Skip to main content
This section describes the best practice data of mainstream LLM models such as DeepSeek and Qwen on the Ascend NPU. If you encounter issues or have any questions, please open an issue.

​
DeepSeek Series Models

​
Low Latency

ModelHardwareCardsDeploy ModeDatasetTPOTQuantizationConfiguration
Deepseek-R1Atlas 800I A332PD Disaggregation6K+1.6K20msW8A8 INT8Optimal Configuration
Deepseek-R1Atlas 800I A332PD Disaggregation3.9K+1K19msW8A8 INT8Optimal Configuration
Deepseek-R1Atlas 800I A332PD Disaggregation3.5K+1.5K19msW8A8 INT8Optimal Configuration
Deepseek-R1Atlas 800I A332PD Disaggregation3.5K+1K19msW8A8 INT8Optimal Configuration
DeepSeek-V3.2Atlas 800I A332PD Disaggregation128K+1K26msW8A8 INT8Optimal Configuration

​
High Throughput

ModelHardwareCardsDeploy ModeDatasetTPOTQuantizationConfiguration
Deepseek-R1Atlas 800I A332PD Disaggregation3.5K+1.5K50msW8A8 INT8Optimal Configuration
Deepseek-R1Atlas 800I A324PD Disaggregation2K+2K50msW8A8 INT8Optimal Configuration
Deepseek-R1Atlas 800I A38PD Mixed2K+2K50msW4A8 INT8Optimal Configuration
Deepseek-R1Atlas 800I A316PD Disaggregation2K+2K50msW4A8 INT8Optimal Configuration
Deepseek-R1Atlas 800I A38PD Mixed3.5K+1.5K50msW4A8 INT8Optimal Configuration
Deepseek-R1Atlas 800I A316PD Disaggregation3.5K+1.5K50msW4A8 INT8Optimal Configuration

​
Qwen Series Models

​
Low Latency

ModelHardwareCardsDeploy ModeDatasetTPOTQuantizationConfiguration
Qwen3-235B-A22BAtlas 800I A38PD Mixed11K+1K10msBF16Optimal Configuration
Qwen3-32BAtlas 800I A34PD Mixed6K+1.5K18msBF16Optimal Configuration
Qwen3-32BAtlas 800I A34PD Mixed4K+1.5K11msBF16Optimal Configuration
Qwen3-32BAtlas 800I A38PD Mixed18K+4K6msBF16Optimal Configuration
Qwen3-32BAtlas 800I A28PD Mixed6K+1.5K18msW8A8 INT8Optimal Configuration
Qwen3-32BAtlas 800I A28PD Mixed4K+1.5K11msBF16Optimal Configuration
Qwen3-32BAtlas 800I A32PD Mixed1K+0.3K12msW8A8 INT8Optimal Configuration
Qwen3-32BAtlas 800I A32PD Mixed6K+1.5K17msW8A8 INT8Optimal Configuration
Qwen3-8BAtlas 800I A31PD Mixed1K+0.3K7msW8A8 INT8Optimal Configuration
Qwen3-8BAtlas 800I A31PD Mixed6K+1.5K12msW8A8 INT8Optimal Configuration
Qwen3-8BAtlas 800I A31PD Mixed3.5K+1.5K5msW8A8 INT8Optimal Configuration
Qwen3-30B-A3BAtlas 800I A31PD Mixed6K+1.5K10msW8A8 INT8Optimal Configuration
Qwen3-30B-A3BAtlas 800I A31PD Mixed1K+0.3K7msW8A8 INT8Optimal Configuration
Qwen3-Next-A3B-InstructAtlas 800I A32PD Mixed1K+0.3K14.21msW8A8 INT8Optimal Configuration
Qwen3-Next-A3B-InstructAtlas 800I A32PD Mixed6K+1.5K15.62msW8A8 INT8Optimal Configuration
Qwen3-Next-A3B-InstructAtlas 800I A32PD Mixed3.5K+1.5K20msW8A8 INT8Optimal Configuration
Qwen3-14BAtlas 800I A31PD Mixed3.5K+1.5K9msW8A8 INT8Optimal Configuration

​
High Throughput

ModelHardwareCardsDeploy ModeDatasetTPOTQuantizationConfiguration
Qwen3-235B-A22BAtlas 800I A324PD Disaggregation3.5K+1.5K50msW8A8 INT8Optimal Configuration
Qwen3-235B-A22BAtlas 800I A38PD Mixed3.5K+1.5K50msW8A8 INT8Optimal Configuration
Qwen3-235B-A22BAtlas 800I A38PD Mixed2K+2K100msW8A8 INT8Optimal Configuration
Qwen3-235B-A22BAtlas 800I A38PD Mixed2K+2K50msW8A8 INT8Optimal Configuration
Qwen3-235B-A22BAtlas 800I A316PD Mixed2K+2K50msW8A8 INT8Optimal Configuration
Qwen3-32BAtlas 800I A32PD Mixed3.5K+1.5K50msW8A8 INT8Optimal Configuration
Qwen3-32BAtlas 800I A32PD Mixed2K+2K50msW8A8 INT8Optimal Configuration
Qwen3-30B-A3BAtlas 800I A31PD Mixed3.5K+1.5K50msW8A8 INT8Optimal Configuration
Qwen3-Coder-480B-A35B-InstructAtlas 800I A324PD Disaggregation3.5K+1.5K50msW8A8 INT8Optimal Configuration
Qwen3-Coder-480B-A35B-InstructAtlas 800I A316PD Mixed3.5K+1.5K50msW8A8 INT8Optimal Configuration
Qwen3-Coder-480B-A35B-InstructAtlas 800I A38PD Mixed3.5K+1.5K50msW8A8 INT8Optimal Configuration
Qwen3-Next-80B-A3B-InstructAtlas 800I A32PD Mixed3.5K+1.5K50msW8A8 INT8Optimal Configuration
Qwen3-32BAtlas 800I A28PD Mixed3.5K+1.5K50msW8A8 INT8Optimal Configuration
Qwen3-32BAtlas 800I A28PD Mixed2K+2K50msW8A8 INT8Optimal Configuration
Qwen3-14BAtlas 800I A31PD Mixed3.5K+1.5K50msW8A8 INT8Optimal Configuration
Qwen3-8BAtlas 800I A31PD Mixed3.5K+1.5K50msW8A8 INT8Optimal Configuration

​
Optimal Configuration

​
DeepSeek-R1 3_5K-1_5K 50ms on A3 32 Cards Disaggregation Mode

Model: Deepseek R1 Hardware: Atlas 800I A3 32Card DeployMode: PD Disaggregation Dataset: random Input Output Length: 3.5K+1.5K TPOT: 50ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000
export SGLANG_SET_CPU_AFFINITY=1
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export STREAMS_PER_DEVICE=32
export HCCL_OP_EXPANSION_MODE=AIV
export SGLANG_NPU_USE_MLAPO=1
export SGLANG_USE_FIA_NZ=1
export SGLANG_NPU_USE_MULTI_STREAM=1

export ASCEND_MF_STORE_URL="tcp://your prefill ip1:24669"

P_IP=('your prefill ip1' 'your prefill ip2')

D_IP=('your decode ip1' 'your decode ip2')

MODEL_PATH=xxx

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`
echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"
# prefill
for i in "${!P_IP[@]}";
do
    if [[ "$LOCAL_HOST1" == "${P_IP[$i]}" || "$LOCAL_HOST2" == "${P_IP[$i]}" ]];
    then
        echo "${P_IP[$i]}"
        export SGLANG_USE_AG_AFTER_QLORA=1
        export HCCL_BUFFSIZE=800
        export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
        export TASK_QUEUE_ENABLE=2
        export SGLANG_NPU_FUSED_MOE_MODE=2
        export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=131072

        export HCCL_SOCKET_IFNAME=lo
        export GLOO_SOCKET_IFNAME=lo
        python -m sglang.launch_server --model-path ${MODEL_PATH}  --disaggregation-mode prefill --host ${P_IP[$i]} \
        --port 8000 --disaggregation-bootstrap-port $((8998+$i)) --trust-remote-code --nnodes 1 --node-rank 0 \
        --tp-size 16 --mem-fraction-static 0.778 --attention-backend ascend --device npu --quantization modelslim \
        --disaggregation-transfer-backend ascend --max-running-requests 16 --disable-radix-cache \
        --chunked-prefill-size -1 --max-prefill-tokens 60000 --moe-a2a-backend ascend_fuseep --deepep-mode normal \
        --speculative-algorithm NEXTN --speculative-num-steps 1 --speculative-eagle-topk 1 --speculative-num-draft-tokens 2  \
        --dp-size 4 --enable-dp-attention --disable-shared-experts-fusion --dtype bfloat16 --enable-attn-tp-input-scattered
        NODE_RANK=$i
        break
    fi
done

# decode
for i in "${!D_IP[@]}";
do
    if [[ "$LOCAL_HOST1" == "${D_IP[$i]}" || "$LOCAL_HOST2" == "${D_IP[$i]}" ]];
    then
        echo "${D_IP[$i]}"
        export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
        export SGLANG_ENABLE_SPEC_V2=1
        export HCCL_BUFFSIZE=600
        export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=64
        export TASK_QUEUE_ENABLE=1
        export SGLANG_NPU_FUSED_MOE_MODE=1
        export SGLANG_LM_HEAD_TP=8
        export HCCL_SOCKET_IFNAME=xxx
        export GLOO_SOCKET_IFNAME=xxx
        python -m sglang.launch_server --model-path ${MODEL_PATH} --disaggregation-mode decode --host ${D_IP[$i]} \
        --port 8001 --trust-remote-code --dist-init-addr ${D_IP[0]}:5000 --nnodes 2 --node-rank $i --tp-size 32 --dp-size 32 \
        --mem-fraction-static 0.82 --max-running-requests 1024 --attention-backend ascend --device npu --quantization modelslim \
        --moe-a2a-backend ascend_fuseep --enable-dp-attention --deepep-mode low_latency --moe-dense-tp 1 \
        --cuda-graph-bs 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 --disaggregation-transfer-backend ascend --watchdog-timeout 9000 --context-length 8192 \
        --speculative-algorithm NEXTN --speculative-num-steps 1 --speculative-eagle-topk 1 --speculative-num-draft-tokens 2  \
        --tokenizer-worker-num 4 --disable-shared-experts-fusion --dtype bfloat16 \
        --load-balance-method round_robin
        NODE_RANK=$i
        break
    fi
done
Command
export SGLANG_DP_ROUND_ROBIN=1
python -m sglang_router.launch_router \
    --pd-disaggregation \
    --policy cache_aware \
    --prefill http://P_IP:8000 8998 \
    --prefill http://P_IP:8000 8999 \
    --decode http://D_IP:8001 \
    --host 127.0.0.1 \
    --port 6688 \
    --mini-lb

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 6688 --max-concurrency 768  --random-input-len 3500 --random-output-len 1500 --num-prompts 3072 --random-range-ratio 1 --request-rate 16

