xLLM 使用 gflags 管理服务启动参数。`--model <PATH>` 是唯一必填参数。使用 `--config_json_file` 时，JSON 文件中的值会覆盖命令行 flag 值。下表按 `/xllm/core/framework/config` 下的 Config 类分组，一个 Config 对应一节；`ConfigJsonUtils` 一节包含配置文件相关的通用参数。
 
+> **设备选择**：xLLM 不再提供 `--devices` / `--device_id` / `--draft_devices` 参数。运行设备由可见设备掩码环境变量决定（NPU 用 `ASCEND_RT_VISIBLE_DEVICES`，NVIDIA 用 `CUDA_VISIBLE_DEVICES`，寒武纪用 `MLU_VISIBLE_DEVICES`，DCU 用 `HIP_VISIBLE_DEVICES`，摩尔线程用 `MUSA_VISIBLE_DEVICES`）。每个进程会使用其可见的全部设备：单卡进程只暴露一张卡，单进程多卡则暴露多张卡。draft 模型始终与 target 模型使用相同设备。
+
 ## ConfigJsonUtils
 
 | 参数名称 | 类型 | 默认值 | 参数含义 |
@@ -37,8 +39,6 @@ xLLM 使用 gflags 管理服务启动参数。`--model <PATH>` 是唯一必填
 | `model` | `string` | `""` | Hugging Face 模型名称或模型路径。 |
 | `backend` | `string` | `""` | 后端模型类型；`llm` 表示纯文本模型，`vlm` 表示多模态模型，`dit` 表示扩散模型。 |
 | `task` | `string` | `"generate"` | 模型任务类型，例如 `generate`、`embed`、`mm_embed`。 |
-| `devices` | `string` | `""` | 已废弃，请改用 `device_id`。当前进程使用的设备，例如 `npu:0`、`npu:0,npu:1`。 |
-| `device_id` | `int32` | `-1` | 运行模型的 device id，例如 `0`。 |
 | `limit_image_per_prompt` | `int32` | `8` | 每个 prompt 允许的最大图片数量，仅用于多模态模型。 |
 | `max_encoder_cache_size` | `int64` | `0` | 每个 worker 的 encoder cache 最大显存大小，单位 MB；`0` 表示禁用 encoder cache。 |
 | `reasoning_parser` | `string` | `""` | reasoning 交互解析器，例如 `auto`、`glm45`、`glm47`、`glm5`、`qwen3`、`qwen35`、`deepseek-r1`。 |
@@ -181,7 +181,6 @@ xLLM 使用 gflags 管理服务启动参数。`--model <PATH>` 是唯一必填
 | 参数名称 | 类型 | 默认值 | 参数含义 |
 |:---------|:-----|:-------|:---------|
 | `draft_model` | `string` | `""` | draft 模型路径；MTP 使用方式详见 [MTP](/zh/features/mtp/)。 |
-| `draft_devices` | `string` | `""` | draft 模型使用的设备，例如 `npu:0`、`npu:0,npu:1`；未指定时，启用 speculative decoding 会使用 target 模型的设备。 |
 | `num_speculative_tokens` | `int32` | `0` | 每轮 speculative decoding 生成的 speculative token 数。 |
 | `speculative_algorithm` | `string` | `"MTP"` | Speculative decoding 算法，支持 `MTP`、`Eagle3`、`Suffix`、`DFlash`。 |
 | `speculative_suffix_cache_max_depth` | `int32` | `64` | Suffix speculative decoding 的后缀树最大深度。 |
diff --git a/docs/src/content/docs/zh/cookbook/autoregressive_models/deepseek/deepseek_v4.md b/docs/src/content/docs/zh/cookbook/autoregressive_models/deepseek/deepseek_v4.md
index 0e2be15dec..d86e557a06 100644
--- a/docs/src/content/docs/zh/cookbook/autoregressive_models/deepseek/deepseek_v4.md
+++ b/docs/src/content/docs/zh/cookbook/autoregressive_models/deepseek/deepseek_v4.md
@@ -131,20 +131,17 @@ MASTER_NODE_ADDR="11.87.49.110:10015"
 LOCAL_HOST="11.87.49.110"
 # Service Port
 START_PORT=18994
-START_DEVICE=0
 LOG_DIR="logs"
 NNODES=8
 
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   nohup $XLLM_PATH -model-id ds \
     --model $MODEL_PATH \
     --host $LOCAL_HOST \
     --port $PORT \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --node_rank=$i \
@@ -166,7 +163,6 @@ done
 
     # 开启mtp时需要的变量
     # --draft_model=$DRAFT_MODEL_PATH \
-    # --draft_devices="npu:$DEVICE" \
     # --num_speculative_tokens=1 \
 
 # numactl -C xxxxx          亲和性绑核(NUMA亲和性查询命令： npu-smi info -t topo)
@@ -179,7 +175,6 @@ done
 #--enable_chunked_prefill   开启chunked_prefill
 #--enable_graph             开启aclgraph
 #--draft_model              mtp - mtp权重路径
-#--draft_devices            mtp - mtp推理设备(与主模型同一)
 #--num_speculative_tokens   mtp - 预测token数
 ```
 
@@ -206,7 +201,6 @@ export ATB_MATMUL_SHUFFLE_K_ENABLE=0
 MASTER_NODE_ADDR="11.87.49.110:19990"
 LOCAL_HOST="11.87.49.110"
 START_PORT=15890
-START_DEVICE=0
 LOG_DIR="logs"
 NNODES=32
 LOCAL_NODES=16
@@ -215,12 +209,11 @@ unset HCCL_OP_EXPANSION_MODE
 
 for (( i=0; i<$LOCAL_NODES; i++ )); do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i)); LOG_FILE="$LOG_DIR/node_$i.log"
+  LOG_FILE="$LOG_DIR/node_$i.log"
   nohup $XLLM_PATH \
     --model $MODEL_PATH \
     --host $LOCAL_HOST \
     --port $PORT \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --node_rank=$i \
@@ -236,7 +229,6 @@ done
 MASTER_NODE_ADDR="11.87.49.110:19990"
 LOCAL_HOST="11.87.49.111"
 START_PORT=15890
-START_DEVICE=0
 LOG_DIR="logs"
 NNODES=32
 LOCAL_NODES=16
@@ -245,12 +237,11 @@ unset HCCL_OP_EXPANSION_MODE
 
 for (( i=0; i<$LOCAL_NODES; i++ )); do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i)); LOG_FILE="$LOG_DIR/node_$i.log"
+  LOG_FILE="$LOG_DIR/node_$i.log"
   nohup  $XLLM_PATH \
     --model $MODEL_PATH \
     --host $LOCAL_HOST \
     --port $PORT \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --node_rank=$((i + LOCAL_NODES)) \
diff --git a/docs/src/content/docs/zh/cookbook/autoregressive_models/glm/glm_5.md b/docs/src/content/docs/zh/cookbook/autoregressive_models/glm/glm_5.md
index 1037900c5c..12b7b0b9a0 100644
--- a/docs/src/content/docs/zh/cookbook/autoregressive_models/glm/glm_5.md
+++ b/docs/src/content/docs/zh/cookbook/autoregressive_models/glm/glm_5.md
@@ -136,7 +136,6 @@ do
   nohup $XLLM_PATH \
     --model $MODEL_PATH \
     --port $PORT \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --node_rank=$i \
@@ -151,7 +150,6 @@ do
     --communication_backend="hccl" \
     --graph_decode_batch_size_limit=2 \
     --draft_model=$DRAFT_MODEL_PATH \
-    --draft_devices="npu:$DEVICE" \
     --num_speculative_tokens=3 \
     --ep_size=16 \
     --dp_size=2 \
@@ -169,7 +167,6 @@ done
 # --enable_graph             开启aclgraph，需要额外显存
 # --acl_graph_decode_batch_size_limit    抓图的最大bs，当前需<= 32 / (预测token数 + 1)
 # --draft_model              mtp - mtp权重路径
-# --draft_devices            mtp - mtp推理设备(与主模型同一)
 # --num_speculative_tokens   mtp - 预测token数
 ```
 