​
DeepSeek-R1 2K-2K 50ms on A3 24 Cards Disaggregation Mode

Model: Deepseek R1 Hardware: Atlas 800I A3 24Card DeployMode: PD Disaggregation Dataset: random Input Output Length: 2K+2K TPOT: 50ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

export SGLANG_SET_CPU_AFFINITY=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export STREAMS_PER_DEVICE=32
export SGLANG_NPU_USE_MLAPO=1
export SGLANG_USE_FIA_NZ=1

export ASCEND_MF_STORE_URL="tcp://your prefill ip1:24669"

P_IP=('your prefill ip1')
D_IP=('your decode ip1' 'your decode ip2')

MODEL_PATH=xxx

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`
echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"
# prefill
for i in "${!P_IP[@]}";
do
    if [[ "$LOCAL_HOST1" == "${P_IP[$i]}" || "$LOCAL_HOST2" == "${P_IP[$i]}" ]];
    then
        echo "${P_IP[$i]}"
        export HCCL_BUFFSIZE=1600
        export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
        export TASK_QUEUE_ENABLE=2
        export SGLANG_USE_AG_AFTER_QLORA=1
        export HCCL_SOCKET_IFNAME=lo
        export GLOO_SOCKET_IFNAME=lo

        python -m sglang.launch_server --model-path ${MODEL_PATH}  --disaggregation-mode prefill --host ${P_IP[$i]} \
        --port 8000 --disaggregation-bootstrap-port $((8998+$i)) --trust-remote-code --nnodes 1 --node-rank 0 \
        --tp-size 16 --mem-fraction-static 0.8 --attention-backend ascend --device npu --quantization modelslim \
        --disaggregation-transfer-backend ascend --max-running-requests 20 --context-length 8192 --disable-radix-cache \
        --chunked-prefill-size -1 --max-prefill-tokens 28680 --moe-a2a-backend deepep --deepep-mode normal \
        --speculative-algorithm NEXTN --speculative-num-steps 1 --speculative-eagle-topk 1 --speculative-num-draft-tokens 2  \
        --dp-size 4 --enable-dp-attention --disable-shared-experts-fusion --dtype bfloat16 --enable-attn-tp-input-scattered
        NODE_RANK=$i
        break
    fi
done

# decode
for i in "${!D_IP[@]}";
do
    if [[ "$LOCAL_HOST1" == "${D_IP[$i]}" || "$LOCAL_HOST2" == "${D_IP[$i]}" ]];
    then
        echo "${D_IP[$i]}"
        export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
        export SGLANG_ENABLE_SPEC_V2=1
        export HCCL_BUFFSIZE=800
        export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=102
        export TASK_QUEUE_ENABLE=1
        export SGLANG_SCHEDULER_SKIP_ALL_GATHER=1
        export SGLANG_NPU_FUSED_MOE_MODE=1
        export HCCL_SOCKET_IFNAME=xxx
        export GLOO_SOCKET_IFNAME=xxx

        python -m sglang.launch_server --model-path ${MODEL_PATH} --disaggregation-mode decode --host ${D_IP[$i]} \
        --port 8001 --trust-remote-code --dist-init-addr ${D_IP[0]}:5000 --nnodes 2 --node-rank $i --tp-size 32 --dp-size 32 \
        --mem-fraction-static 0.81 --max-running-requests 1088 --attention-backend ascend --device npu --quantization modelslim \
        --moe-a2a-backend ascend_fuseep --enable-dp-attention --deepep-mode low_latency --enable-dp-lm-head --moe-dense-tp 1 \
        --cuda-graph-bs 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 --disaggregation-transfer-backend ascend --watchdog-timeout 9000 --context-length 8192 \
        --speculative-algorithm NEXTN --speculative-num-steps 2 --speculative-eagle-topk 1 --speculative-num-draft-tokens 3  \
        --tokenizer-worker-num 4 --disable-shared-experts-fusion --dtype bfloat16 \
        --load-balance-method round_robin
        NODE_RANK=$i
        break
    fi
done
Command
python -m sglang_router.launch_router \
    --pd-disaggregation \
    --policy cache_aware \
    --prefill http://P_IP:8000 8998 \
    --prefill http://P_IP:8000 8999 \
    --decode http://D_IP:8001 \
    --host 127.0.0.1 \
    --port 6688 \
    --mini-lb

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python -m sglang.bench_serving --dataset-name random --backend sglang \
--host 127.0.0.1 \
--port 6688 \
--max-concurrency 1088 \
--random-input-len 2048 \
--random-output-len 2048 \
--num-prompts 12800 \
--random-range-ratio 1 \
--request-rate 24

​
DeepSeek-R1 6K-1_6K 20ms on A3 32 Cards Disaggregation Mode

Model: Deepseek R1 Hardware: Atlas 800I A3 32Card DeployMode: PD Disaggregation Dataset: random Input Output Length: 6K+1.6K TPOT: 20ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000
export SGLANG_SET_CPU_AFFINITY=1
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export STREAMS_PER_DEVICE=32
export ASCEND_MF_STORE_URL="tcp://your prefill ip1:24669"

P_IP=('your prefill ip1' 'your prefill ip2')

D_IP=('your decode ip1' 'your decode ip2')

MODEL_PATH=xxx

export SGLANG_NPU_USE_MLAPO=1
export SGLANG_USE_FIA_NZ=1

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`
echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"
# prefill
for i in "${!P_IP[@]}";
do
    if [[ "$LOCAL_HOST1" == "${P_IP[$i]}" || "$LOCAL_HOST2" == "${P_IP[$i]}" ]];
    then
        echo "${P_IP[$i]}"
        export HCCL_BUFFSIZE=1536
        export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
        export TASK_QUEUE_ENABLE=2

        export HCCL_SOCKET_IFNAME=lo
        export GLOO_SOCKET_IFNAME=lo
        python -m sglang.launch_server --model-path ${MODEL_PATH}  --disaggregation-mode prefill --host ${P_IP[$i]} \
        --port 8000 --disaggregation-bootstrap-port $((8998+$i)) --trust-remote-code --nnodes 1 --node-rank 0 \
        --tp-size 16 --mem-fraction-static 0.81 --attention-backend ascend --device npu --quantization modelslim \
        --disaggregation-transfer-backend ascend --max-running-requests 4 --disable-radix-cache \
        --chunked-prefill-size -1 --max-prefill-tokens 28680 --moe-a2a-backend deepep --deepep-mode normal \
        --speculative-algorithm NEXTN --speculative-num-steps 1 --speculative-eagle-topk 1 --speculative-num-draft-tokens 2  \
        --dp-size 2 --enable-dp-attention --disable-shared-experts-fusion --dtype bfloat16 --enable-attn-tp-input-scattered
        NODE_RANK=$i
        break
    fi
done

# decode
for i in "${!D_IP[@]}";
do
    if [[ "$LOCAL_HOST1" == "${D_IP[$i]}" || "$LOCAL_HOST2" == "${D_IP[$i]}" ]];
    then
        echo "${D_IP[$i]}"
        export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
        export SGLANG_ENABLE_SPEC_V2=1
        export HCCL_BUFFSIZE=650
        export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=16
        export TASK_QUEUE_ENABLE=1
        export SGLANG_SCHEDULER_SKIP_ALL_GATHER=1
        export HCCL_SOCKET_IFNAME=xxx
        export GLOO_SOCKET_IFNAME=xxx
        python -m sglang.launch_server --model-path ${MODEL_PATH} --disaggregation-mode decode --host ${D_IP[$i]} \
        --port 8001 --trust-remote-code --dist-init-addr DIP1:5000 --nnodes 2 --node-rank $i --tp-size 32 --dp-size 8 \
        --mem-fraction-static 0.75 --max-running-requests 32 --attention-backend ascend --device npu --quantization modelslim \
        --moe-a2a-backend deepep --enable-dp-attention --deepep-mode low_latency --enable-dp-lm-head --moe-dense-tp 1 \
        --cuda-graph-bs 2 4 6 --disaggregation-transfer-backend ascend --watchdog-timeout 9000 \
        --speculative-algorithm NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4  \
        --tokenizer-worker-num 4 --disable-shared-experts-fusion --dtype bfloat16 \
        --load-balance-method round_robin
        NODE_RANK=$i
        break
    fi
done
Command
export SGLANG_DP_ROUND_ROBIN=1
python -m sglang_router.launch_router \
    --pd-disaggregation \
    --policy cache_aware \
    --prefill http://P_IP:8000 8998 \
    --prefill http://P_IP:8000 8999 \
    --decode http://D_IP:8001 \
    --host 127.0.0.1 \
    --port 6688 \
    --mini-lb

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python -m sglang.bench_serving --dataset-name random --backend sglang \
    --host 127.0.0.1 \
    --port 6688 \
    --max-concurrency 32 \
    --random-input-len 6000 \
    --random-output-len 1600 \
    --num-prompts 32 \
    --random-range-ratio 1 \
    --request-rate 16

​
DeepSeek-R1 3_9K-1K 19ms on A3 32 Cards Disaggregation Mode

Model: Deepseek R1 Hardware: Atlas 800I A3 32Card DeployMode: PD Disaggregation Dataset: random Input Output Length: 3.9K+1K TPOT: 19ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

export SGLANG_SET_CPU_AFFINITY=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export STREAMS_PER_DEVICE=32
export SGLANG_NPU_USE_MLAPO=1
export SGLANG_USE_FIA_NZ=1
export ASCEND_MF_STORE_URL="tcp://your prefill ip1:24669"

P_IP=('your prefill ip1' 'your prefill ip2')
D_IP=('your decode ip1' 'your decode ip2')

MODEL_PATH=xxx

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`
echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

# prefill
for i in "${!P_IP[@]}";
do
    if [[ "$LOCAL_HOST1" == "${P_IP[$i]}" || "$LOCAL_HOST2" == "${P_IP[$i]}" ]];
    then
        echo "${P_IP[$i]}"
        export HCCL_BUFFSIZE=1536
        export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
        export TASK_QUEUE_ENABLE=2
        export HCCL_SOCKET_IFNAME=lo
        export GLOO_SOCKET_IFNAME=lo

        python -m sglang.launch_server --model-path ${MODEL_PATH}  --disaggregation-mode prefill --host ${P_IP[$i]} \
        --port 8000 --disaggregation-bootstrap-port $((8998+$i)) --trust-remote-code --nnodes 1 --node-rank 0 \
        --tp-size 16 --mem-fraction-static 0.81 --attention-backend ascend --device npu --quantization modelslim \
        --disaggregation-transfer-backend ascend --max-running-requests 4 --context-length 8192 --disable-radix-cache \
        --chunked-prefill-size -1 --max-prefill-tokens 28680 --moe-a2a-backend deepep --deepep-mode normal \
        --speculative-algorithm NEXTN --speculative-num-steps 1 --speculative-eagle-topk 1 --speculative-num-draft-tokens 2  \
        --dp-size 2 --enable-dp-attention --disable-shared-experts-fusion --dtype bfloat16 --enable-attn-tp-input-scattered
        NODE_RANK=$i
        break
    fi
done

# decode
for i in "${!D_IP[@]}";
do
    if [[ "$LOCAL_HOST1" == "${D_IP[$i]}" || "$LOCAL_HOST2" == "${D_IP[$i]}" ]];
    then
        echo "${D_IP[$i]}"
        export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
        export SGLANG_ENABLE_SPEC_V2=1
        export HCCL_BUFFSIZE=650
        export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=12
        export TASK_QUEUE_ENABLE=1
        export SGLANG_SCHEDULER_SKIP_ALL_GATHER=1
        export HCCL_SOCKET_IFNAME=xxx
        export GLOO_SOCKET_IFNAME=xxx
        python -m sglang.launch_server --model-path ${MODEL_PATH} --disaggregation-mode decode --host ${D_IP[$i]} \
        --port 8001 --trust-remote-code --dist-init-addr DIP1:5000 --nnodes 2 --node-rank $i --tp-size 32 --dp-size 16 \
        --mem-fraction-static 0.75 --max-running-requests 32 --attention-backend ascend --device npu --quantization modelslim \
        --moe-a2a-backend deepep --enable-dp-attention --deepep-mode low_latency --enable-dp-lm-head --moe-dense-tp 1 \
        --cuda-graph-bs 2 4 6 --disaggregation-transfer-backend ascend --watchdog-timeout 9000 --context-length 8192 \
        --speculative-algorithm NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4  \
        --tokenizer-worker-num 4 --disable-shared-experts-fusion --dtype bfloat16 \
        --load-balance-method round_robin
        NODE_RANK=$i
        break
    fi
done
Command
export SGLANG_DP_ROUND_ROBIN=1
python -m sglang_router.launch_router \
    --pd-disaggregation \
    --policy cache_aware \
    --prefill http://P_IP:8000 8998 \
    --prefill http://P_IP:8000 8999 \
    --decode http://D_IP:8001 \
    --host 127.0.0.1 \
    --port 6688 \
    --mini-lb

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python -m sglang.bench_serving --dataset-name random --backend sglang \
    --host 127.0.0.1 \
    --port 6688 \
    --max-concurrency 32 \
    --random-input-len 3900 \
    --random-output-len 1024 \
    --num-prompts 32 \
    --random-range-ratio 1 \
    --request-rate 16

​
DeepSeek-R1 3_5K-1_5K 19ms on A3 32 Cards Disaggregation Mode

Model: Deepseek R1 Hardware: Atlas 800I A3 32Card DeployMode: PD Disaggregation Dataset: random Input Output Length: 3.5K+1.5K TPOT: 19ms

​
Model Deployment

Please Turn to DeepSeek-R1 3_9K-1K 19ms on A3 32 Cards Disaggregation Mode

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python -m sglang.bench_serving --dataset-name random --backend sglang \
    --host 127.0.0.1 \
    --port 6688 \
    --max-concurrency 32 \
    --random-input-len 3500 \
    --random-output-len 1500 \
    --num-prompts 32 \
    --random-range-ratio 1 \
    --request-rate 16

​
DeepSeek-R1 3_5K-1K 19ms on A3 32 Cards Disaggregation Mode

Model: Deepseek R1 Hardware: Atlas 800I A3 32Card DeployMode: PD Disaggregation Dataset: random Input Output Length: 3.5K+1K TPOT: 19ms

​
Model Deployment

Please Turn to DeepSeek-R1 3_9K-1K 19ms on A3 32 Cards Disaggregation Mode

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python -m sglang.bench_serving --dataset-name random --backend sglang \
    --host 127.0.0.1 \
    --port 6688 \
    --max-concurrency 32 \
    --random-input-len 3500 \
    --random-output-len 1024 \
    --num-prompts 32 \
    --random-range-ratio 1 \
    --request-rate 16

​
DeepSeek-R1 2K-2K 50ms on A3 8 Cards Mixed Mode

Model: Deepseek R1 Hardware: Atlas 800I A3 8Card DeployMode: PD Mixed Dataset: random Input Output Length: 2K+2K TPOT: 50ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

export SGLANG_SET_CPU_AFFINITY=1
unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export STREAMS_PER_DEVICE=32
export SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE=1
export SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES=200

export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo

export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=88
export HCCL_BUFFSIZE=1600
export DEEPEP_NORMAL_LONG_SEQ_ROUND=10
export DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS=512

MODEL_PATH=xxx

export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
export SGLANG_NPU_USE_MLAPO=1
export SGLANG_ENABLE_SPEC_V2=1
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_USE_FIA_NZ=1

python3 -m sglang.launch_server --model-path ${MODEL_PATH} \
--tp 16 \
--trust-remote-code \
--attention-backend ascend \
--device npu \
--quantization modelslim \
--watchdog-timeout 9000 \
--host 127.0.0.1 --port 6699 \
--cuda-graph-bs 4 8 20 21 22 \
--mem-fraction-static 0.78 \
--max-running-requests 352 \
--disable-radix-cache --chunked-prefill-size -1 --max-prefill-tokens 1500 \
--moe-a2a-backend deepep --deepep-mode auto \
--enable-dp-attention --dp-size 16 --enable-dp-lm-head \
--speculative-algorithm NEXTN --speculative-num-steps 2 --speculative-eagle-topk 1 --speculative-num-draft-tokens 3 \
--dtype bfloat16

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 6699 --max-concurrency 352  --random-input-len 2048 --random-output-len 2048 --num-prompts 1408 --random-range-ratio 1

​
DeepSeek-R1 2K-2K 50ms on A3 16 Cards Disaggregation Mode

Model: Deepseek R1 Hardware: Atlas 800I A3 16Card DeployMode: PD Disaggregation Dataset: random Input Output Length: 2K+2K TPOT: 50ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

export SGLANG_SET_CPU_AFFINITY=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export STREAMS_PER_DEVICE=32

export ASCEND_MF_STORE_URL="tcp://your prefill ip1:24667"

P_IP=('your prefill ip1')

D_IP=('your decode ip1')

MODEL_PATH=xxx

export SGLANG_NPU_USE_MLAPO=1
export SGLANG_USE_FIA_NZ=1
export ENABLE_MOE_NZ=1

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`
echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

# prefill
for i in "${!P_IP[@]}";
do
    if [[ "$LOCAL_HOST1" == "${P_IP[$i]}" || "$LOCAL_HOST2" == "${P_IP[$i]}" ]];
    then
        echo "${P_IP[$i]}"
        export HCCL_BUFFSIZE=2600
        export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
        export TASK_QUEUE_ENABLE=2

        export HCCL_SOCKET_IFNAME=lo
        export GLOO_SOCKET_IFNAME=lo
        python -m sglang.launch_server --model-path ${MODEL_PATH}  --disaggregation-mode prefill --host ${P_IP[$i]} \
        --port 8000 --disaggregation-bootstrap-port $((8998+$i)) --trust-remote-code --nnodes 1 --node-rank 0 \
        --tp-size 16 --mem-fraction-static 0.7 --attention-backend ascend --device npu --quantization modelslim \
        --disaggregation-transfer-backend ascend --max-running-requests 32 --context-length 8192  --disable-radix-cache \
        --chunked-prefill-size -1 --max-prefill-tokens 10240 --moe-a2a-backend deepep --deepep-mode normal \
        --speculative-algorithm NEXTN --speculative-num-steps 1 --speculative-eagle-topk 1 --speculative-num-draft-tokens 2  \
        --dp-size 8 --enable-dp-attention --disable-shared-experts-fusion --dtype bfloat16
        NODE_RANK=$i
        break
    fi
done

# decode
for i in "${!D_IP[@]}";
do
    if [[ "$LOCAL_HOST1" == "${D_IP[$i]}" || "$LOCAL_HOST2" == "${D_IP[$i]}" ]];
    then
        echo "${D_IP[$i]}"
        export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
        export SGLANG_ENABLE_SPEC_V2=1
        export HCCL_BUFFSIZE=900
        export SGLANG_DP_ROUND_ROBIN=1
        export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=112
        export TASK_QUEUE_ENABLE=1
        export HCCL_SOCKET_IFNAME=xxx
        export GLOO_SOCKET_IFNAME=xxx
        python -m sglang.launch_server --model-path ${MODEL_PATH} --disaggregation-mode decode --host ${D_IP[$i]} \
        --port 8001 --trust-remote-code --nnodes 1 --node-rank 0 --tp-size 16 --dp-size 16 \
        --mem-fraction-static 0.8 --max-running-requests 448 --attention-backend ascend --device npu --quantization modelslim \
        --moe-a2a-backend deepep --enable-dp-attention --deepep-mode low_latency --enable-dp-lm-head \
        --cuda-graph-bs 2 4 6 8 10 12 14 16 18 20 22 24 26 28 --disaggregation-transfer-backend ascend --watchdog-timeout 9000 --context-length 8192 \
        --speculative-algorithm NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4  \
        --disable-shared-experts-fusion --dtype bfloat16 --tokenizer-worker-num 4 \
        --load-balance-method round_robin
        NODE_RANK=$i
        break
    fi
done
Command
python -m sglang_router.launch_router \
    --pd-disaggregation \
    --policy cache_aware \
    --prefill http://P_IP:8000 8998 \
    --decode http://D_IP:8001 \
    --host 127.0.0.1 \
    --port 6688 \
    --mini-lb

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 6688 --max-concurrency 448  --random-input-len 2048 --random-output-len 2048 --num-prompts 1792 --random-range-ratio 1 --request-rate 32

​
DeepSeek-R1 3_5K-1_5K 50ms on A3 8 Cards Mixed Mode

Model: Deepseek R1 Hardware: Atlas 800I A3 8Card DeployMode: PD Mixed Dataset: random Input Output Length: 3.5K+1.5K TPOT: 50ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True

export STREAMS_PER_DEVICE=32
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE=1
export SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES=200
export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=56
export HCCL_BUFFSIZE=1200
export DEEPEP_NORMAL_LONG_SEQ_ROUND=10
export DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS=512
export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
export SGLANG_NPU_USE_MLAPO=1
export SGLANG_ENABLE_SPEC_V2=1
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_USE_FIA_NZ=1

MODEL_PATH=xxx

python3 -m sglang.launch_server --model-path ${MODEL_PATH} \
--tp 16 \
--trust-remote-code \
--attention-backend ascend \
--device npu \
--quantization modelslim \
--watchdog-timeout 9000 \
--host 127.0.0.1 --port 6699 \
--cuda-graph-bs 4 8 12 14 \
--mem-fraction-static 0.77 \
--max-running-requests 224 \
--context-length 8188  --disable-radix-cache --chunked-prefill-size -1 --max-prefill-tokens 3000 \
--moe-a2a-backend deepep --deepep-mode auto \
--enable-dp-attention --dp-size 16 --enable-dp-lm-head \
--speculative-algorithm NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 \
--dtype bfloat16

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 6699 --max-concurrency 224  --random-input-len 3500 --random-output-len 1500 --num-prompts 896 --random-range-ratio 1