@@ -206,10 +203,10 @@ unset HCCL_OP_EXPANSION_MODE
 for (( i=0; i<$LOCAL_NODES; i++ ))do
   PORT=$((START_PORT + i))
   DEVICE=$((START_DEVICE + i));  LOG_FILE="$LOG_DIR/node_$i.log"
-  nohup numactl -C $((DEVICE*40))-$((DEVICE*40+39)) $XLLM_PATH \    --model $MODEL_PATH \
+  nohup numactl -C $((DEVICE*40))-$((DEVICE*40+39)) $XLLM_PATH \
+    --model $MODEL_PATH \
     --host $LOCAL_HOST \
     --port $PORT \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --node_rank=$i \
@@ -224,7 +221,6 @@ for (( i=0; i<$LOCAL_NODES; i++ ))do
     --enable_graph=true \
     --acl_graph_decode_batch_size_limit=4 \
     --draft_model=$DRAFT_MODEL_PATH \
-    --draft_devices="npu:$DEVICE" \
     --num_speculative_tokens=3 \
     --ep_size=32 \
     --dp_size=4 \
@@ -250,10 +246,10 @@ unset HCCL_OP_EXPANSION_MODE
 for (( i=0; i<$LOCAL_NODES; i++ ))do
   PORT=$((START_PORT + i))
   DEVICE=$((START_DEVICE + i));  LOG_FILE="$LOG_DIR/node_$i.log"
-  nohup numactl -C $((DEVICE*40))-$((DEVICE*40+39)) $XLLM_PATH \    --model $MODEL_PATH \
+  nohup numactl -C $((DEVICE*40))-$((DEVICE*40+39)) $XLLM_PATH \
+    --model $MODEL_PATH \
     --host $LOCAL_HOST \
     --port $PORT \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --node_rank=$((i + LOCAL_NODES)) \
@@ -268,7 +264,6 @@ for (( i=0; i<$LOCAL_NODES; i++ ))do
     --enable_graph=true \
     --acl_graph_decode_batch_size_limit=4 \
     --draft_model=$DRAFT_MODEL_PATH \
-    --draft_devices="npu:$DEVICE" \
     --num_speculative_tokens=3 \
     --ep_size=32 \
     --dp_size=4 \
@@ -430,7 +425,6 @@ ENABLE_DECODE_RESPONSE_TO_SERVICE=true ../xllm-service/build/xllm_service/xllm_m
       --model $MODEL_PATH  --model_id glmmoe \
       --host $LOCAL_HOST \
       --port $PORT \
-      --devices="npu:$DEVICE" \
       --master_node_addr=$MASTER_NODE_ADDR \
       --nnodes=$NNODES \
       --node_rank=$i \
@@ -443,7 +437,6 @@ ENABLE_DECODE_RESPONSE_TO_SERVICE=true ../xllm-service/build/xllm_service/xllm_m
       --enable_chunked_prefill=false \
       --enable_graph=true \
       --draft_model $DRAFT_MODEL_PATH \
-      --draft_devices="npu:$DEVICE" \
       --num_speculative_tokens 1 \
       --tool_call_parser=auto \
       --enable_disagg_pd=true \
@@ -486,7 +479,6 @@ ENABLE_DECODE_RESPONSE_TO_SERVICE=true ../xllm-service/build/xllm_service/xllm_m
       --model $MODEL_PATH  --model_id glmmoe \
       --host $LOCAL_HOST \
       --port $PORT \
-      --devices="npu:$DEVICE" \
       --master_node_addr=$MASTER_NODE_ADDR \
       --nnodes=$NNODES \
       --node_rank=$i \
@@ -499,7 +491,6 @@ ENABLE_DECODE_RESPONSE_TO_SERVICE=true ../xllm-service/build/xllm_service/xllm_m
       --enable_chunked_prefill=false \
       --enable_graph=true \
       --draft_model $DRAFT_MODEL_PATH \
-      --draft_devices="npu:$DEVICE" \
       --num_speculative_tokens 1 \
       --tool_call_parser=auto \
       --enable_disagg_pd=true \
diff --git a/docs/src/content/docs/zh/cookbook/autoregressive_models/kimi/kimi2_5.md b/docs/src/content/docs/zh/cookbook/autoregressive_models/kimi/kimi2_5.md
index 67c50d772b..ca689cc45e 100644
--- a/docs/src/content/docs/zh/cookbook/autoregressive_models/kimi/kimi2_5.md
+++ b/docs/src/content/docs/zh/cookbook/autoregressive_models/kimi/kimi2_5.md
@@ -153,10 +153,10 @@ unset HCCL_OP_EXPANSION_MODE
 for (( i=0; i<$LOCAL_NODES; i++ ))do
   PORT=$((START_PORT + i))
   DEVICE=$((START_DEVICE + i));  LOG_FILE="$LOG_DIR/node_$i.log"
-  nohup numactl -C $((DEVICE*40))-$((DEVICE*40+39)) $XLLM_PATH \    --model $MODEL_PATH \
+  nohup numactl -C $((DEVICE*40))-$((DEVICE*40+39)) $XLLM_PATH \
+    --model $MODEL_PATH \
     --host $LOCAL_HOST \
     --port $PORT \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --node_rank=$i \
@@ -194,10 +194,10 @@ unset HCCL_OP_EXPANSION_MODE
 for (( i=0; i<$LOCAL_NODES; i++ ))do
   PORT=$((START_PORT + i))
   DEVICE=$((START_DEVICE + i));  LOG_FILE="$LOG_DIR/node_$i.log"
-  nohup numactl -C $((DEVICE*40))-$((DEVICE*40+39)) $XLLM_PATH \    --model $MODEL_PATH \
+  nohup numactl -C $((DEVICE*40))-$((DEVICE*40+39)) $XLLM_PATH \
+    --model $MODEL_PATH \
     --host $LOCAL_HOST \
     --port $PORT \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --node_rank=$((i + LOCAL_NODES)) \
diff --git a/docs/src/content/docs/zh/cookbook/autoregressive_models/minmax/minmax_m2_7.md b/docs/src/content/docs/zh/cookbook/autoregressive_models/minmax/minmax_m2_7.md
index 2fc02f1c05..d2e290badf 100644
--- a/docs/src/content/docs/zh/cookbook/autoregressive_models/minmax/minmax_m2_7.md
+++ b/docs/src/content/docs/zh/cookbook/autoregressive_models/minmax/minmax_m2_7.md
@@ -167,20 +167,17 @@ MASTER_NODE_ADDR="10.143.3.204:10015"
 LOCAL_HOST="10.143.3.204"
 # Service Port
 START_PORT=18994
-START_DEVICE=0
 LOG_DIR="logs"
 NNODES=16
 
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   nohup numactl -C $((i*40))-$((i*40+39)) $XLLM_PATH \
     --model $MODEL_PATH \
     --host $LOCAL_HOST \
     --port $PORT \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --node_rank=$i \
diff --git a/docs/src/content/docs/zh/cookbook/autoregressive_models/qwen/qwen3_5.md b/docs/src/content/docs/zh/cookbook/autoregressive_models/qwen/qwen3_5.md
index 39342f4493..9130936415 100644
--- a/docs/src/content/docs/zh/cookbook/autoregressive_models/qwen/qwen3_5.md
+++ b/docs/src/content/docs/zh/cookbook/autoregressive_models/qwen/qwen3_5.md
@@ -91,7 +91,6 @@ export LD_LIBRARY_PATH=/usr/local/libtorch_npu/lib:$LD_LIBRARY_PATH
 source /usr/local/Ascend/ascend-toolkit/set_env.sh
 source /usr/local/Ascend/nnal/atb/set_env.sh
 
-export ASCEND_RT_VISIBLE_DEVICES=14,15
 export ASDOPS_LOG_TO_STDOUT=1
 export ASDOPS_LOG_LEVEL=0
 export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
@@ -120,19 +119,17 @@ DRAFT_MODEL_PATH="/path/to/Qwen3.5-27B-mtp"
 
 MASTER_NODE_ADDR="<master-host>:32764"
 START_PORT=18076
-START_DEVICE=0
 NNODES=2
 
+export ASCEND_RT_VISIBLE_DEVICES=14,15
 export HCCL_IF_BASE_PORT=53433
 
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   ./xllm/build/xllm/core/server/xllm \
     --model $MODEL_PATH \
-    --devices="npu:$DEVICE" \
     --port $PORT \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
@@ -151,7 +148,6 @@ do
     --max_concurrent_requests=8 \
     --backend llm \
     --draft_model $DRAFT_MODEL_PATH \
-    --draft_devices="npu:$DEVICE" \
     --num_speculative_tokens 3 \
     >> $LOG_FILE 2>&1 &
 done
diff --git a/docs/src/content/docs/zh/cookbook/diffusion_models/flux/flux2.md b/docs/src/content/docs/zh/cookbook/diffusion_models/flux/flux2.md
index 950945fda0..19d9588f44 100644
--- a/docs/src/content/docs/zh/cookbook/diffusion_models/flux/flux2.md
+++ b/docs/src/content/docs/zh/cookbook/diffusion_models/flux/flux2.md
@@ -115,7 +115,6 @@ export HCCL_IF_BASE_PORT=43432  # HCCL 通信基础端口
 MODEL_PATH="/path/to/flux2/text_encoder/"          # text_encoder路径
 MASTER_NODE_ADDR="127.0.0.1:9748"                  # Master 节点地址（需全局一致）
 START_PORT=18001                                   # 服务起始端口
-START_DEVICE=2                                     # 起始逻辑设备号
 LOG_DIR="log"                                      # 日志目录
 NNODES=2                                           # 节点数（当前脚本启动 1 个进程）
 