​
DeepSeek-R1 3_5K-1_5K 50ms on A3 16 Cards Disaggregation Mode

Model: Deepseek R1 Hardware: Atlas 800I A3 16Card DeployMode: PD Disaggregation Dataset: random Input Output Length: 3.5K+1.5K TPOT: 50ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

export SGLANG_SET_CPU_AFFINITY=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export STREAMS_PER_DEVICE=32

export ASCEND_MF_STORE_URL="tcp://your prefill ip1:24667"

P_IP=('your prefill ip1')

D_IP=('your decode ip1')

MODEL_PATH=xxx

export SGLANG_NPU_USE_MLAPO=1
export SGLANG_USE_FIA_NZ=1
export ENABLE_MOE_NZ=1

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`
echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

# prefill
for i in "${!P_IP[@]}";
do
    if [[ "$LOCAL_HOST1" == "${P_IP[$i]}" || "$LOCAL_HOST2" == "${P_IP[$i]}" ]];
    then
        echo "${P_IP[$i]}"
        export HCCL_BUFFSIZE=3500
        export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
        export TASK_QUEUE_ENABLE=2

        export HCCL_SOCKET_IFNAME=lo
        export GLOO_SOCKET_IFNAME=lo
        python -m sglang.launch_server --model-path ${MODEL_PATH}  --disaggregation-mode prefill --host ${P_IP[$i]} \
        --port 8000 --disaggregation-bootstrap-port $((8998+$i)) --trust-remote-code --nnodes 1 --node-rank 0 \
        --tp-size 16 --mem-fraction-static 0.62 --attention-backend ascend --device npu --quantization modelslim \
        --disaggregation-transfer-backend ascend --max-running-requests 32 --context-length 8192  --disable-radix-cache \
        --chunked-prefill-size -1 --max-prefill-tokens 20480 --moe-a2a-backend deepep --deepep-mode normal \
        --speculative-algorithm NEXTN --speculative-num-steps 1 --speculative-eagle-topk 1 --speculative-num-draft-tokens 2  \
        --dp-size 8 --enable-dp-attention --disable-shared-experts-fusion --dtype bfloat16
        NODE_RANK=$i
        break
    fi
done

# decode
for i in "${!D_IP[@]}";
do
    if [[ "$LOCAL_HOST1" == "${D_IP[$i]}" || "$LOCAL_HOST2" == "${D_IP[$i]}" ]];
    then
        echo "${D_IP[$i]}"
        export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
        export SGLANG_ENABLE_SPEC_V2=1
        export HCCL_BUFFSIZE=800
        export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=78
        export TASK_QUEUE_ENABLE=1
        export HCCL_SOCKET_IFNAME=xxx
        export GLOO_SOCKET_IFNAME=xxx
        python -m sglang.launch_server --model-path ${MODEL_PATH} --disaggregation-mode decode --host ${D_IP[$i]} \
        --port 8001 --trust-remote-code --nnodes 1 --node-rank 0 --tp-size 16 --dp-size 16 \
        --mem-fraction-static 0.805 --max-running-requests 416 --attention-backend ascend --device npu --quantization modelslim \
        --moe-a2a-backend deepep --enable-dp-attention --deepep-mode low_latency --enable-dp-lm-head \
        --cuda-graph-bs 2 4 6 8 10 12 14 16 18 20 22 24 26 --disaggregation-transfer-backend ascend --watchdog-timeout 9000 --context-length 8192 \
        --speculative-algorithm NEXTN --speculative-num-steps 2 --speculative-eagle-topk 1 --speculative-num-draft-tokens 3  \
        --disable-shared-experts-fusion --dtype bfloat16 --tokenizer-worker-num 4 \
		--load-balance-method round_robin
        NODE_RANK=$i
        break
    fi
done
Command
python -m sglang_router.launch_router \
    --pd-disaggregation \
    --policy cache_aware \
    --prefill http://P_IP:8000 8998 \
    --decode http://D_IP:8001 \
    --host 127.0.0.1 \
    --port 6688 \
    --mini-lb

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 6688 --max-concurrency 416  --random-input-len 3500 --random-output-len 1500 --num-prompts 1664 --random-range-ratio 1

​
DeepSeek-V3.2 128K-1K 26ms on A3 32 Cards Disaggregation Mode

Model: DeepSeek-V3.2-W8A8 Hardware: Atlas 800I A3 32Card DeployMode: PD Disaggregation Dataset: random Input Output Length: 128K+1K TPOT: 26ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING

source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/op_api/lib/:${LD_LIBRARY_PATH}
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

export SGLANG_SET_CPU_AFFINITY=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export STREAMS_PER_DEVICE=32
export ASCEND_MF_STORE_URL="tcp://your prefill ip1:24670"

P_IP=('your prefill ip1' 'your prefill ip2')
D_IP=('your decode ip1' 'your decode ip2')
MODEL_PATH=xxx

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`
echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

# prefill
for i in "${!P_IP[@]}";
do
    if [[ "$LOCAL_HOST1" == "${P_IP[$i]}" || "$LOCAL_HOST2" == "${P_IP[$i]}" ]];
    then
        echo "${P_IP[$i]}"
        export HCCL_BUFFSIZE=1200
        export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
        export TASK_QUEUE_ENABLE=2
        export HCCL_SOCKET_IFNAME=xxx
        export GLOO_SOCKET_IFNAME=xxx

        python3 -m sglang.launch_server --model-path ${MODEL_PATH} \
        --tp 32 \
        --trust-remote-code \
        --attention-backend ascend \
        --device npu \
        --watchdog-timeout 9000 \
        --host ${P_IP[$i]} --port 8000 \
        --mem-fraction-static 0.73 \
        --disable-radix-cache --chunked-prefill-size -1 --max-prefill-tokens 68000 \
        --max-running-requests 1 \
        --moe-a2a-backend deepep --deepep-mode normal \
        --quantization modelslim \
        --disaggregation-transfer-backend ascend \
        --disaggregation-mode prefill \
        --disable-cuda-graph \
        --nnodes 2 --node-rank $i \
        --disaggregation-bootstrap-port 8995 \
        --moe-dense-tp-size 1 \
	    --enable-nsa-prefill-context-parallel \
        --nsa-prefill-cp-mode in-seq-split \
        --attn-cp-size 32 \
        --speculative-algorithm NEXTN --speculative-num-steps 1 --speculative-eagle-topk 1 --speculative-num-draft-tokens 2 \
        --dist-init-addr ${P_IP[0]}:10000
        break
    fi
done


# decode
for i in "${!D_IP[@]}";
do
    if [[ "$LOCAL_HOST1" == "${D_IP[$i]}" || "$LOCAL_HOST2" == "${D_IP[$i]}" ]];
    then
        echo "${D_IP[$i]}"
        export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
        export SGLANG_ENABLE_SPEC_V2=1

        export TASK_QUEUE_ENABLE=0
        export SGLANG_SCHEDULER_SKIP_ALL_GATHER=1

        export HCCL_SOCKET_IFNAME=xxx
        export GLOO_SOCKET_IFNAME=xxx

        DP=8
        export HCCL_BUFFSIZE=400
        export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=8

        python3 -m sglang.launch_server --model-path ${MODEL_PATH} \
        --tp 32 \
        --dp ${DP} \
        --ep 32 \
        --moe-dense-tp-size 1 \
        --enable-dp-attention \
        --enable-dp-lm-head \
        --trust-remote-code \
        --attention-backend ascend \
        --device npu \
        --watchdog-timeout 9000 \
        --host ${D_IP[$i]} --port 8001 \
        --mem-fraction-static 0.79 \
        --disable-radix-cache \
        --chunked-prefill-size -1 --max-prefill-tokens 68000 \
        --max-running-requests 32 \
        --cuda-graph-max-bs 4 \
        --moe-a2a-backend deepep \
        --deepep-mode low_latency \
        --quantization modelslim \
        --speculative-algorithm NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 \
        --disaggregation-transfer-backend ascend \
        --disaggregation-mode decode \
        --nnodes 2 --node-rank $i \
        --dist-init-addr ${D_IP[0]}:10000
        break
    fi
done
Command
python -m sglang_router.launch_router \
    --pd-disaggregation \
    --policy cache_aware \
    --prefill http://P_IP1:8000 8995 \
    --decode http://D_IP1:8001 \
    --host 127.0.0.1 \
    --port 6688 \
    --mini-lb

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 6688 --max-concurrency 8  --random-input-len 131076 --random-output-len 1024 --num-prompts 8 --random-range-ratio 1

​
Qwen3-235B-A22B 3_5K-1_5K 50ms on A3 24 Cards Disaggregation Mode

Model: Qwen3-235B-A22B-W8A8 Hardware: Atlas 800I A3 24Card DeployMode: PD Disaggregation Dataset: random Input Output Length: 3.5K+1.5K TPOT: 50ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING

source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash

export SGLANG_SET_CPU_AFFINITY=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1
export SGLANG_DP_ROUND_ROBIN=1
export SGLANG_NPU_FUSED_MOE_MODE=2

MODEL_PATH=xxx
export ASCEND_MF_STORE_URL="tcp://your prefill ip1:24667"
P_IP=('your prefill ip1')
D_IP=('your decode ip1' 'your decode ip2')

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"


for i in "${!P_IP[@]}";
do
    if [[ "$LOCAL_HOST1" == "${P_IP[$i]}" || "$LOCAL_HOST2" == "${P_IP[$i]}" ]];
    then
        echo "${P_IP[$i]}"
        source /usr/local/Ascend/ascend-toolkit/set_env.sh
        source /usr/local/Ascend/nnal/atb/set_env.sh
        export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=188416
        export DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS=1024
        export DEEPEP_NORMAL_LONG_SEQ_ROUND=16
        export HCCL_BUFFSIZE=4300
        export TASK_QUEUE_ENABLE=2
        export HCCL_SOCKET_IFNAME=lo
        export GLOO_SOCKET_IFNAME=lo
        export STREAMS_PER_DEVICE=32
        export DEEP_NORMAL_MODE_USE_INT8_QUANT=1

        # P节点
        python -m sglang.launch_server --model-path ${MODEL_PATH} --disaggregation-mode prefill \
        --host ${P_IP[$i]} --port 8000 --disaggregation-bootstrap-port 8995 --trust-remote-code \
        --nnodes 1 --node-rank $i --tp-size 16 --dp-size 16 --mem-fraction-static 0.6 \
        --disable-radix-cache \
        --attention-backend ascend --device npu --quantization modelslim --disaggregation-transfer-backend ascend \
        --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
        --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 \
        --speculative-draft-model-quantization unquant \
        --max-running-requests 128 --chunked-prefill-size 94208 --max-prefill-tokens 262144 \
        --enable-dp-attention  \
        --moe-a2a-backend ascend_fuseep --dtype bfloat16
        NODE_RANK=$i
        break
    fi
done


for i in "${!D_IP[@]}";
do
    if [[ "$LOCAL_HOST1" == "${D_IP[$i]}" || "$LOCAL_HOST2" == "${D_IP[$i]}" ]];
    then
        echo "${D_IP[$i]}"
        source /usr/local/Ascend/ascend-toolkit/set_env.sh
        source /usr/local/Ascend/nnal/atb/set_env.sh
        export DP_ROUND_ROBIN=1
        export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=65536
        export HCCL_BUFFSIZE=800
        export HCCL_SOCKET_IFNAME=data0.3001
        export GLOO_SOCKET_IFNAME=data0.3001
        export STREAMS_PER_DEVICE=32

        python -m sglang.launch_server --model-path ${MODEL_PATH} --disaggregation-mode decode \
        --host ${D_IP[$i]} --port 8001 --trust-remote-code \
        --nnodes 2 --node-rank $i --tp-size 32 --dp-size 32 --mem-fraction-static 0.83 --max-running-requests 768 \
        --attention-backend ascend --device npu --quantization modelslim --enable-dp-attention \
        --moe-a2a-backend ascend_fuseep --cuda-graph-bs 6 8 12 15 18 20 22 24 \
        --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
        --speculative-draft-model-quantization unquant \
        --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 \
        --dist-init-addr xxx:5000 \
        --disaggregation-transfer-backend ascend --watchdog-timeout 9000 --context-length 8192 \
        --enable-dp-lm-head --dtype bfloat16 --tokenizer-worker-num 4 \
        --load-balance-method round_robin
        NODE_RANK=$i
        break
    fi
done
Command
export SGLANG_DP_ROUND_ROBIN=1
python -m sglang_router.launch_router \
    --pd-disaggregation \
    --policy cache_aware \
    --prefill http://PIP:8000 8995 \
    --decode http://DIP:8001 \
    --host 127.0.0.1 \
    --port 6688 \
    --mini-lb

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python -m sglang.bench_serving --dataset-name random --backend sglang-oai --host 127.0.0.1 --port 7239 --max-concurrency 860 --random-input-len 3500 --random-output-len 1500 --num-prompts 3440 --random-range-ratio 1

​
Qwen3-235B-A22B 3_5K-1_5K 50ms on A3 8 Cards Mixed Mode

Model: Qwen3-235B-A22B-W8A8 Hardware: Atlas 800I A3 8Card DeployMode: PD Mixed Dataset: random Input Output Length: 3.5K+1.5K TPOT: 50ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

export SGLANG_SET_CPU_AFFINITY=1
unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True

MODEL_PATH=xxx

export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_BUFFSIZE=570
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1
export SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE=1
export SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES=100

export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=188416
export SGLANG_NPU_FUSED_MOE_MODE=2

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7439 --trust-remote-code --nnodes 1 --node-rank 0  \
    --attention-backend ascend --device npu --quantization modelslim  \
    --max-running-requests 432 --context-length 8192 --dtype bfloat16 \
    --chunked-prefill-size 94208 --max-prefill-tokens 458880 --sampling-backend ascend \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 \
    --disable-radix-cache --moe-a2a-backend ascend_fuseep --speculative-draft-model-quantization unquant \
    --tp 16 --dp-size 16 --enable-dp-attention --enable-dp-lm-head --mem-fraction-static 0.8 --cuda-graph-bs 1 2 4 8 16 20 24 26 27

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 7439 --max-concurrency 272 --random-input-len 3500 --random-output-len 1500 --num-prompts 1088 --random-range-ratio 1

​
Qwen3-235B-A22B 2K-2K 100ms on A3 8 Cards Mixed Mode

Model: Qwen3-235B-A22B-W8A8 Hardware: Atlas 800I A3 8Card DeployMode: PD Mixed Dataset: random Input Output Length: 2K+2K TPOT: 100ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

export SGLANG_SET_CPU_AFFINITY=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

MODEL_PATH=xxx

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_BUFFSIZE=1200
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1
export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=144

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7439 --trust-remote-code --nnodes 1 --node-rank 0  \
    --attention-backend ascend --device npu --quantization modelslim  \
    --max-running-requests 576 --context-length 8192 --dtype bfloat16 \
    --chunked-prefill-size 32768 --max-prefill-tokens 458880  \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 \
    --disable-radix-cache --moe-a2a-backend deepep  --deepep-mode auto --speculative-draft-model-quantization unquant  \
    --tp 16 --dp-size 16 --enable-dp-attention --enable-dp-lm-head --mem-fraction-static 0.84 --cuda-graph-bs 8 16 20 24 32 36

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 7439 --max-concurrency 576 --random-input-len 2000 --random-output-len 2000 --num-prompts 576 --random-range-ratio 1

​
Qwen3-235B-A22B 2K-2K 50ms on A3 8 Cards Mixed Mode

Model: Qwen3-235B-A22B-W8A8 Hardware: Atlas 800I A3 8Card DeployMode: PD Mixed Dataset: random Input Output Length: 2K+2K TPOT: 50ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

export SGLANG_SET_CPU_AFFINITY=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

MODEL_PATH=xxx

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_BUFFSIZE=450
export HCCL_SOCKET_IFNAME=xxx
export GLOO_SOCKET_IFNAME=xxx
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1
export SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE=1
export SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES=100
export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=147456
export SGLANG_NPU_FUSED_MOE_MODE=2

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7439 --trust-remote-code --nnodes 1 --node-rank 0  \
    --attention-backend ascend --device npu --quantization modelslim  \
    --max-running-requests 624 --context-length 8192 --dtype bfloat16 \
    --chunked-prefill-size 73728 --max-prefill-tokens 458880 --speculative-draft-model-quantization unquant  \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 \
    --disable-radix-cache --moe-a2a-backend ascend_fuseep \
    --tp 16 --dp-size 16 --enable-dp-attention --enable-dp-lm-head --mem-fraction-static 0.83 --cuda-graph-bs 4 8 16 24 28 29 30 32 34 36 37 38 39

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 7439 --max-concurrency 480 --random-input-len 2048 --random-output-len 2048 --num-prompts 480 --random-range-ratio 1

​
Qwen3-235B-A22B 2K-2K 50ms on A3 16 Cards Mixed Mode

Model: Qwen3-235B-A22B-W8A8 Hardware: Atlas 800I A3 16Card DeployMode: PD Mixed Dataset: random Input Output Length: 2K+2K TPOT: 50ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash

export SGLANG_SET_CPU_AFFINITY=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

MODEL_PATH=xxx

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_BUFFSIZE=1600
export HCCL_SOCKET_IFNAME=xxx
export GLOO_SOCKET_IFNAME=xxx
export HCCL_OP_EXPANSION_MODE="AIV"

MIX_IP=('IP1' 'IP2')

for i in "${!MIX_IP[@]}";
do
    if [[ "$LOCAL_HOST1" == "${MIX_IP[$i]}" || "$LOCAL_HOST2" == "${MIX_IP[$i]}" ]];
    then
        echo "${MIX_IP[$i]}"
        export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
        export SGLANG_ENABLE_SPEC_V2=1

        python -m sglang.launch_server --model-path ${MODEL_PATH} \
        --host 127.0.0.1 --port 7439 --trust-remote-code \
        --nnodes 2 --node-rank $i --tp-size 32 --dp-size 32 --mem-fraction-static 0.8 --max-running-requests 768 \
        --attention-backend ascend --device npu --quantization modelslim --enable-dp-attention \
        --moe-a2a-backend deepep --deepep-mode auto --cuda-graph-bs 6 8 10 12 18 24 \
        --dist-init-addr ${MIX_IP[0]}:5000 --chunked-prefill-size 131072 --max-prefill-tokens 458880 \
        --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx --speculative-draft-model-quantization= unquant \
        --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 \
        --context-length 8192 --disable-radix-cache \
        --enable-dp-lm-head --dtype bfloat16
        NODE_RANK=$i
        break
    fi
done

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 7439 --max-concurrency 768 --random-input-len 2000 --random-output-len 2000 --num-prompts 768 --random-range-ratio 1

​
Qwen3-235B-A22B 11K-1K 10ms on A3 8 Cards Mixed Mode

Model: Qwen3-235B-A22B-W8A8 Hardware: Atlas 800I A3 8Card DeployMode: PD Mixed Dataset: random Input Output Length: 11K+1K TPOT: 10ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

export SGLANG_SET_CPU_AFFINITY=1

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True

MODEL_PATH=xxx

export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_BUFFSIZE=1600
export HCCL_SOCKET_IFNAME=xxx
export GLOO_SOCKET_IFNAME=xxx
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7439 --trust-remote-code --nnodes 1 --node-rank 0  \
    --attention-backend ascend --device npu --quantization modelslim  \
    --max-running-requests 1  --dtype bfloat16 \
    --chunked-prefill-size -1 --max-prefill-tokens 16384 --speculative-draft-model-quantization unquant  \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 4 --speculative-eagle-topk 1 --speculative-num-draft-tokens 5 \
    --disable-radix-cache --enable-dp-lm-head \
    --tp 16 --mem-fraction-static 0.78 --cuda-graph-bs 1

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 7439 --max-concurrency 1 --random-input-len 11000 --random-output-len 1000 --num-prompts 1 --random-range-ratio 1

​
Qwen3-32B 6K-1_5K 18ms on A3 4 Cards Mixed Mode

Model: Qwen3-32B Hardware: Atlas 800I A3 4Card DeployMode: PD Mixed Dataset: random Input Output Length: 6K+1.5K TPOT: 18ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

export SGLANG_SET_CPU_AFFINITY=1

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

MODEL_PATH=xxx

export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_BUFFSIZE=400
export HCCL_SOCKET_IFNAME=xxx
export GLOO_SOCKET_IFNAME=xxx
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7439 --trust-remote-code --nnodes 1 --node-rank 0  \
    --attention-backend ascend --device npu \
    --max-running-requests 32 \
    --disable-radix-cache \
    --chunked-prefill-size 24576 --max-prefill-tokens 65536 \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 4 --speculative-eagle-topk 1 --speculative-num-draft-tokens 5 \
    --tp-size 8 --mem-fraction-static 0.72 --cuda-graph-bs 8 16 24 32  --dtype bfloat16

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 7439 --max-concurrency 32 --random-output-len 1500 --random-input-len 6000 --num-prompts 32 --random-range-ratio 1

​
Qwen3-32B 4K-1_5K 11ms on A3 4 Cards Mixed Mode

Model: Qwen3-32B Hardware: Atlas 800I A3 4Card DeployMode: PD Mixed Dataset: random Input Output Length: 4K+1.5K TPOT: 11ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

export SGLANG_SET_CPU_AFFINITY=1
unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

MODEL_PATH=xxx

export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_BUFFSIZE=400
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7339 --trust-remote-code --nnodes 1 --node-rank 0  \
    --attention-backend ascend --device npu   \
    --max-running-requests 1 \
    --disable-radix-cache \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 4 --speculative-eagle-topk 1 --speculative-num-draft-tokens 5 \
    --chunked-prefill-size 24576 --max-prefill-tokens 65536  \
    --tp-size 8 --mem-fraction-static 0.72 --cuda-graph-bs 1 --dtype bfloat16