@@ -124,11 +123,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/mistral_node_$i.log"
   ./build/xllm/core/server/xllm \
     --model $MODEL_PATH \
-    --devices="npu:$DEVICE" \
     --port $PORT \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
@@ -186,21 +183,18 @@ export HCCL_IF_BASE_PORT=43432  # HCCL 通信基础端口
 ```bash
 MASTER_NODE_ADDR="127.0.0.1:8999"                  # Master 节点地址（需全局一致）
 START_PORT=18018                                   # 服务起始端口
-START_DEVICE=4                                     # 起始 NPU 逻辑设备号
 LOG_DIR="log"                                      # 日志目录
 NNODES=2                                           # 节点数（当前脚本启动 2 个进程）
 
 for (( i=0; i<2; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/dit_node_$i.log"
   ./build/xllm/core/server/xllm \
     --model="/path/to/flux2/" \
     --max_memory_utilization=0.6 \
     --backend="dit" \
     --tp_size=2 \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --port $PORT \
diff --git a/docs/src/content/docs/zh/features/disagg_pd.md b/docs/src/content/docs/zh/features/disagg_pd.md
index 4dbdd6b310..db66c93d51 100644
--- a/docs/src/content/docs/zh/features/disagg_pd.md
+++ b/docs/src/content/docs/zh/features/disagg_pd.md
@@ -34,9 +34,8 @@ ENABLE_DECODE_RESPONSE_TO_SERVICE=true ./xllm_master_serving --etcd_addr="127.0.
 4. 以Qwen2-7B为例
     - 启动Prefill实例
         ```bash
-        /path/to/xllm --model=Qwen2-7B-Instruct \
+        ASCEND_RT_VISIBLE_DEVICES=0 /path/to/xllm --model=Qwen2-7B-Instruct \
                --port=8010 \
-               --devices="npu:0" \
                --master_node_addr="127.0.0.1:18888" \
                --enable_prefix_cache=false \
                --enable_chunked_prefill=false \
@@ -50,9 +49,8 @@ ENABLE_DECODE_RESPONSE_TO_SERVICE=true ./xllm_master_serving --etcd_addr="127.0.
         ```
     - 启动Decode实例
         ```bash
-        /path/to/xllm --model=Qwen2-7B-Instruct \
+        ASCEND_RT_VISIBLE_DEVICES=1 /path/to/xllm --model=Qwen2-7B-Instruct \
                --port=8020 \
-               --devices="npu:1" \
                --master_node_addr="127.0.0.1:18898" \
                --enable_prefix_cache=false \
                --enable_chunked_prefill=false \
diff --git a/docs/src/content/docs/zh/features/mtp.md b/docs/src/content/docs/zh/features/mtp.md
index eb06c21898..6df813375e 100644
--- a/docs/src/content/docs/zh/features/mtp.md
+++ b/docs/src/content/docs/zh/features/mtp.md
@@ -92,23 +92,19 @@ MODEL_PATH="/models/DeepSeek-V3"
 DRAFT_MODEL_PATH="/models/DeepSeek-V3-mtp"
 MASTER_NODE_ADDR="127.0.0.1:42123"
 START_PORT=13222
-START_DEVICE=0
 LOG_DIR="log"
 NNODES=16
 
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   nohup ./xllm \
     --model $MODEL_PATH \
-    --devices="npu:$DEVICE" \
     --port $PORT \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --draft_model $DRAFT_MODEL_PATH \
-    --draft_devices="npu:$DEVICE" \
     --num_speculative_tokens 1 \
     --max_memory_utilization=0.90 \
     --max_tokens_per_batch=10000 \
diff --git a/docs/src/content/docs/zh/getting_started/launch_xllm.md b/docs/src/content/docs/zh/getting_started/launch_xllm.md
index 2855124710..d46e881992 100644
--- a/docs/src/content/docs/zh/getting_started/launch_xllm.md
+++ b/docs/src/content/docs/zh/getting_started/launch_xllm.md
@@ -21,7 +21,6 @@ export HCCL_IF_BASE_PORT=43432  # HCCL 通信基础端口
 MODEL_PATH="/path/to/model/Qwen3-8B"               # 模型路径
 MASTER_NODE_ADDR="127.0.0.1:9748"                  # Master 节点地址（需全局一致）
 START_PORT=18000                                   # 服务起始端口
-START_DEVICE=0                                     # 起始逻辑设备号
 LOG_DIR="log"                                      # 日志目录
 NNODES=1                                           # 节点数（当前脚本启动 1 个进程）
 
@@ -30,11 +29,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="npu:$DEVICE" \
     --port $PORT \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
@@ -64,7 +61,6 @@ export CUDA_VISIBLE_DEVICES=0
 MODEL_PATH="/path/to/model/Qwen3-8B"
 MASTER_NODE_ADDR="127.0.0.1:9748"
 START_PORT=18000
-START_DEVICE=0
 LOG_DIR="log"
 NNODES=1
 
@@ -73,11 +69,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="cuda:$DEVICE" \
     --port $PORT \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
@@ -103,7 +97,6 @@ export MLU_VISIBLE_DEVICES=0
 MODEL_PATH="/path/to/model/Qwen3-8B"
 MASTER_NODE_ADDR="127.0.0.1:9748"
 START_PORT=18000
-START_DEVICE=0
 LOG_DIR="log"
 NNODES=1
 
@@ -112,11 +105,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="mlu:$DEVICE" \
     --port $PORT \
     --nnodes=$NNODES \
     --master_node_addr=$MASTER_NODE_ADDR \
@@ -138,7 +129,6 @@ export HIP_VISIBLE_DEVICES=0
 MODEL_PATH="/path/to/model/Qwen3-8B"
 MASTER_NODE_ADDR="127.0.0.1:9748"
 START_PORT=18000
-START_DEVICE=0
 LOG_DIR="log"
 NNODES=1
 
@@ -147,11 +137,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="dcu:$DEVICE" \
     --port $PORT \
     --nnodes=$NNODES \
     --master_node_addr=$MASTER_NODE_ADDR \
@@ -174,7 +162,6 @@ export FLASHINFER_OPS_PATH=/opt/conda/lib/python3.10/site-packages/flashinfer/da
 MODEL_PATH="/path/to/model/Qwen3-8B"
 MASTER_NODE_ADDR="127.0.0.1:9748"
 START_PORT=18000
-START_DEVICE=0
 LOG_DIR="log"
 NNODES=1
 
@@ -183,11 +170,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="cuda:$DEVICE" \
     --port $PORT \
     --nnodes=$NNODES \
     --master_node_addr=$MASTER_NODE_ADDR \
@@ -213,7 +198,6 @@ export MUSA_VISIBLE_DEVICES=0
 MODEL_PATH="/path/to/model/Qwen3.5-27B"
 MASTER_NODE_ADDR="127.0.0.1:9748"
 START_PORT=18000
-START_DEVICE=0
 LOG_DIR="log"
 NNODES=1
 
@@ -222,11 +206,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="musa:$DEVICE" \
     --port $PORT \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
diff --git a/docs/src/content/docs/zh/getting_started/multi_machine.md b/docs/src/content/docs/zh/getting_started/multi_machine.md
index 7eed9e114a..0683e95ac3 100644
--- a/docs/src/content/docs/zh/getting_started/multi_machine.md
+++ b/docs/src/content/docs/zh/getting_started/multi_machine.md
@@ -22,7 +22,6 @@ MODEL_PATH="/path/to/your/DeepSeek-R1"             # 模型路径
 MASTER_NODE_ADDR="123.123.123.123:9748"            # Master 节点地址（需全局一致）
 LOCAL_HOST=123.123.123.123                         # 本机服务启动IP
 START_PORT=18000                                   # 服务起始端口
-START_DEVICE=0                                     # 起始 NPU 逻辑设备号
 LOG_DIR="log"                                      # 日志目录
 LOCAL_NODES=16                                     # 单机节点数（当前脚本启动 16 个进程）
 NNODES=32                                          # 总卡数（该示例为2机32卡）
@@ -32,13 +31,11 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$LOCAL_NODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   /path/to/xllm \
     --model $MODEL_PATH \
     --host $LOCAL_HOST \
     --port $PORT \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --max_memory_utilization=0.86 \
@@ -74,7 +71,6 @@ MODEL_PATH="/path/to/your/DeepSeek-R1"             # 模型路径
 MASTER_NODE_ADDR="123.123.123.123:9748"            # Master 节点地址（需全局一致）
 LOCAL_HOST=456.456.456.456                         # 本机服务启动IP
 START_PORT=18000                                   # 服务起始端口
-START_DEVICE=0                                     # 起始 NPU 逻辑设备号
 LOG_DIR="log"                                      # 日志目录
 LOCAL_NODES=16                                     # 单机节点数（当前脚本启动 16 个进程）
 NNODES=32                                          # 总卡数（该示例为2机32卡）
@@ -84,13 +80,11 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$LOCAL_NODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   /path/to/xllm \
     --model $MODEL_PATH \
     --host $LOCAL_HOST \
     --port $PORT \
-    --devices="npu:$DEVICE" \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
     --max_memory_utilization=0.86 \
diff --git a/docs/src/content/docs/zh/getting_started/offline_service.md b/docs/src/content/docs/zh/getting_started/offline_service.md
index c9b9fbd9a2..3a9430dd28 100644
--- a/docs/src/content/docs/zh/getting_started/offline_service.md
+++ b/docs/src/content/docs/zh/getting_started/offline_service.md
@@ -16,7 +16,9 @@ LLM Beam Search 示例：[:simple-github: https://github.com/xLLM-AI/xllm/blob/m
 ```python
 from xllm import BeamSearchParams, LLM
 
-llm = LLM(model="/path/models/Qwen2-7B-Instruct", devices="npu:0")
+# 运行设备通过 ASCEND_RT_VISIBLE_DEVICES 选择
+# （其他平台使用 CUDA_VISIBLE_DEVICES / MLU_VISIBLE_DEVICES 等）。
+llm = LLM(model="/path/models/Qwen2-7B-Instruct")
 params = BeamSearchParams(
     beam_width=2,
     top_logprobs=4,
diff --git a/docs/src/content/docs/zh/hardware/ascend_npu.md b/docs/src/content/docs/zh/hardware/ascend_npu.md
index e87fecac9d..df9b744100 100644
--- a/docs/src/content/docs/zh/hardware/ascend_npu.md
+++ b/docs/src/content/docs/zh/hardware/ascend_npu.md
@@ -63,7 +63,6 @@ export HCCL_IF_BASE_PORT=43432  # HCCL 通信基础端口
 MODEL_PATH="/path/to/model/Qwen3-8B"               # 模型路径
 MASTER_NODE_ADDR="127.0.0.1:9748"                  # Master 节点地址（需全局一致）
 START_PORT=18000                                   # 服务起始端口
-START_DEVICE=0                                     # 起始逻辑设备号
 LOG_DIR="log"                                      # 日志目录
 NNODES=1                                           # 节点数（当前脚本启动 1 个进程）
 
@@ -72,11 +71,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="npu:$DEVICE" \
     --port $PORT \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
diff --git a/docs/src/content/docs/zh/hardware/cambricon_mlu.md b/docs/src/content/docs/zh/hardware/cambricon_mlu.md
index a8e4e57726..312377b0d9 100644
--- a/docs/src/content/docs/zh/hardware/cambricon_mlu.md
+++ b/docs/src/content/docs/zh/hardware/cambricon_mlu.md
@@ -38,7 +38,6 @@ export MLU_VISIBLE_DEVICES=0
 MODEL_PATH="/path/to/model/Qwen3-8B"
 MASTER_NODE_ADDR="127.0.0.1:9748"
 START_PORT=18000
-START_DEVICE=0
 LOG_DIR="log"
 NNODES=1
 
@@ -47,11 +46,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="mlu:$DEVICE" \
     --port $PORT \
     --nnodes=$NNODES \
     --master_node_addr=$MASTER_NODE_ADDR \
diff --git a/docs/src/content/docs/zh/hardware/dcu.md b/docs/src/content/docs/zh/hardware/dcu.md
index 5397b7f9a9..8720e9ecb1 100644
--- a/docs/src/content/docs/zh/hardware/dcu.md
+++ b/docs/src/content/docs/zh/hardware/dcu.md
@@ -50,7 +50,6 @@ export HIP_VISIBLE_DEVICES=0
 MODEL_PATH="/path/to/model/Qwen3-8B"
 MASTER_NODE_ADDR="127.0.0.1:9748"
 START_PORT=18000
-START_DEVICE=0
 LOG_DIR="log"
 NNODES=1
 
@@ -59,11 +58,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="dcu:$DEVICE" \
     --port $PORT \
     --nnodes=$NNODES \
     --master_node_addr=$MASTER_NODE_ADDR \
diff --git a/docs/src/content/docs/zh/hardware/metax_maca.md b/docs/src/content/docs/zh/hardware/metax_maca.md
index 4f9890a8bb..c9f53d53a0 100644
--- a/docs/src/content/docs/zh/hardware/metax_maca.md
+++ b/docs/src/content/docs/zh/hardware/metax_maca.md
@@ -53,7 +53,6 @@ export FLASHINFER_OPS_PATH=/opt/conda/lib/python3.10/site-packages/flashinfer/da
 MODEL_PATH="/path/to/model/Qwen3-8B"
 MASTER_NODE_ADDR="127.0.0.1:9748"
 START_PORT=18000
-START_DEVICE=0
 LOG_DIR="log"
 NNODES=1
 
@@ -62,11 +61,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="cuda:$DEVICE" \
     --port $PORT \
     --nnodes=$NNODES \
     --master_node_addr=$MASTER_NODE_ADDR \
diff --git a/docs/src/content/docs/zh/hardware/musa.md b/docs/src/content/docs/zh/hardware/musa.md
index 2a2fae5242..0b6b2a606c 100644
--- a/docs/src/content/docs/zh/hardware/musa.md
+++ b/docs/src/content/docs/zh/hardware/musa.md
@@ -5,7 +5,7 @@ sidebar:
   order: 7
 ---
 
-在摩尔线程（Mthreads）MUSA GPU 上部署 xLLM 时使用 MUSA 设备后端（`--devices=musa:<id>`）。
+在摩尔线程（Mthreads）MUSA GPU 上部署 xLLM 时使用 MUSA 设备后端；通过 `MUSA_VISIBLE_DEVICES` 选择可见的设备卡。
 
 ## 镜像和容器启动命令
 
@@ -49,7 +49,6 @@ export MUSA_VISIBLE_DEVICES=0
 MODEL_PATH="/path/to/model/Qwen3.5-27B"
 MASTER_NODE_ADDR="127.0.0.1:9748"
 START_PORT=18000
-START_DEVICE=0
 LOG_DIR="log"
 NNODES=1
 
@@ -58,11 +57,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="musa:$DEVICE" \
     --port $PORT \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
diff --git a/docs/src/content/docs/zh/hardware/nvidia_gpu.md b/docs/src/content/docs/zh/hardware/nvidia_gpu.md
index 120510fd72..52ea0b7bd8 100644
--- a/docs/src/content/docs/zh/hardware/nvidia_gpu.md
+++ b/docs/src/content/docs/zh/hardware/nvidia_gpu.md
@@ -46,7 +46,6 @@ export CUDA_VISIBLE_DEVICES=0
 MODEL_PATH="/path/to/model/Qwen3-8B"
 MASTER_NODE_ADDR="127.0.0.1:9748"
 START_PORT=18000
-START_DEVICE=0
 LOG_DIR="log"
 NNODES=1
 