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 7239 --random-range-ratio 1 --max-concurrency 1 --random-output-len 1500 --random-input-len 4096 --num-prompts 4

​
Qwen3-32B 18K-4K 6ms on A3 8 Cards Mixed Mode

Model: Qwen3-32B Hardware: Atlas 800I A3 8Card DeployMode: PD Mixed Dataset: random Input Output Length: 18K+4K TPOT: 6ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

export SGLANG_SET_CPU_AFFINITY=1
unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

MODEL_PATH=xxx

export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_BUFFSIZE=400
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7339 --trust-remote-code --nnodes 1 --node-rank 0  \
    --attention-backend ascend --device npu   \
    --max-running-requests 1 \
    --disable-radix-cache --speculative-draft-model-quantization unquant \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 4 --speculative-eagle-topk 1 --speculative-num-draft-tokens 5 \
    --chunked-prefill-size -1 --max-prefill-tokens 65536  \
    --tp-size 16 --mem-fraction-static 0.72 --cuda-graph-bs 1 --dtype bfloat16

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 7339 --random-range-ratio 1 --max-concurrency 1 --random-output-len 18000 --random-input-len 4000 --num-prompts 1

​
Qwen3-32B 3_5K-1_5K 50ms on A3 2 Cards Mixed Mode

Model: Qwen3-32B Hardware: Atlas 800I A3 2Card DeployMode: PD Mixed Dataset: random Input Output Length: 3.5K+1.5K TPOT: 50ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

export SGLANG_SET_CPU_AFFINITY=1
unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH


MODEL_PATH=xxx

export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_BUFFSIZE=400
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7239 --trust-remote-code --nnodes 1 --node-rank 0  \
    --attention-backend ascend --device npu  --quantization modelslim  \
    --max-running-requests 78 \
    --disable-radix-cache --speculative-draft-model-quantization unquant \
    --chunked-prefill-size -1 --max-prefill-tokens 49152  \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 \
    --tp-size 4  --mem-fraction-static 0.72 --cuda-graph-bs 16 32 64 68 72 78 --dtype bfloat16

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 7239 --max-concurrency 78 --random-output-len 1500 --random-input-len 3500 --num-prompts 312 --random-range-ratio 1

​
Qwen3-32B 2K-2K 50ms on A3 2 Cards Mixed Mode

Model: Qwen3-32B Hardware: Atlas 800I A3 2Card DeployMode: PD Mixed Dataset: random Input Output Length: 2K+2K TPOT: 50ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

export SGLANG_SET_CPU_AFFINITY=1
unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

MODEL_PATH=xxx

export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_BUFFSIZE=400
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7239 --trust-remote-code --nnodes 1 --node-rank 0  \
    --attention-backend ascend --device npu  --quantization modelslim  \
    --max-running-requests 120 \
    --disable-radix-cache --speculative-draft-model-quantization unquant \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 \
    --chunked-prefill-size -1 --max-prefill-tokens 49152 \
    --tp-size 4 --mem-fraction-static 0.7 --cuda-graph-bs 54 60 66 72 78 84 90 108 114 120 --dtype bfloat16

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 7239 --max-concurrency 120 --random-output-len 2000 --random-input-len 2000 --num-prompts 480 --random-range-ratio 1

​
Qwen3-30B-A3B 3_5K-1_5K 50ms on A3 1 Card Mixed Mode

Model: Qwen3-30B-A3B-Instruct-2507 Hardware: Atlas 800I A3 1Card DeployMode: PD Mixed Dataset: random Input Output Length: 3.5K+1.5K TPOT: 50ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

MODEL_PATH=xxx

export SGLANG_SET_CPU_AFFINITY=1
export ASCEND_LAUNCH_BLOCKING=0
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_BUFFSIZE=400
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE=1
export SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES=200
export SGLANG_ENABLE_SPEC_V2=1

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7239 --trust-remote-code --nnodes 1 --node-rank 0  \
    --attention-backend ascend --device npu  --quantization modelslim  \
    --max-running-requests 162 \
    --disable-radix-cache \
    --speculative-draft-model-quantization unquant \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 \
    --chunked-prefill-size -1 --max-prefill-tokens 35000 \
    --tp-size 2 --mem-fraction-static 0.87 --cuda-graph-bs 1 5 15 40 70 100 120 130 140 146 150 154 156 158 160 162 \
    --dtype bfloat16

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 7239 --max-concurrency 156 --random-input-len 3500 --random-output-len 1500 --num-prompts 624 --random-range-ratio 1

​
Qwen3-Coder-480B-A35B-Instruct 3_5K-1_5K 50ms on A3 24 Cards Disaggregation Mode

Model: Qwen3-Coder-480B-A35B-Instruct Hardware: Atlas 800I A3 24Card DeployMode: PD Disaggregation Dataset: random Input Output Length: 3.5K+1.5K TPOT: 50ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING

source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash

export SGLANG_SET_CPU_AFFINITY=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export STREAMS_PER_DEVICE=32
export SGLANG_NPU_FUSED_MOE_MODE=2

MODEL_PATH=xxx
export ASCEND_MF_STORE_URL="tcp://PIP:24667"
P_IP=('PIP')
D_IP=('DIP1' 'DIP2')
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"


for i in "${!P_IP[@]}";
do
    if [[ "$LOCAL_HOST1" == "${P_IP[$i]}" || "$LOCAL_HOST2" == "${P_IP[$i]}" ]];
    then
        echo "${P_IP[$i]}"
        source /usr/local/Ascend/ascend-toolkit/set_env.sh
        source /usr/local/Ascend/nnal/atb/set_env.sh
        export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=327680
        export HCCL_BUFFSIZE=1550
        export TASK_QUEUE_ENABLE=2
        export HCCL_SOCKET_IFNAME=lo
        export GLOO_SOCKET_IFNAME=lo
        export DEEP_NORMAL_MODE_USE_INT8_QUANT=1

        python -m sglang.launch_server --model-path ${MODEL_PATH} --disaggregation-mode prefill \
        --host ${P_IP[$i]} --port 8000 --disaggregation-bootstrap-port 8995 --trust-remote-code \
        --nnodes 1 --node-rank $i --tp-size 16 --dp-size 2 --mem-fraction-static 0.7 \
        --disable-radix-cache \
	    --attention-backend ascend --device npu --quantization modelslim --disaggregation-transfer-backend ascend \
	    --max-running-requests 16 --chunked-prefill-size 20480 --max-prefill-tokens 20480 \
        --enable-dp-attention  \
        --moe-a2a-backend ascend_fuseep --dtype bfloat16 \
        --disable-overlap-schedule
        NODE_RANK=$i
        break
    fi
done

for i in "${!D_IP[@]}";
do
    if [[ "$LOCAL_HOST1" == "${D_IP[$i]}" || "$LOCAL_HOST2" == "${D_IP[$i]}" ]];
    then
        echo "${D_IP[$i]}"
        source /usr/local/Ascend/ascend-toolkit/set_env.sh
        source /usr/local/Ascend/nnal/atb/set_env.sh
        export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=65536
        export HCCL_BUFFSIZE=600
        export SGLANG_NPU_FUSED_MOE_MODE=2
        export HCCL_SOCKET_IFNAME=xxx
        export GLOO_SOCKET_IFNAME=xxx

        python -m sglang.launch_server --model-path ${MODEL_PATH} --disaggregation-mode decode \
        --host ${D_IP[$i]} --port 8001 --trust-remote-code \
        --nnodes 2 --node-rank $i --tp-size 32 --dp-size 4 --mem-fraction-static 0.75 --max-running-requests 544 \
        --attention-backend ascend --device npu --quantization modelslim --enable-dp-attention \
        --moe-a2a-backend ascend_fuseep --cuda-graph-bs 16 32 56 72 80 88 96 104 112 120 128 136 \
        --dist-init-addr DIP1:5000 \
	    --disaggregation-transfer-backend ascend --watchdog-timeout 9000 --context-length 8192 \
        --enable-dp-lm-head --dtype bfloat16 --tokenizer-worker-num 4 --load-balance-method round_robin
        NODE_RANK=$i
        break
    fi
done
Command
python -m sglang_router.launch_router \
    --pd-disaggregation \
    --policy cache_aware \
    --prefill http://PIP:8000 8995 \
    --decode http://DIP:8001 \
    --host 127.0.0.1 \
    --port 6688 \
    --mini-lb

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 7239 --max-concurrency 410 --random-input-len 3500 --random-output-len 1500 --num-prompts 1640 --random-range-ratio 1 --request-rate 8

​
Qwen3-Coder-480B-A35B-Instruct 3_5K-1_5K 50ms on A3 16 Cards Mixed Mode

Model: Qwen3-Coder-480B-A35B-Instruct Hardware: Atlas 800I A3 16Card DeployMode: PD Mixed Dataset: random Input Output Length: 3.5K+1.5K TPOT: 50ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash

export SGLANG_SET_CPU_AFFINITY=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=72
export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

MODEL_PATH=xxx

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_BUFFSIZE=1800
export HCCL_SOCKET_IFNAME=xxx
export GLOO_SOCKET_IFNAME=xxx
export HCCL_OP_EXPANSION_MODE="AIV"

MIX_IP=('IP1' 'IP2')

for i in "${!MIX_IP[@]}";
do
    if [[ "$LOCAL_HOST1" == "${MIX_IP[$i]}" || "$LOCAL_HOST2" == "${MIX_IP[$i]}" ]];
    then
        echo "${MIX_IP[$i]}"

        python -m sglang.launch_server --model-path $MODEL_PATH \
        --host 127.0.0.1 --port 7439 --trust-remote-code --nnodes 2 --node-rank $i  \
        --dist-init-addr 141.61.133.128:5000 \
        --attention-backend ascend --device npu --quantization modelslim  \
        --max-running-requests 288 --context-length 8192 --dtype bfloat16  \
        --chunked-prefill-size 114688 --max-prefill-tokens 458880  \
        --disable-radix-cache --moe-a2a-backend deepep  --deepep-mode auto  \
        --tp 32 --dp-size 4 --enable-dp-attention --enable-dp-lm-head --mem-fraction-static 0.7 --cuda-graph-bs 56 64 72
        NODE_RANK=$i
        break
    fi
done

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 7439 --max-concurrency 288 --random-input-len 3500 --random-output-len 1500 --num-prompts 1152 --random-range-ratio 1 --request-rate 20

​
Qwen3-Coder-480B-A35B-Instruct 3_5K-1_5K 50ms on A3 8 Cards Mixed Mode

Model: Qwen3-Coder-480B-A35B-Instruct Hardware: Atlas 800I A3 8Card DeployMode: PD Mixed Dataset: random Input Output Length: 3.5K+1.5K TPOT: 50ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

export SGLANG_SET_CPU_AFFINITY=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