@@ -55,11 +54,9 @@ mkdir -p $LOG_DIR
 for (( i=0; i<$NNODES; i++ ))
 do
   PORT=$((START_PORT + i))
-  DEVICE=$((START_DEVICE + i))
   LOG_FILE="$LOG_DIR/node_$i.log"
   xllm \
     --model $MODEL_PATH \
-    --devices="cuda:$DEVICE" \
     --port $PORT \
     --master_node_addr=$MASTER_NODE_ADDR \
     --nnodes=$NNODES \
diff --git a/examples/generate.py b/examples/generate.py
index 2a804e475d..55dfd3cf97 100644
--- a/examples/generate.py
+++ b/examples/generate.py
@@ -1,5 +1,5 @@
-# python examples/generate.py --model='/path/models/Qwen2-7B-Instruct' --devices='npu:0'
-# python generate.py --model='/path/models/Qwen2-7B-Instruct' --devices='npu:0,npu:1'
+# ASCEND_RT_VISIBLE_DEVICES=0 python examples/generate.py --model='/path/models/Qwen2-7B-Instruct'
+# ASCEND_RT_VISIBLE_DEVICES=0,1 python generate.py --model='/path/models/Qwen2-7B-Instruct'
 
 from xllm import ArgumentParser, LLM, SamplingParams
 
diff --git a/examples/generate_beam_search.py b/examples/generate_beam_search.py
index a39a606ff3..78baa3801f 100644
--- a/examples/generate_beam_search.py
+++ b/examples/generate_beam_search.py
@@ -1,5 +1,5 @@
-# python examples/generate_beam_search.py --model='/path/models/Qwen2-7B-Instruct' --devices='npu:0'
-# python generate_beam_search.py --model='/path/models/Qwen2-7B-Instruct' --devices='npu:0,npu:1'
+# ASCEND_RT_VISIBLE_DEVICES=0 python examples/generate_beam_search.py --model='/path/models/Qwen2-7B-Instruct'
+# ASCEND_RT_VISIBLE_DEVICES=0,1 python generate_beam_search.py --model='/path/models/Qwen2-7B-Instruct'
 
 from xllm import ArgumentParser, BeamSearchParams, LLM
 
diff --git a/examples/generate_embedding.py b/examples/generate_embedding.py
index e0e3c55c95..7a6354fcb9 100644
--- a/examples/generate_embedding.py
+++ b/examples/generate_embedding.py
@@ -1,5 +1,5 @@
-# python examples/generate_embedding.py --model='/path/models/Qwen3-8B' --devices='npu:0' --runner pooling
-# python generate_embedding.py --model='/path/models/Qwen3-8B' --devices='npu:0,npu:1'
+# ASCEND_RT_VISIBLE_DEVICES=0 python examples/generate_embedding.py --model='/path/models/Qwen3-8B' --runner pooling
+# ASCEND_RT_VISIBLE_DEVICES=0,1 python generate_embedding.py --model='/path/models/Qwen3-8B'
 
 from xllm import ArgumentParser, LLM, PoolingParams
 
diff --git a/examples/generate_vlm.py b/examples/generate_vlm.py
index 406d1591c6..05231be3f8 100755
--- a/examples/generate_vlm.py
+++ b/examples/generate_vlm.py
@@ -1,4 +1,4 @@
-# python generate_vlm.py --model /path/to/Qwen2.5-VL-7B-Instruct/ --max_seqs_per_batch 4 --devices='npu:0' --enable_shm
+# ASCEND_RT_VISIBLE_DEVICES=0 python generate_vlm.py --model /path/to/Qwen2.5-VL-7B-Instruct/ --max_seqs_per_batch 4 --enable_shm
 
 from xllm import ArgumentParser, SamplingParams
 from xllm import LLM
diff --git a/examples/sample.py b/examples/sample.py
index 5c6cab869b..7613f67026 100644
--- a/examples/sample.py
+++ b/examples/sample.py
@@ -1,5 +1,5 @@
-# python examples/sample.py --model='/path/models/Qwen3-8B' --devices='npu:0'
-# python examples/sample.py --model='/path/models/Qwen3-8B' --devices='npu:0,npu:1'
+# ASCEND_RT_VISIBLE_DEVICES=0 python examples/sample.py --model='/path/models/Qwen3-8B'
+# ASCEND_RT_VISIBLE_DEVICES=0,1 python examples/sample.py --model='/path/models/Qwen3-8B'
 
 from xllm import ArgumentParser, LLM, RequestParams
 
diff --git a/tests/core/util/device_name_utils_test.cpp b/tests/core/util/device_name_utils_test.cpp
index d0c4f7079c..093e71fe61 100644
--- a/tests/core/util/device_name_utils_test.cpp
+++ b/tests/core/util/device_name_utils_test.cpp
@@ -26,15 +26,9 @@ limitations under the License.
 namespace xllm {
 namespace {
 
-TEST(DeviceNameUtilsTest, ToDeviceStringUsesCompiledDeviceType) {
-  const std::string device_string = DeviceNameUtils::to_device_string(3);
-
-  EXPECT_EQ(device_string, Platform::type_str() + ":3");
-}
-
 TEST(DeviceNameUtilsTest, ParseGeneratedDeviceString) {
   const std::vector<torch::Device> devices =
-      DeviceNameUtils::parse_devices(DeviceNameUtils::to_device_string(2));
+      DeviceNameUtils::parse_devices(Platform::type_str() + ":2");
 
   ASSERT_EQ(devices.size(), 1);
   EXPECT_EQ(devices[0].type(), Platform::type_torch());
diff --git a/xllm/c_api/default.h b/xllm/c_api/default.h
index 4ee08e327e..fa735fbe81 100644
--- a/xllm/c_api/default.h
+++ b/xllm/c_api/default.h
@@ -58,7 +58,6 @@ const XLLM_InitOptions XLLM_INIT_LLM_OPTIONS_DEFAULT = {
     .kv_cache_transfer_mode = "PUSH",
     .log_dir = "",
     .draft_model = "",
-    .draft_devices = "",
     .cpu_affinity = ""};
 
 const XLLM_RequestParams XLLM_LLM_REQUEST_PARAMS_DEFAULT = {
@@ -132,7 +131,6 @@ const XLLM_InitOptions XLLM_INIT_REC_OPTIONS_DEFAULT = {
     .kv_cache_transfer_mode = "PUSH",
     .log_dir = "",
     .draft_model = "",
-    .draft_devices = "",
     .cpu_affinity = ""};
 
 const XLLM_RequestParams XLLM_REC_REQUEST_PARAMS_DEFAULT = {
diff --git a/xllm/c_api/examples/test_press_rec_multimodal_completions.cpp b/xllm/c_api/examples/test_press_rec_multimodal_completions.cpp
index 54a30e7693..5bc85309da 100644
--- a/xllm/c_api/examples/test_press_rec_multimodal_completions.cpp
+++ b/xllm/c_api/examples/test_press_rec_multimodal_completions.cpp
@@ -56,7 +56,6 @@ constexpr int32_t kMaxClientParallelism = 128;
 
 struct CliOptions {
   std::string model_path;
-  std::string devices = "cuda:0";
   std::string master_node_addr = "127.0.0.1:18899";
   int32_t prompt_size = 128;
   int32_t token_min_size = 1024;
@@ -737,7 +736,6 @@ void PrintSummary(const CliOptions& options,
   std::cout << "=== stress_rec_multimodal_completions summary ===\n";
   std::cout << "model_id=" << ResolveModelId(options.model_path) << "\n";
   std::cout << "model_type=" << config.model_type << "\n";
-  std::cout << "devices=" << options.devices << "\n";
   std::cout << "master_node_addr=" << options.master_node_addr << "\n";
   std::cout << "hidden_size=" << config.hidden_size
             << ", vocab_size=" << config.vocab_size
@@ -802,10 +800,8 @@ int main(int argc, char** argv) {
                   sizeof(init_options.master_node_addr),
                   "%s",
                   options.master_node_addr.c_str());
-    const bool init_ok = xllm_rec_initialize(rec_handler,
-                                             options.model_path.c_str(),
-                                             options.devices.c_str(),
-                                             &init_options);
+    const bool init_ok = xllm_rec_initialize(
+        rec_handler, options.model_path.c_str(), &init_options);
     if (!init_ok) {
       xllm_rec_destroy(rec_handler);
       throw std::runtime_error("xllm_rec_initialize failed");
diff --git a/xllm/c_api/examples/test_query_llm_chat_completions.cpp b/xllm/c_api/examples/test_query_llm_chat_completions.cpp
index 90ed928b85..aa7b193689 100644
--- a/xllm/c_api/examples/test_query_llm_chat_completions.cpp
+++ b/xllm/c_api/examples/test_query_llm_chat_completions.cpp
@@ -20,7 +20,6 @@ limitations under the License.
 
 #include "llm.h"
 
-std::string devices = "cuda:4";
 std::string model_name = "Qwen3-8B";
 std::string model_path = "/export/home/models/Qwen3-8B";
 
@@ -34,8 +33,8 @@ XLLM_LLM_Handler* service_startup_hook() {
   snprintf(
       init_options.log_dir, sizeof(init_options.log_dir), "/export/xllm/log");
 
-  bool ret = xllm_llm_initialize(
-      llm_handler, model_path.c_str(), devices.c_str(), &init_options);
+  bool ret =
+      xllm_llm_initialize(llm_handler, model_path.c_str(), &init_options);
   if (!ret) {
     std::cout << "LLM init failed" << std::endl;
     xllm_llm_destroy(llm_handler);
diff --git a/xllm/c_api/examples/test_query_llm_completions.cpp b/xllm/c_api/examples/test_query_llm_completions.cpp
index 70e56e9d30..5021238cef 100644
--- a/xllm/c_api/examples/test_query_llm_completions.cpp
+++ b/xllm/c_api/examples/test_query_llm_completions.cpp
@@ -20,7 +20,6 @@ limitations under the License.
 
 #include "llm.h"
 
-std::string devices = "cuda:1";
 std::string model_name = "Qwen3-8B";
 std::string model_path = "/export/home/models/Qwen3-8B";
 
@@ -35,8 +34,8 @@ XLLM_LLM_Handler* service_startup_hook() {
   snprintf(
       init_options.log_dir, sizeof(init_options.log_dir), "/export/xllm/log");
 
-  bool ret = xllm_llm_initialize(
-      llm_handler, model_path.c_str(), devices.c_str(), &init_options);
+  bool ret =
+      xllm_llm_initialize(llm_handler, model_path.c_str(), &init_options);
   if (!ret) {
     std::cout << "LLM init failed" << std::endl;
     xllm_llm_destroy(llm_handler);
diff --git a/xllm/c_api/examples/test_query_rec_completions.cpp b/xllm/c_api/examples/test_query_rec_completions.cpp
index a1926f9d12..78d7f67a8f 100644
--- a/xllm/c_api/examples/test_query_rec_completions.cpp
+++ b/xllm/c_api/examples/test_query_rec_completions.cpp
@@ -21,13 +21,6 @@ limitations under the License.
 