MODEL_PATH=xxx

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_BUFFSIZE=2100
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export HCCL_OP_EXPANSION_MODE="AIV"

python -m sglang.launch_server --model-path $MODEL_PATH \
--host 127.0.0.1 --port 7439 --trust-remote-code --nnodes 1 --node-rank 0  \
--attention-backend ascend --device npu --quantization modelslim  \
--max-running-requests 80 --context-length 8192 --dtype bfloat16 \
--chunked-prefill-size 28672 --max-prefill-tokens 458880  \
--disable-radix-cache --moe-a2a-backend deepep  --deepep-mode auto --enable-dp-attention --enable-dp-lm-head \
--tp 16 --dp-size 4 --mem-fraction-static 0.7 --cuda-graph-bs  16 20 24

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 7439 --max-concurrency 80 --random-input-len 3500 --random-output-len 1500 --num-prompts 320 --random-range-ratio 1

​
Qwen3-Next-80B-A3B-Instruct 3_5K-1_5K 50ms on A3 2 Cards Mixed Mode

Model: Qwen3-Next-80B-A3B-Instruct Hardware: Atlas 800I A3 2Card DeployMode: PD Mixed Dataset: random Input Output Length: 3.5K+1.5K TPOT: 50ms

​
Model Deployment

Command
export cann_path=/usr/local/Ascend/ascend-toolkit/latest
source /usr/local/Ascend/driver/bin/setenv.bash
source ${cann_path}/../set_env.sh
source ${cann_path}/../../nnal/atb/set_env.sh
source ${cann_path}/opp/vendors/customize/bin/set_env.bash
export ASCEND_HOME_PATH=${cann_path}
source /usr/local/Ascend/8.5.0/bisheng_toolkit/set_env.sh

echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

export SGLANG_SET_CPU_AFFINITY=1

export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export STREAMS_PER_DEVICE=32
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo

export HCCL_OP_EXPANSION_MODE=AIV
export HCCL_ALGO="level0:NA;level1:ring"

export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=20
export HCCL_BUFFSIZE=2000

python -m sglang.launch_server \
        --model-path /path/to/Qwen3-Next-80B-A3B-Instruct-W8A8-3 \
        --host 127.0.0.1 \
        --port 6699 \
        --tp-size 4 \
        --device npu \
        --attention-backend ascend \
        --mem-fraction-static 0.685 \
        --max-running-requests 80 \
        --watchdog-timeout 3600 \
        --disable-radix-cache \
        --cuda-graph-bs 80 \
        --max-prefill-tokens 28672  --max-total-tokens 450560 \
        --moe-a2a-backend deepep --deepep-mode auto \
        --quantization modelslim \
        --chunked-prefill-size -1

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 6699 --max-concurrency 80 --random-output-len 1536 --random-input-len 3584 --num-prompts 160 --random-range-ratio 1

​
Qwen3-32B 6K-1_5K 18ms on A2 8 Cards Mixed Mode

Model: Qwen3-32B Hardware: Atlas 800I A2 8Card DeployMode: PD Mixed Dataset: random Input Output Length: 6K+1.5K TPOT: 18ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

export SGLANG_SET_CPU_AFFINITY=1
unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

MODEL_PATH=xxx

export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_BUFFSIZE=400
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7439 --trust-remote-code --nnodes 1 --node-rank 0  \
    --attention-backend ascend --device npu  --quantization modelslim  \
    --max-running-requests 32 \
    --disable-radix-cache \
    --chunked-prefill-size 24576 --max-prefill-tokens 65536 \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 4 --speculative-eagle-topk 1 --speculative-num-draft-tokens 5 \
    --tp-size 8 --mem-fraction-static 0.72 --cuda-graph-bs 8 16 24 32 --dtype bfloat16

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 7439 --max-concurrency 32 --random-output-len 1500 --random-input-len 6000 --num-prompts 32 --random-range-ratio 1

​
Qwen3-32B 4K-1_5K 11ms on A2 8 Cards Mixed Mode

Model: Qwen3-32B Hardware: Atlas 800I A2 8Card DeployMode: PD Mixed Dataset: random Input Output Length: 4K+1.5K TPOT: 11ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

export SGLANG_SET_CPU_AFFINITY=1

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True

MODEL_PATH=xxx

export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_BUFFSIZE=400
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7339 --trust-remote-code --nnodes 1 --node-rank 0  \
    --attention-backend ascend --device npu   \
    --max-running-requests 32 \
    --disable-radix-cache \
    --speculative-draft-model-quantization unquant \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx  \
    --speculative-num-steps 4 --speculative-eagle-topk 1 --speculative-num-draft-tokens 5 \
    --chunked-prefill-size -1 --max-prefill-tokens 65536  \
    --tp-size 8 --mem-fraction-static 0.72 --cuda-graph-bs 1 4 6 12 18 24 30 32 --dtype bfloat16

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving  --dataset-name random --backend sglang --host 127.0.0.1 --port 7339 --random-range-ratio 1 --max-concurrency 1 --random-output-len 1500 --random-input-len 4096 --num-prompts 4

​
Qwen3-32B 1K-0_3K 12ms on A3 2 Cards Mixed Mode

Model: Qwen3-32B Hardware: Atlas 800I A3 2Card DeployMode: PD Mixed Dataset: random Input Output Length: 1K+0.3K TPOT: 12ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

export SGLANG_SET_CPU_AFFINITY=1

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True

MODEL_PATH=xxx

export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7339 --trust-remote-code --nnodes 1 --node-rank 0  \
    --attention-backend ascend --device npu --quantization modelslim \
    --max-running-requests 16 \
    --disable-radix-cache \
    --speculative-draft-model-quantization unquant \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 \
    --chunked-prefill-size -1 --max-prefill-tokens 16384  \
    --tp-size 4 --mem-fraction-static 0.843 --cuda-graph-bs 1 4 8 16 --dtype bfloat16

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving  --dataset-name random --backend sglang --host 127.0.0.1 --port 7339 --random-range-ratio 1 --max-concurrency 16 --random-output-len 300 --random-input-len 1024 --num-prompts 16

​
Qwen3-32B 6K-1_5K 17ms on A3 2 Cards Mixed Mode

Model: Qwen3-32B Hardware: Atlas 800I A3 2Card DeployMode: PD Mixed Dataset: random Input Output Length: 6K+1.5K TPOT: 17ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

export SGLANG_SET_CPU_AFFINITY=1

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True

MODEL_PATH=xxx

export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7339 --trust-remote-code --nnodes 1 --node-rank 0  \
    --attention-backend ascend --device npu --quantization modelslim \
    --max-running-requests 16 \
    --disable-radix-cache \
    --speculative-draft-model-quantization unquant \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 \
    --chunked-prefill-size -1 --max-prefill-tokens 16384  \
    --tp-size 4 --mem-fraction-static 0.843 --cuda-graph-bs 1 4 10 15 16 --dtype bfloat16

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving  --dataset-name random --backend sglang --host 127.0.0.1 --port 7339 --random-range-ratio 1 --max-concurrency 16 --random-output-len 1500 --random-input-len 6144 --num-prompts 16

​
Qwen3-8B 1K-0_3K 7ms on A3 1 Cards Mixed Mode

Model: Qwen3-8B Hardware: Atlas 800I A3 1Card DeployMode: PD Mixed Dataset: random Input Output Length: 1K+0.3K TPOT: 7ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

export SGLANG_SET_CPU_AFFINITY=1

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True

MODEL_PATH=xxx

export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7339 --trust-remote-code --nnodes 1 --node-rank 0  \
    --attention-backend ascend --device npu --quantization modelslim \
    --max-running-requests 16 \
    --disable-radix-cache \
    --speculative-draft-model-quantization unquant \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 4 --speculative-eagle-topk 1 --speculative-num-draft-tokens 5 \
    --chunked-prefill-size -1 --max-prefill-tokens 16384  \
    --tp-size 2 --mem-fraction-static 0.894 --cuda-graph-bs 1 2 4 6 9 10 15 16 --dtype bfloat16

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving  --dataset-name random --backend sglang --host 127.0.0.1 --port 7339 --random-range-ratio 1 --max-concurrency 16 --random-output-len 300 --random-input-len 1024 --num-prompts 16

​
Qwen3-8B 6K-1_5K 12ms on A3 1 Cards Mixed Mode

Model: Qwen3-8B Hardware: Atlas 800I A3 1Card DeployMode: PD Mixed Dataset: random Input Output Length: 6K+1.5K TPOT: 12ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

export SGLANG_SET_CPU_AFFINITY=1

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True

MODEL_PATH=xxx

export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7339 --trust-remote-code --nnodes 1 --node-rank 0  \
    --attention-backend ascend --device npu --quantization modelslim \
    --max-running-requests 16 \
    --disable-radix-cache \
    --speculative-draft-model-quantization unquant \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 4 --speculative-eagle-topk 1 --speculative-num-draft-tokens 5 \
    --chunked-prefill-size -1 --max-prefill-tokens 16384  \
    --tp-size 2 --mem-fraction-static 0.894 --cuda-graph-bs 1 5 15 16 --dtype bfloat16

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving  --dataset-name random --backend sglang --host 127.0.0.1 --port 7339 --random-range-ratio 1 --max-concurrency 16 --random-output-len 1500 --random-input-len 6144 --num-prompts 16

​
Qwen3-32B 3_5K-1_5K 50ms on A2 8 Cards Mixed Mode

Model: Qwen3-32B Hardware: Atlas 800I A2 8Card DeployMode: PD Mixed Dataset: random Input Output Length: 3.5K+1.5K TPOT: 50ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

export SGLANG_SET_CPU_AFFINITY=1

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

MODEL_PATH=xxx

export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_BUFFSIZE=400
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7239 --trust-remote-code --nnodes 1 --node-rank 0  \
    --attention-backend ascend --device npu  --quantization modelslim  \
    --max-running-requests 78 \
    --disable-radix-cache --speculative-draft-model-quantization unquant \
    --chunked-prefill-size -1 --max-prefill-tokens 65536  \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 \
    --tp-size 4  --mem-fraction-static 0.72 --cuda-graph-bs 1 4 8 16 32 64 68 72 78 --dtype bfloat16 --base-gpu-id 4

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 7239 --max-concurrency 78 --random-output-len 1500 --random-input-len 3500 --num-prompts 312 --random-range-ratio 1

​
Qwen3-32B 2K-2K 50ms on A2 8 Cards Mixed Mode

Model: Qwen3-32B Hardware: Atlas 800I A2 8Card DeployMode: PD Mixed Dataset: random Input Output Length: 2K+2K TPOT: 50ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

export SGLANG_SET_CPU_AFFINITY=1
unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

MODEL_PATH=xxx

export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_BUFFSIZE=400
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7239 --trust-remote-code --nnodes 1 --node-rank 0  \
    --attention-backend ascend --device npu  --quantization modelslim  \
    --max-running-requests 120 \
    --disable-radix-cache \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 --speculative-draft-model-quantization unquant \
    --chunked-prefill-size -1 --max-prefill-tokens 49152 --base-gpu-id 4 \
    --tp-size 4 --mem-fraction-static 0.7 --cuda-graph-bs 54 60 66 72 78 84 90 108 114 120 --dtype bfloat16