 #include "rec.h"
 
-#if defined(USE_NPU)
-std::string devices = "npu:14";
-#elif defined(USE_CUDA)
-std::string devices = "cuda:0";
-#else
-std::string devices = "npu:14";
-#endif
 std::string model_name = "Qwen3-0.6B";
 std::string model_path = "/export/home/models/Qwen3-0.6B";
 
@@ -54,8 +47,8 @@ XLLM_REC_Handler* service_startup_hook() {
   init_options.rec_worker_max_concurrency = 1;
 #endif
 
-  bool ret = xllm_rec_initialize(
-      rec_handler, model_path.c_str(), devices.c_str(), &init_options);
+  bool ret =
+      xllm_rec_initialize(rec_handler, model_path.c_str(), &init_options);
   if (!ret) {
     std::cout << "REC init failed" << std::endl;
     xllm_rec_destroy(rec_handler);
@@ -74,11 +67,10 @@ void service_stop_hook(XLLM_REC_Handler* rec_handler) {
 
 int main(int argc, char** argv) {
   if (argc > 1) {
-    devices = argv[1];
+    model_path = argv[1];
   }
 
   std::cout << "Using model path: " << model_path << std::endl;
-  std::cout << "Using devices: " << devices << std::endl;
 
   XLLM_REC_Handler* rec_handler = service_startup_hook();
   if (nullptr == rec_handler) {
diff --git a/xllm/c_api/examples/test_query_rec_multimodal_completions.cpp b/xllm/c_api/examples/test_query_rec_multimodal_completions.cpp
index 56d2bd323e..59127f9401 100644
--- a/xllm/c_api/examples/test_query_rec_multimodal_completions.cpp
+++ b/xllm/c_api/examples/test_query_rec_multimodal_completions.cpp
@@ -27,7 +27,6 @@ limitations under the License.
 
 #include "rec.h"
 
-std::string devices = "cuda:0";
 std::string model_name = "homepage_qwen_06b_6_raw";
 std::string model_path = "/export/home/models/homepage_qwen_06b_6_raw";
 
@@ -185,8 +184,8 @@ XLLM_REC_Handler* service_startup_hook() {
            sizeof(init_options.log_dir),
            "/export/home/huheng7/log");
 
-  bool ret = xllm_rec_initialize(
-      rec_handler, model_path.c_str(), devices.c_str(), &init_options);
+  bool ret =
+      xllm_rec_initialize(rec_handler, model_path.c_str(), &init_options);
   if (!ret) {
     std::cout << "REC init failed" << std::endl;
     xllm_rec_destroy(rec_handler);
diff --git a/xllm/c_api/internal/llm.cpp b/xllm/c_api/internal/llm.cpp
index 3c2ffd18bb..5683a52be8 100644
--- a/xllm/c_api/internal/llm.cpp
+++ b/xllm/c_api/internal/llm.cpp
@@ -59,9 +59,8 @@ XLLM_CAPI_EXPORT void xllm_llm_init_options_default(
 XLLM_CAPI_EXPORT bool xllm_llm_initialize(
     XLLM_LLM_Handler* handler,
     const char* model_path,
-    const char* devices,
     const XLLM_InitOptions* init_options) {
-  if (!handler || !model_path || !devices) return false;
+  if (!handler || !model_path) return false;
 
   try {
     XLLM_InitOptions xllm_init_options;
@@ -81,9 +80,7 @@ XLLM_CAPI_EXPORT bool xllm_llm_initialize(
     xllm::Options options;
     options.model_path(model_path)
         .task_type(xllm_init_options.task)
-        .devices(devices)
         .draft_model_path(xllm_init_options.draft_model)
-        .draft_devices(xllm_init_options.draft_devices)
         .backend("llm")
         .block_size(xllm_init_options.block_size)
         .max_cache_size(xllm_init_options.max_cache_size)
diff --git a/xllm/c_api/internal/rec.cpp b/xllm/c_api/internal/rec.cpp
index bc0348c941..6048a63b17 100644
--- a/xllm/c_api/internal/rec.cpp
+++ b/xllm/c_api/internal/rec.cpp
@@ -66,9 +66,8 @@ XLLM_CAPI_EXPORT void xllm_rec_init_options_default(
 XLLM_CAPI_EXPORT bool xllm_rec_initialize(
     XLLM_REC_Handler* handler,
     const char* model_path,
-    const char* devices,
     const XLLM_InitOptions* init_options) {
-  if (!handler || !model_path || !devices) return false;
+  if (!handler || !model_path) return false;
 
   try {
     xllm::CpuAffinity::get_instance().set_cpu_affinity(
@@ -91,9 +90,7 @@ XLLM_CAPI_EXPORT bool xllm_rec_initialize(
     xllm::Options options;
     options.model_path(model_path)
         .task_type(xllm_init_options.task)
-        .devices(devices)
         .draft_model_path(xllm_init_options.draft_model)
-        .draft_devices(xllm_init_options.draft_devices)
         .backend("rec")
         .block_size(xllm_init_options.block_size)
         .max_cache_size(xllm_init_options.max_cache_size)
diff --git a/xllm/c_api/llm.h b/xllm/c_api/llm.h
index 5460ae0da5..72d321c14c 100644
--- a/xllm/c_api/llm.h
+++ b/xllm/c_api/llm.h
@@ -82,6 +82,8 @@ XLLM_CAPI_EXPORT void xllm_llm_init_options_default(
  * Loads model weights from the specified path, configures target devices,
  * initializes compute contexts, and prepares the inference runtime.
  * Must be called exactly once per handler before using completion/chat APIs.
+ * The target devices are auto-detected from the visible device mask
+ * (e.g. ASCEND_RT_VISIBLE_DEVICES / CUDA_VISIBLE_DEVICES).
  *
  * If init_options is NULL, this function automatically uses the default values
  * from XLLM_INIT_LLM_OPTIONS_DEFAULT (via xllm_llm_init_options_default()).
@@ -89,9 +91,6 @@ XLLM_CAPI_EXPORT void xllm_llm_init_options_default(
  * @param handler Valid LLM instance handle (must not be NULL)
  * @param model_path Null-terminated string of the model directory/file path
  *                   (supports .bin/.pth/.safetensors formats)
- * @param devices Null-terminated string specifying target devices (format:
- *                "npu:0,1" (specific NPUs), "cuda:0" (single GPU), "auto"
- * (automatic selection))
  * @param init_options Advanced initialization options (NULL = use defaults)
  *
  * @return true if initialization succeeds; false on failure (see failure causes
@@ -100,7 +99,6 @@ XLLM_CAPI_EXPORT void xllm_llm_init_options_default(
  * @failure_causes
  * - Invalid handler (NULL or already destroyed)
  * - Invalid model_path (non-existent, corrupted, or unsupported format)
- * - Invalid devices string (malformed format or unavailable devices)
  * - Model load error (mismatched model architecture or weight corruption)
  * - Device initialization failure (out of memory, driver error)
  *
@@ -109,7 +107,6 @@ XLLM_CAPI_EXPORT void xllm_llm_init_options_default(
  */
 XLLM_CAPI_EXPORT bool xllm_llm_initialize(XLLM_LLM_Handler* handler,
                                           const char* model_path,
-                                          const char* devices,
                                           const XLLM_InitOptions* init_options);
 
 /**
diff --git a/xllm/c_api/rec.h b/xllm/c_api/rec.h
index ef795006f6..2d8fa73657 100644
--- a/xllm/c_api/rec.h
+++ b/xllm/c_api/rec.h
@@ -75,13 +75,12 @@ XLLM_CAPI_EXPORT void xllm_rec_init_options_default(
  * @brief Initialize the Generative Recommendation (REC) model and runtime
  * environment Loads generative recommendation model weights from the specified
  * path, configures target devices, initializes compute contexts, and prepares
- * the recommendation inference runtime
+ * the recommendation inference runtime. The target devices are auto-detected
+ * from the visible device mask (e.g. ASCEND_RT_VISIBLE_DEVICES /
+ * CUDA_VISIBLE_DEVICES).
  * @param handler Valid REC inference instance handle (must not be NULL)
  * @param model_path Null-terminated string of the REC model directory/file path
  *                   (supports .bin/.pth/.safetensors formats with ranking head)
- * @param devices Null-terminated string specifying target devices (format:
- *                "npu:0,1" (specific NPUs), "cuda:0" (single GPU), "auto"
- * (automatic selection))
  * @param init_options Advanced initialization options (NULL = use REC defaults)
  * @return true if initialization succeeds; false on failure (see failure causes
  * below)
@@ -89,7 +88,6 @@ XLLM_CAPI_EXPORT void xllm_rec_init_options_default(
  * - Invalid handler (NULL or already destroyed)
  * - Invalid model_path (non-existent, corrupted, or missing ranking head
  * weights)
- * - Invalid devices string (malformed format or unavailable devices)
  * - Model load error (mismatched REC model architecture or embedding table
  * corruption)
  * - Device initialization failure (out of memory, driver error, insufficient
@@ -99,7 +97,6 @@ XLLM_CAPI_EXPORT void xllm_rec_init_options_default(
  */
 XLLM_CAPI_EXPORT bool xllm_rec_initialize(XLLM_REC_Handler* handler,
                                           const char* model_path,
-                                          const char* devices,
                                           const XLLM_InitOptions* init_options);
 
 /**
diff --git a/xllm/c_api/test/utils.cpp b/xllm/c_api/test/utils.cpp
index e0724df61b..a8e58f9536 100644
--- a/xllm/c_api/test/utils.cpp
+++ b/xllm/c_api/test/utils.cpp
@@ -23,7 +23,6 @@ limitations under the License.
 
 // --- Server + XLLM_InitOptions gflags (defaults aligned with REC defaults) ---
 DEFINE_string(model_path, "", "Path to REC model weights");
-DEFINE_string(devices, "auto", "Devices string, e.g. npu:0 or auto");
 DEFINE_int32(port, 8000, "brpc TCP port");
 DEFINE_string(listen_addr,
               "",
@@ -74,7 +73,6 @@ DEFINE_string(kv_cache_transfer_mode, "PUSH", "");
 // Not named "log_dir": glog already registers FLAGS_log_dir.
 DEFINE_string(xllm_init_log_dir, "", "");
 DEFINE_string(draft_model, "", "");
-DEFINE_string(draft_devices, "", "");
 
 namespace xllm_capi_test {
 
@@ -192,8 +190,6 @@ void ApplyGflagsToXllmInitOptions(XLLM_InitOptions* o) {
       o->log_dir, FLAGS_xllm_init_log_dir, XLLM_META_STRING_FIELD_MAX_LEN);
   CopyToFixed(
       o->draft_model, FLAGS_draft_model, XLLM_META_STRING_FIELD_MAX_LEN);
-  CopyToFixed(
-      o->draft_devices, FLAGS_draft_devices, XLLM_META_STRING_FIELD_MAX_LEN);
 }
 
 void PbToXllmDims(const c_api_test::XLLM_Dims& pb, XLLM_Dims* out) {
diff --git a/xllm/c_api/test/utils.h b/xllm/c_api/test/utils.h
index a875e733f9..5246291c76 100644
--- a/xllm/c_api/test/utils.h
+++ b/xllm/c_api/test/utils.h
@@ -26,7 +26,6 @@ limitations under the License.
 #include "xllm_test.pb.h"
 
 DECLARE_string(model_path);
-DECLARE_string(devices);
 DECLARE_int32(port);
 DECLARE_string(listen_addr);
 DECLARE_int32(idle_timeout_s);
diff --git a/xllm/c_api/test/xllm_test.cpp b/xllm/c_api/test/xllm_test.cpp
index b0d0a608d9..8d2873b697 100644
--- a/xllm/c_api/test/xllm_test.cpp
+++ b/xllm/c_api/test/xllm_test.cpp
@@ -259,12 +259,10 @@ int main(int argc, char* argv[]) {
     XLLM_InitOptions init{};
     xllm_rec_init_options_default(&init);
     xllm_capi_test::ApplyGflagsToXllmInitOptions(&init);
-    if (!xllm_rec_initialize(rec_holder.get(),
-                             FLAGS_model_path.c_str(),
-                             FLAGS_devices.c_str(),
-                             &init)) {
-      LOG(ERROR) << "xllm_rec_initialize failed model_path=" << FLAGS_model_path
-                 << " devices=" << FLAGS_devices;
+    if (!xllm_rec_initialize(
+            rec_holder.get(), FLAGS_model_path.c_str(), &init)) {
+      LOG(ERROR) << "xllm_rec_initialize failed model_path="
+                 << FLAGS_model_path;
       return -4;
     }
     rec_ptr = rec_holder.get();
@@ -277,12 +275,10 @@ int main(int argc, char* argv[]) {
     XLLM_InitOptions init{};
     xllm_llm_init_options_default(&init);
     xllm_capi_test::ApplyGflagsToXllmInitOptions(&init);
-    if (!xllm_llm_initialize(llm_holder.get(),
-                             FLAGS_model_path.c_str(),
-                             FLAGS_devices.c_str(),
-                             &init)) {
-      LOG(ERROR) << "xllm_llm_initialize failed model_path=" << FLAGS_model_path
-                 << " devices=" << FLAGS_devices;
+    if (!xllm_llm_initialize(
+            llm_holder.get(), FLAGS_model_path.c_str(), &init)) {
+      LOG(ERROR) << "xllm_llm_initialize failed model_path="
+                 << FLAGS_model_path;
       return -6;
     }
     llm_ptr = llm_holder.get();
diff --git a/xllm/c_api/types.h b/xllm/c_api/types.h
index d4a9c2ce01..ad2e4e7a2e 100644
--- a/xllm/c_api/types.h
+++ b/xllm/c_api/types.h
@@ -178,12 +178,6 @@ typedef struct XLLM_CAPI_EXPORT XLLM_InitOptions {
   /** Draft hf model path (empty string = no draft model) */
   char draft_model[XLLM_META_STRING_FIELD_MAX_LEN];
 
-  /**
-   * Devices to run the draft model on (e.g. npu:0, npu:0,npu:1).
-   * Empty string = use the same devices as main model
-   */
-  char draft_devices[XLLM_META_STRING_FIELD_MAX_LEN];
-
   // CPU affinity settings, format be like:
   // 1,2,3
   // meaing ith thread attach to logical processor 1,2,3
diff --git a/xllm/cc_api/examples/multiple_llm_instances.cpp b/xllm/cc_api/examples/multiple_llm_instances.cpp
index eb460e8fc4..49814cb563 100644
--- a/xllm/cc_api/examples/multiple_llm_instances.cpp
+++ b/xllm/cc_api/examples/multiple_llm_instances.cpp
@@ -26,7 +26,6 @@ limitations under the License.
  * multiple models needs to be supported, then you can follow this example.
  */
 
-std::string devices = "npu:0";
 std::string model_path = "/export/home/models/Qwen3-4B";
 
 int main(int argc, char** argv) {
@@ -35,7 +34,7 @@ int main(int argc, char** argv) {
   std::shared_ptr<xllm::LLM> llm_instance_01 = std::make_shared<xllm::LLM>();
   xllm::XLLM_InitLLMOptions options_01;
   options_01.max_memory_utilization = 0.45;
-  bool ret = llm_instance_01->Initialize(model_path, devices, options_01);
+  bool ret = llm_instance_01->Initialize(model_path, options_01);
   if (!ret) {
     std::cout << "LLM instance init failed." << std::endl;
     return -1;
@@ -60,7 +59,7 @@ int main(int argc, char** argv) {
   options_02.server_idx = 1;
   options_02.max_memory_utilization = 0.9;
 
-  ret = llm_instance_02->Initialize(model_path, devices, options_02);
+  ret = llm_instance_02->Initialize(model_path, options_02);
   if (!ret) {
     std::cout << "LLM instance init failed." << std::endl;
     return -1;
diff --git a/xllm/cc_api/examples/single_llm_instance.cpp b/xllm/cc_api/examples/single_llm_instance.cpp
index 8490d4fdd3..b3bbf45013 100644
--- a/xllm/cc_api/examples/single_llm_instance.cpp
+++ b/xllm/cc_api/examples/single_llm_instance.cpp
@@ -25,14 +25,13 @@ limitations under the License.
  * inference engine.
  */
 
-std::string devices = "npu:0";
 std::string model_path = "/export/home/models/Qwen3-4B";
 
 int main(int argc, char** argv) {
   xllm::LLM llm_instance;
   xllm::XLLM_InitLLMOptions options;
   options.cp_size = 1;
-  bool ret = llm_instance.Initialize(model_path, devices, options);
+  bool ret = llm_instance.Initialize(model_path, options);
   if (!ret) {
     std::cout << "LLM init failed." << std::endl;
     return -1;
diff --git a/xllm/cc_api/llm.cpp b/xllm/cc_api/llm.cpp
index e6c45dc8f6..37d59a3b6d 100644
--- a/xllm/cc_api/llm.cpp
+++ b/xllm/cc_api/llm.cpp
@@ -59,7 +59,6 @@ LLM::~LLM() {
 }
 
 bool LLM::Initialize(const std::string& model_path,
-                     const std::string& devices,
                      const XLLM_InitLLMOptions& init_options) {
   if (!init_options.log_dir.empty()) {
     InitGlog(init_options.log_dir);
@@ -74,9 +73,7 @@ bool LLM::Initialize(const std::string& model_path,
     Options options;
     options.model_path(model_path)
         .task_type(init_options.task)
-        .devices(devices)
         .draft_model_path(init_options.draft_model)
-        .draft_devices(init_options.draft_devices)
         .backend("llm")
         .block_size(init_options.block_size)
         .max_cache_size(init_options.max_cache_size)
diff --git a/xllm/cc_api/llm.h b/xllm/cc_api/llm.h
index 541ab3c110..9e6ab32646 100644
--- a/xllm/cc_api/llm.h
+++ b/xllm/cc_api/llm.h
@@ -44,16 +44,14 @@ class XLLM_CAPI_EXPORT LLM {
    * @brief Initialize the model: Load model files and configure runtime
    * environment
    * @param model_path Path to model files
-   * @param devices Device configuration (format: "npu:1" for specific NPU,
-   * "auto" for auto-selection)
    * @param init_options Advanced initialization options, Provided default
    * configuration
    * @return bool true if initialization succeeds; false if fails
    * @note Must be called before Completions/ChatCompletions, and only needs to
-   * be called once
+   * be called once. The devices to run on are auto-detected from the visible
+   * device mask (e.g. ASCEND_RT_VISIBLE_DEVICES / CUDA_VISIBLE_DEVICES).
    */
   bool Initialize(const std::string& model_path,
-                  const std::string& devices,
                   const XLLM_InitLLMOptions& init_options);
 
   /**
diff --git a/xllm/cc_api/types.h b/xllm/cc_api/types.h
index f690c82a38..d8ea2b9145 100644
--- a/xllm/cc_api/types.h
+++ b/xllm/cc_api/types.h
@@ -137,9 +137,6 @@ struct XLLM_CAPI_EXPORT XLLM_InitLLMOptions {
   // draft hf model path to the model file
   std::optional<std::string> draft_model = std::nullopt;
 
-  // Devices to run the draft model on, e.g. npu:0, npu:0,npu:1
-  std::optional<std::string> draft_devices = std::nullopt;
-
   // Index ID for internal server ID, which must be set different values
   // if the model supports multiple version or there are multiple models.
   int64_t server_idx = 0;
diff --git a/xllm/core/common/global_flags.h b/xllm/core/common/global_flags.h
index ecc6a13d03..a9d65171fd 100755
--- a/xllm/core/common/global_flags.h
+++ b/xllm/core/common/global_flags.h
@@ -43,10 +43,6 @@ DECLARE_string(python_model_path);
 
 DECLARE_string(task);
 
-DECLARE_string(devices);
-
-DECLARE_int32(device_id);
-
 DECLARE_int32(limit_image_per_prompt);
 
 DECLARE_string(mm_download_headers);
@@ -76,8 +72,6 @@ DECLARE_int32(max_tokens_per_chunk_for_prefill);
 // --- speculative inference config ---
 DECLARE_string(draft_model);
 
-DECLARE_string(draft_devices);
-
 DECLARE_int32(num_speculative_tokens);
 
 DECLARE_string(speculative_algorithm);
diff --git a/xllm/core/common/options.cpp b/xllm/core/common/options.cpp
index 7bbc65a030..bfe237b1f3 100644
--- a/xllm/core/common/options.cpp
+++ b/xllm/core/common/options.cpp
@@ -20,9 +20,7 @@ std::string Options::to_string() const {
   std::stringstream ss;
   ss << "Options: [";
   ss << "model_path: " << model_path()
-     << ", devices: " << devices().value_or("null")
      << ", draft_model_path: " << draft_model_path().value_or("null")
-     << ", draft_devices: " << draft_devices().value_or("null")
      << ", backend: " << backend()
      << ", limit_image_per_prompt: " << limit_image_per_prompt()
      << ", block_size: " << block_size()
diff --git a/xllm/core/common/options.h b/xllm/core/common/options.h
index 63deb6116a..e1c56661fc 100644
--- a/xllm/core/common/options.h
+++ b/xllm/core/common/options.h
@@ -39,12 +39,8 @@ class Options {
 
   PROPERTY(std::string, model_id);
 
-  PROPERTY(std::optional<std::string>, devices);
-
   PROPERTY(std::optional<std::string>, draft_model_path);
 
-  PROPERTY(std::optional<std::string>, draft_devices);
-
   // model backend
   PROPERTY(std::string, backend);
 
diff --git a/xllm/core/distributed_runtime/master.cpp b/xllm/core/distributed_runtime/master.cpp
index 0012e6b15e..b1743ddde2 100644
--- a/xllm/core/distributed_runtime/master.cpp
+++ b/xllm/core/distributed_runtime/master.cpp
@@ -199,10 +199,19 @@ Master::Master(const Options& options, EngineType type)
       master_status_(options.master_status()) {
   const auto model_path =
       std::filesystem::path(options_.model_path()).lexically_normal();
-  const auto devices =
-      DeviceNameUtils::parse_devices(options_.devices().value_or("auto"));
-  const int32_t global_world_size =
-      options_.nnodes() * static_cast<int32_t>(devices.size());
+  // Multi-process serving runs one worker per process. Enumerate the visible
+  // cards (honoring *_VISIBLE_DEVICES) and select the single card this process
+  // owns by its node_rank, so `devices` holds exactly the card this process
+  // uses -- mirroring the historical single-element devices semantics.
+  const auto visible_devices = DeviceNameUtils::parse_devices("auto");
+  CHECK_LT(options_.node_rank(), static_cast<int32_t>(visible_devices.size()))
+      << "node_rank " << options_.node_rank()
+      << " exceeds the number of visible devices " << visible_devices.size()
+      << ". Ensure *_VISIBLE_DEVICES exposes all cards used across processes.";
+  const std::vector<torch::Device> devices = {
+      visible_devices[options_.node_rank()]};
+  // World size is the node count (one worker per process).
+  const int32_t global_world_size = options_.nnodes();
   std::string cp_model_type;
   if (options_.cp_size() > 1 && Platform::uses_model_cp_partition()) {
     cp_model_type = util::get_model_type(model_path, options_.backend());
@@ -341,8 +350,8 @@ Master::Master(const Options& options, EngineType type)
     const bool use_suffix_spec = options_.speculative_algorithm() == "Suffix";
     CHECK(use_suffix_spec || !draft_model_path.empty())
         << "draft model path is required unless --speculative_algorithm=Suffix";
-    const auto draft_devices = DeviceNameUtils::parse_devices(
-        options_.draft_devices().value_or("auto"));
+    // Draft model shares the same devices as the target model.
+    const auto& draft_devices = devices;
     LOG(INFO) << "Using draft devices: "
               << DeviceNameUtils::to_string(draft_devices);
     runtime::Options spec_options;
diff --git a/xllm/core/framework/config/model_config.cpp b/xllm/core/framework/config/model_config.cpp
index 66bd75a3e3..c4719a75c0 100644
--- a/xllm/core/framework/config/model_config.cpp
+++ b/xllm/core/framework/config/model_config.cpp
@@ -19,7 +19,6 @@ limitations under the License.
 
 #include "core/common/global_flags.h"
 #include "core/framework/config/config_utils.h"
-#include "core/platform/device_name_utils.h"
 
 DEFINE_string(model_id, "", "hf model name.");
 
@@ -48,13 +47,6 @@ DEFINE_string(task,
               "generate",
               "The task to use the model for(e.g. generate, embed, mm_embed).");
 
-DEFINE_string(devices,
-              "",
-              "Deprecated. Use --device_id instead. Devices to run the model "
-              "on, e.g. npu:0, npu:0,npu:1.");
-
-DEFINE_int32(device_id, -1, "Device id to run the model on, e.g. 0.");
-
 DEFINE_int32(limit_image_per_prompt,
              8,
              "Maximum number of image per prompt. Only applicable for "
@@ -123,19 +115,6 @@ void ModelConfig::from_flags() {
   XLLM_CONFIG_ASSIGN_FROM_FLAG(python_model_path);
   XLLM_CONFIG_ASSIGN_FROM_FLAG(backend);
   XLLM_CONFIG_ASSIGN_FROM_FLAG(task);
-  XLLM_CONFIG_ASSIGN_FROM_FLAG(device_id);
-  const bool devices_specified = config::is_flag_specified("devices");
-  const bool device_id_specified = config::is_flag_specified("device_id");
-  if (devices_specified) {
-    LOG(WARNING) << "--devices is deprecated and will be removed in a future "
-                    "release. Use --device_id instead.";
-  }
-  if (devices_specified && !device_id_specified) {
-    XLLM_CONFIG_ASSIGN_FROM_FLAG(devices);
-  } else {
-    CHECK(device_id() >= 0) << "--device_id must be >= 0.";
-    devices(DeviceNameUtils::to_device_string(device_id()));
-  }
   XLLM_CONFIG_ASSIGN_FROM_FLAG(limit_image_per_prompt);
   XLLM_CONFIG_ASSIGN_FROM_FLAG(max_encoder_cache_size);
   XLLM_CONFIG_ASSIGN_FROM_FLAG(reasoning_parser);
@@ -176,9 +155,6 @@ void ModelConfig::from_json(const JsonReader& json) {
   XLLM_CONFIG_ASSIGN_FROM_JSON(python_model_path);
   XLLM_CONFIG_ASSIGN_FROM_JSON(backend);
   XLLM_CONFIG_ASSIGN_FROM_JSON(task);
-  // don't read rank-related config
-  // XLLM_CONFIG_ASSIGN_FROM_JSON(device_id);
-  // XLLM_CONFIG_ASSIGN_FROM_JSON(devices);
   XLLM_CONFIG_ASSIGN_FROM_JSON(limit_image_per_prompt);
   XLLM_CONFIG_ASSIGN_FROM_JSON(max_encoder_cache_size);
   XLLM_CONFIG_ASSIGN_FROM_JSON(reasoning_parser);
@@ -203,11 +179,6 @@ void ModelConfig::append_config_json(
       config_json, default_config, python_model_path);
   APPEND_CONFIG_JSON_VALUE_IF_NOT_DEFAULT(config_json, default_config, backend);
   APPEND_CONFIG_JSON_VALUE_IF_NOT_DEFAULT(config_json, default_config, task);
-  // don't dump rank-related config
-  //  APPEND_CONFIG_JSON_VALUE_IF_NOT_DEFAULT(config_json, default_config,
-  //  device_id);
-  //  APPEND_CONFIG_JSON_VALUE_IF_NOT_DEFAULT(config_json, default_config,
-  //  devices);
   APPEND_CONFIG_JSON_VALUE_IF_NOT_DEFAULT(
       config_json, default_config, limit_image_per_prompt);
   APPEND_CONFIG_JSON_VALUE_IF_NOT_DEFAULT(
diff --git a/xllm/core/framework/config/model_config.h b/xllm/core/framework/config/model_config.h
index f2a4f29b42..8566de744c 100644
--- a/xllm/core/framework/config/model_config.h
+++ b/xllm/core/framework/config/model_config.h
@@ -50,8 +50,6 @@ class ModelConfig final {
          "model_impl",
          "backend",
          "task",
-         "device_id",
-         "devices",
          "python_model_path",
          "limit_image_per_prompt",
          "max_encoder_cache_size",
@@ -78,10 +76,6 @@ class ModelConfig final {
 
   PROPERTY(std::string, task) = "generate";
 
-  PROPERTY(int32_t, device_id) = -1;
-
-  PROPERTY(std::string, devices) = "";
-
   PROPERTY(int32_t, limit_image_per_prompt) = 8;
 
   PROPERTY(int64_t, max_encoder_cache_size) = 0;
diff --git a/xllm/core/framework/config/speculative_config.cpp b/xllm/core/framework/config/speculative_config.cpp
index d74b498e0b..beb20e93ce 100644
--- a/xllm/core/framework/config/speculative_config.cpp
+++ b/xllm/core/framework/config/speculative_config.cpp
@@ -22,12 +22,6 @@ limitations under the License.
 
 DEFINE_string(draft_model, "", "draft hf model path to the model file.");
 
-DEFINE_string(draft_devices,
-              "",
-              "Devices to run the draft model on, e.g. npu:0, npu:0,npu:1. "
-              "If omitted, uses the target model devices when speculative "
-              "decoding is enabled.");
-
 DEFINE_int32(num_speculative_tokens, 0, "Number of speculative tokens.");
 
 DEFINE_string(speculative_algorithm,
@@ -80,11 +74,6 @@ namespace xllm {
 
 void SpeculativeConfig::from_flags() {
   XLLM_CONFIG_ASSIGN_FROM_FLAG(draft_model);
-  if (config::is_flag_specified("draft_devices")) {
-    LOG(WARNING) << "--draft_devices is deprecated and will be removed in a "
-                    "future release. Because it's same as --devices.";
-  }
-  XLLM_CONFIG_ASSIGN_FROM_FLAG(draft_devices);
   XLLM_CONFIG_ASSIGN_FROM_FLAG(num_speculative_tokens);
   XLLM_CONFIG_ASSIGN_FROM_FLAG(speculative_algorithm);
   XLLM_CONFIG_ASSIGN_FROM_FLAG(speculative_suffix_cache_max_depth);
@@ -100,8 +89,6 @@ void SpeculativeConfig::from_flags() {
 
 void SpeculativeConfig::from_json(const JsonReader& json) {
   XLLM_CONFIG_ASSIGN_FROM_JSON(draft_model);
-  // don't read rank-related config
-  // XLLM_CONFIG_ASSIGN_FROM_JSON(draft_devices);
   XLLM_CONFIG_ASSIGN_FROM_JSON(num_speculative_tokens);
   XLLM_CONFIG_ASSIGN_FROM_JSON(speculative_algorithm);
   XLLM_CONFIG_ASSIGN_FROM_JSON(speculative_suffix_cache_max_depth);
@@ -120,9 +107,6 @@ void SpeculativeConfig::append_config_json(
   const SpeculativeConfig default_config;
   APPEND_CONFIG_JSON_VALUE_IF_NOT_DEFAULT(
       config_json, default_config, draft_model);
-  // don't dump rank-related config
-  //  APPEND_CONFIG_JSON_VALUE_IF_NOT_DEFAULT(
-  //      config_json, default_config, draft_devices);
   APPEND_CONFIG_JSON_VALUE_IF_NOT_DEFAULT(
       config_json, default_config, num_speculative_tokens);
   APPEND_CONFIG_JSON_VALUE_IF_NOT_DEFAULT(
diff --git a/xllm/core/framework/config/speculative_config.h b/xllm/core/framework/config/speculative_config.h
index 54c98b9db9..b259236b52 100644
--- a/xllm/core/framework/config/speculative_config.h
+++ b/xllm/core/framework/config/speculative_config.h
@@ -55,7 +55,6 @@ class SpeculativeConfig final {
     static const OptionCategory kOptionCategory = {
         "SPECULATIVE OPTIONS",
         {"draft_model",
-         "draft_devices",
          "num_speculative_tokens",
          "speculative_algorithm",
          "speculative_suffix_cache_max_depth",
@@ -72,8 +71,6 @@ class SpeculativeConfig final {
 
   PROPERTY(std::string, draft_model);
 
-  PROPERTY(std::string, draft_devices) = "";
-
   PROPERTY(int32_t, num_speculative_tokens) = 0;
 
   PROPERTY(std::string, speculative_algorithm) = "MTP";
diff --git a/xllm/core/platform/device_name_utils.cpp b/xllm/core/platform/device_name_utils.cpp
index 6efdb523dc..6b06324875 100644
--- a/xllm/core/platform/device_name_utils.cpp
+++ b/xllm/core/platform/device_name_utils.cpp
@@ -29,10 +29,6 @@ limitations under the License.
 
 namespace xllm {
 
-std::string DeviceNameUtils::to_device_string(int32_t device_id) {
-  return Platform::type_str() + ":" + std::to_string(device_id);
-}
-
 std::vector<torch::Device> DeviceNameUtils::parse_devices(
     const std::string& device_str) {
   std::vector<torch::Device> devices;
diff --git a/xllm/core/platform/device_name_utils.h b/xllm/core/platform/device_name_utils.h
index ca4c0aa872..3adc3849ab 100644
--- a/xllm/core/platform/device_name_utils.h
+++ b/xllm/core/platform/device_name_utils.h
@@ -27,8 +27,6 @@ namespace xllm {
 
 class DeviceNameUtils {
  public:
-  static std::string to_device_string(int32_t device_id);
-
   static std::vector<torch::Device> parse_devices(
       const std::string& device_str);
 
diff --git a/xllm/core/platform/platform.cpp b/xllm/core/platform/platform.cpp
index 6503a60ebc..a12d3910e4 100644
--- a/xllm/core/platform/platform.cpp
+++ b/xllm/core/platform/platform.cpp
@@ -84,6 +84,20 @@ int32_t Platform::device_count() {
 #endif
 }
 
+int32_t Platform::current_device() {
+#if defined(USE_NPU)
+  return static_cast<int32_t>(c10_npu::current_device());
+#elif defined(USE_MLU)
+  return static_cast<int32_t>(torch_mlu::current_device());
+#elif defined(USE_CUDA) || defined(USE_ILU)
+  return static_cast<int32_t>(c10::cuda::current_device());
+#elif defined(USE_MUSA)
+  return static_cast<int32_t>(c10::musa::current_device());
+#elif defined(USE_DCU)
+  return static_cast<int32_t>(c10::hip::current_device());
+#endif
+}
+
 bool Platform::is_enable_pdl() { return enable_pdl_; }
 
 int32_t Platform::sm_count() { return sm_count_; }
diff --git a/xllm/core/platform/platform.h b/xllm/core/platform/platform.h
index c2a29ea7e3..67c2ed77c6 100644
--- a/xllm/core/platform/platform.h
+++ b/xllm/core/platform/platform.h
@@ -81,6 +81,9 @@ class Platform final {
   }
 
   static int32_t device_count();
+  // Returns the logical index of the device bound to the current thread.
+  // Valid only after the device has been set (e.g. via Device::set_device).
+  static int32_t current_device();
   static int32_t sm_count();
   static bool is_enable_pdl();
   static bool is_support_sm90a();
diff --git a/xllm/launch_server.py b/xllm/launch_server.py
index 142b47beb7..c4c5c0c888 100644
--- a/xllm/launch_server.py
+++ b/xllm/launch_server.py
@@ -100,14 +100,6 @@ def _build_parser() -> argparse.ArgumentParser:
         default=None,
         help="Launch only this rank. If omitted, local ranks 0..nnodes-1 are launched.",
     )
-    parser.add_argument(
-        "--start-device-id",
-        "--start_device_id",
-        dest="start_device_id",
-        type=int,
-        default=0,
-        help="Base logical device id. Local multi-rank launch uses id + rank.",
-    )
     parser.add_argument(
         "--log-dir",
         "--log_dir",
@@ -195,8 +187,6 @@ def _validate_args(parser: argparse.ArgumentParser, args: argparse.Namespace) ->
         parser.error("--nnodes must be greater than 0")
     if args.start_port < 1 or args.start_port > 65535:
         parser.error("--port/--start-port must be in range [1, 65535]")
-    if args.start_device_id < 0:
-        parser.error("--start-device-id must be greater than or equal to 0")
     if args.node_rank is not None and (
         args.node_rank < 0 or args.node_rank >= args.nnodes
     ):
@@ -224,15 +214,6 @@ def _resolve_port(
     return args.start_port
 
 
-def _resolve_device_id(
-    args: argparse.Namespace,
-    rank: int,
-    launches_all_local_ranks: bool,
-) -> int:
-    rank_offset = rank if launches_all_local_ranks else 0
-    return args.start_device_id + rank_offset
-
-
 def _build_command(
     binary_path: str,
     args: argparse.Namespace,
@@ -241,7 +222,6 @@ def _build_command(
     launches_all_local_ranks: bool,
 ) -> list[str]:
     port = _resolve_port(args, rank, launches_all_local_ranks)
-    device_id = _resolve_device_id(args, rank, launches_all_local_ranks)
 
     command = [binary_path]
     if args.config_json_file is not None:
@@ -249,7 +229,6 @@ def _build_command(
     command.append(f"--port={port}")
     command.append(f"--nnodes={args.nnodes}")
     command.append(f"--node_rank={rank}")
-    command.append(f"--device_id={device_id}")
     command.extend(extra_args)
     return command
 
diff --git a/xllm/pybind/args.py b/xllm/pybind/args.py
index 7e4767c6a7..6dd34a4936 100644
--- a/xllm/pybind/args.py
+++ b/xllm/pybind/args.py
@@ -34,9 +34,7 @@ def __init__(self) -> None:
         self.parser.add_argument('--model', type=str, help='"Name or path of the huggingface model to use."')
         self.parser.add_argument('--task', type=str, default="generate", help='The task to use the model for. generate/embed.')
         self.parser.add_argument('--runner', type=str, choices=['pooling'], default=None, help='Optional runner mode for LLM. Currently supports: pooling.')
-        self.parser.add_argument('--devices', type=str, default='npu:0', help='Devices to run the model on, e.g. npu:0,npu:1.')
         self.parser.add_argument('--draft_model', type=str, default='', help='draft hf model path to the model file.')
-        self.parser.add_argument('--draft_devices', type=str, default='npu:0', help='Devices to run the draft model on, e.g. npu:0,npu:1.')
         self.parser.add_argument('--limit_image_per_prompt', type=int, default=8, help='Maximum number of images per prompt.')
         self.parser.add_argument('--block_size', type=int, default=128, help='Number of slots per kv cache block. Default is 128.')
         self.parser.add_argument('--max_cache_size', type=int, default=0, help='Max gpu memory size for kv cache. Default is 0, which means cache size is calculated by available memory.')
diff --git a/xllm/pybind/bind.cpp b/xllm/pybind/bind.cpp
index 412535cd24..aab3abf560 100644
--- a/xllm/pybind/bind.cpp
+++ b/xllm/pybind/bind.cpp
@@ -40,9 +40,7 @@ PYBIND11_MODULE(xllm_export, m) {
   py::class_<Options>(m, "Options")
       .def(py::init())
       .def_readwrite("model_path", &Options::model_path_)
-      .def_readwrite("devices", &Options::devices_)
       .def_readwrite("draft_model_path", &Options::draft_model_path_)
-      .def_readwrite("draft_devices", &Options::draft_devices_)
       .def_readwrite("backend", &Options::backend_)
       .def_readwrite("limit_image_per_prompt",
                      &Options::limit_image_per_prompt_)
diff --git a/xllm/pybind/embedding.py b/xllm/pybind/embedding.py
index 024f7f88b3..3593b7d41d 100644
--- a/xllm/pybind/embedding.py
+++ b/xllm/pybind/embedding.py
@@ -28,7 +28,6 @@ class Embedding:
     def __init__(
         self,
         model: str,
-        devices: str = 'npu:0',
         limit_image_per_prompt: int = 8,
         block_size: int = 128,
         max_cache_size: int = 0,
@@ -76,9 +75,7 @@ def __init__(
         options = Options()
         options.model_path = model
         options.task_type = "embed"
-        options.devices = devices
         options.draft_model_path = None
-        options.draft_devices = None
         options.backend = "llm"
         options.limit_image_per_prompt = limit_image_per_prompt
         options.block_size = block_size
diff --git a/xllm/pybind/llm.py b/xllm/pybind/llm.py
index c15fdf51f7..969fb7f3b1 100644
--- a/xllm/pybind/llm.py
+++ b/xllm/pybind/llm.py
@@ -93,9 +93,7 @@ def __init__(
         model: str,
         task: str = "generate",
         runner: Optional[str] = None,
-        devices: str = 'npu:0',
         draft_model: Optional[str] = '',
-        draft_devices: Optional[str] = 'npu:0',
         limit_image_per_prompt: int = 8,
         block_size: int = 128,
         max_cache_size: int = 0,
@@ -165,9 +163,7 @@ def __init__(
         options = Options()
         options.model_path = model
         options.task_type = task
-        options.devices = devices
         options.draft_model_path = draft_model
-        options.draft_devices = draft_devices
         options.backend = backend
         options.limit_image_per_prompt = limit_image_per_prompt
         options.block_size = block_size
diff --git a/xllm/pybind/vlm.py b/xllm/pybind/vlm.py
index 2f52ea5560..84c85da03b 100644
--- a/xllm/pybind/vlm.py
+++ b/xllm/pybind/vlm.py
@@ -50,9 +50,7 @@ def __init__(
         self,
         model: str,
         task: str = "generate",
-        devices: str = 'npu:0',
         draft_model: Optional[str] = '',
-        draft_devices: Optional[str] = 'npu:0',
         limit_image_per_prompt: int = 8,
         block_size: int = 128,
         max_cache_size: int = 0,
@@ -107,9 +105,7 @@ def __init__(
         options = Options()
         options.model_path = model
         options.task_type = task
-        options.devices = devices
         options.draft_model_path = draft_model
-        options.draft_devices = draft_devices
         options.backend ="vlm"
         options.limit_image_per_prompt = limit_image_per_prompt
         options.block_size = block_size
diff --git a/xllm/xllm.cpp b/xllm/xllm.cpp
index e39787d061..d0ebb126fe 100644
--- a/xllm/xllm.cpp
+++ b/xllm/xllm.cpp
@@ -119,7 +119,6 @@ Options create_options(const std::string& instance_name, bool is_local) {
   options.model_path(model_config.model())
       .model_id(model_config.model_id())
       .task_type(model_config.task())
-      .devices(model_config.devices())
       .draft_model_path(speculative_config.draft_model())
       .backend(model_config.backend())
       .limit_image_per_prompt(model_config.limit_image_per_prompt())
@@ -224,13 +223,6 @@ Options create_options(const std::string& instance_name, bool is_local) {
           static_cast<int32_t>(rec_config.rec_worker_max_concurrency()))
       .is_local(is_local);
 
-  if (speculative_config.num_speculative_tokens() > 0) {
-    const std::string draft_devices = speculative_config.draft_devices().empty()
-                                          ? model_config.devices()
-                                          : speculative_config.draft_devices();
-    options.draft_devices(draft_devices);
-  }
-
   return options;
 }
 
@@ -427,8 +419,7 @@ int run() {
   // init XTensor allocator and PhyPagePool for xtensor mode
   if (kv_cache_config.enable_xtensor()) {
     // Parse devices
-    const auto devices =
-        DeviceNameUtils::parse_devices(options.devices().value_or("auto"));
+    const auto devices = DeviceNameUtils::parse_devices("auto");
 
     // Initialize XTensorAllocator with first device
     auto& allocator = XTensorAllocator::get_instance();