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 7239 --max-concurrency 120 --random-output-len 2000 --random-input-len 2000 --num-prompts 120 --random-range-ratio 1

​
Qwen3-30B-A3B 6K-1_5K 10ms on A3 1 Cards Mixed Mode

Model: Qwen3-30B-A3B Hardware: Atlas 800I A3 1Card DeployMode: PD Mixed Dataset: random Input Output Length: 6K+1.5K TPOT: 10ms

​
Model Deployment

Command
export SGLANG_SET_CPU_AFFINITY=1

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING

export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True

MODEL_PATH=xxx

export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_BUFFSIZE=400
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7339 --trust-remote-code --nnodes 1 --node-rank 0  \
    --attention-backend ascend --device npu --quantization modelslim \
    --max-running-requests 16 \
    --disable-radix-cache \
    --speculative-draft-model-quantization unquant \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 4 --speculative-eagle-topk 1 --speculative-num-draft-tokens 5 \
    --chunked-prefill-size -1 --max-prefill-tokens 35000  \
    --tp-size 2 --mem-fraction-static 0.6 --cuda-graph-bs 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 --dtype bfloat16

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving  --dataset-name random --backend sglang --host 127.0.0.1 --port 7339 --random-range-ratio 1 --max-concurrency 16 --random-output-len 1500 --random-input-len 6144 --num-prompts 16

​
Qwen3-30B-A3B 1K-0_3K 7ms on A3 1 Cards Mixed Mode

Model: Qwen3-30B-A3B Hardware: Atlas 800I A3 1Card DeployMode: PD Mixed Dataset: random Input Output Length: 1K+0.3K TPOT: 7ms

​
Model Deployment

Command
export SGLANG_SET_CPU_AFFINITY=1

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING

export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True

MODEL_PATH=xxx

export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_BUFFSIZE=400
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7339 --trust-remote-code --nnodes 1 --node-rank 0  \
    --attention-backend ascend --device npu --quantization modelslim \
    --max-running-requests 8 \
    --disable-radix-cache \
    --speculative-draft-model-quantization unquant \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 4 --speculative-eagle-topk 1 --speculative-num-draft-tokens 5 \
    --chunked-prefill-size -1 --max-prefill-tokens 35000  \
    --tp-size 2 --mem-fraction-static 0.7 --cuda-graph-bs 1 2 3 4 5 6 7 8 --dtype bfloat16

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving  --dataset-name random --backend sglang --host 127.0.0.1 --port 7339 --random-range-ratio 1 --max-concurrency 8 --random-output-len 300 --random-input-len 1024 --num-prompts 8

​
Qwen3-Next 1K-0_3K 14_21ms on A3 2 Cards Mixed Mode

Model: Qwen3-Next-80B-A3B-Instruct Hardware: Atlas 800I A3 2Card DeployMode: PD Mixed Dataset: random Input Output Length: 1K+0.3K TPOT: 14.21ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

export SGLANG_SET_CPU_AFFINITY=1

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export STREAMS_PER_DEVICE=32
export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=330
export DEEPEP_NORMAL_LONG_SEQ_ROUND=5
export DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS=3000
export DEEPEP_NORMAL_COMBINE_ENABLE_LONG_SEQ=1

export ASCEND_USE_FIA=1
export SGLANG_NPU_USE_MULTI_STREAM=1

export SGLANG_WARMUP_TIMEOUT=3600
export SGLANG_ENABLE_SPEC_V2=1
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export FORCE_DRAFT_MODEL_NON_QUANT=1

MODEL_PATH=xxx

export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_BUFFSIZE=2000
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1

python3 -m sglang.launch_server --model-path ${MODEL_PATH} \
    --page-size 128 \
    --tp-size 4 \
    --trust-remote-code \
    --attention-backend ascend \
    --device npu \
    --watchdog-timeout 9000 \
    --host 127.0.0.1 --port 6699 \
    --mem-fraction-static 0.75 \
    --disable-radix-cache --max-prefill-tokens 14080 --context-length 26384 \
    --speculative-algorithm NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4  --speculative-draft-model-quantization  unquant \
    --chunked-prefill-size -1 --max-running-requests 312 \
    --cuda-graph-bs 2 4 16 32 48 64 80 96 128 140 156 \
    --mamba-ssm-dtype bfloat16 \
    --base-gpu-id 0 \
    --speculative-draft-model-path /home/weights/Qwen3-Next-80B-A3B-Instruct \
    --quantization modelslim \
    --moe-a2a-backend deepep --deepep-mode auto \

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving  --dataset-name random --backend sglang --host 127.0.0.1 --port 6699 --random-range-ratio 1 --max-concurrency 16 --random-output-len 300 --random-input-len 1024 --num-prompts 16

​
Qwen3-Next 6K-1_5K 15_62ms on A3 2 Cards Mixed Mode

Model: Qwen3-Next-80B-A3B-Instruct Hardware: Atlas 800I A3 2Card DeployMode: PD Mixed Dataset: random Input Output Length: 6K+1.5K TPOT: 15.62ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

export SGLANG_SET_CPU_AFFINITY=1

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export STREAMS_PER_DEVICE=32
export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=330
export DEEPEP_NORMAL_LONG_SEQ_ROUND=5
export DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS=3000
export DEEPEP_NORMAL_COMBINE_ENABLE_LONG_SEQ=1

export ASCEND_USE_FIA=1
export SGLANG_NPU_USE_MULTI_STREAM=1

export SGLANG_WARMUP_TIMEOUT=3600
export SGLANG_ENABLE_SPEC_V2=1
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export FORCE_DRAFT_MODEL_NON_QUANT=1

MODEL_PATH=xxx

export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

export HCCL_BUFFSIZE=2000
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1

python3 -m sglang.launch_server --model-path ${MODEL_PATH} \
    --page-size 128 \
    --tp-size 4 \
    --trust-remote-code \
    --attention-backend ascend \
    --device npu \
    --watchdog-timeout 9000 \
    --host 127.0.0.1 --port 6699 \
    --mem-fraction-static 0.75 \
    --disable-radix-cache --max-prefill-tokens 14080 --context-length 26384 \
    --speculative-algorithm NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4  --speculative-draft-model-quantization  unquant \
    --chunked-prefill-size -1 --max-running-requests 312 \
    --cuda-graph-bs 2 4 16 32 48 64 80 96 128 140 156 \
    --mamba-ssm-dtype bfloat16 \
    --base-gpu-id 0 \
    --speculative-draft-model-path /home/weights/Qwen3-Next-80B-A3B-Instruct \
    --quantization modelslim \
    --moe-a2a-backend deepep --deepep-mode auto \

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving  --dataset-name random --backend sglang --host 127.0.0.1 --port 6699 --random-range-ratio 1 --max-concurrency 16 --random-output-len 1500 --random-input-len 6144 --num-prompts 16

​
Qwen3-14B 3_5K-1_5K 9ms on A3 1 Cards Mixed Mode

Model: Qwen3-14B Hardware: Atlas 800I A3 1Card DeployMode: PD Mixed Dataset: random Input Output Length: 3.5K+1.5K TPOT: 9ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

MODEL_PATH=xxx

export SGLANG_SET_CPU_AFFINITY=1
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export HCCL_OP_EXPANSION_MODE="AIV"
export STREAMS_PER_DEVICE=32
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export ASCEND_USE_FIA=0
export SGLANG_ENABLE_SPEC_V2=1
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7239 --trust-remote-code --nnodes 1 --node-rank 0 \
    --attention-backend ascend --device npu --quantization modelslim \
    --disable-radix-cache --mem-fraction-static 0.8 \
    --tp-size 1 --dp-size 1 \
    --sampling-backend ascend --max-running-requests 8 \
    --served-model-name Qwen3-14B \
    --chunked-prefill-size -1 \
    --cuda-graph-bs 8 \
    --dtype bfloat16 \
    --speculative-draft-model-quantization unquant \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 \
    --schedule-conservativeness 0.01

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 7239 --max-concurrency 1 --random-output-len 1500 --random-input-len 3500 --num-prompts 8 --random-range-ratio 1

​
Qwen3-14B 3_5K-1_5K 50ms on A3 1 Cards Mixed Mode

Model: Qwen3-14B Hardware: Atlas 800I A3 1Card DeployMode: PD Mixed Dataset: random Input Output Length: 3.5K+1.5K TPOT: 50ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

MODEL_PATH=xxx

export SGLANG_SET_CPU_AFFINITY=1
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export HCCL_OP_EXPANSION_MODE="AIV"
export STREAMS_PER_DEVICE=32
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
export ASCEND_USE_FIA=0
export SGLANG_ENABLE_SPEC_V2=1
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE=1
export SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES=200

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7239 --trust-remote-code --nnodes 1 --node-rank 0 \
    --attention-backend ascend --device npu --quantization modelslim \
    --disable-radix-cache --mem-fraction-static 0.89 \
    --tp-size 1 --dp-size 2 \
    --sampling-backend ascend --max-running-requests 144 \
    --max-prefill-tokens 12288 \
    --served-model-name Qwen3-14B \
    --chunked-prefill-size -1 \
    --cuda-graph-bs 8 16 32 44 48 50 52 \
    --dtype bfloat16 \
    --speculative-draft-model-quantization unquant \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 \
    --schedule-conservativeness 0.01

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 7239 --max-concurrency 144 --random-output-len 1500 --random-input-len 3500 --num-prompts 576 --random-range-ratio 1

​
Qwen3-8B 3_5K-1_5K 50ms on A3 1 Cards Mixed Mode

Model: Qwen3-8B Hardware: Atlas 800I A3 1Card DeployMode: PD Mixed Dataset: random Input Output Length: 3.5K+1.5K TPOT: 50ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

MODEL_PATH=xxx

export SGLANG_SET_CPU_AFFINITY=1
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1
export SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE=1
export SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES=50
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7239 --trust-remote-code --nnodes 1 --node-rank 0 \
    --attention-backend ascend --device npu --quantization modelslim \
    --disable-radix-cache --mem-fraction-static 0.9 \
    --tp-size 1 \
    --max-running-requests 70 \
    --max-prefill-tokens 16384 \
    --served-model-name Qwen3-8B \
    --chunked-prefill-size 16384 \
    --cuda-graph-bs 8 12 24 36 48 51 55 60 63 64 66 68 70 \
    --dtype bfloat16 \
    --speculative-draft-model-quantization unquant \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 7239 --max-concurrency 64 --random-output-len 1500 --random-input-len 3500 --num-prompts 256 --random-range-ratio 1

​
Qwen3-8B 3_5K-1_5K 5ms on A3 1 Cards Mixed Mode

Model: Qwen3-8B Hardware: Atlas 800I A3 1Card DeployMode: PD Mixed Dataset: random Input Output Length: 3.5K+1.5K TPOT: 5ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

MODEL_PATH=xxx

export SGLANG_SET_CPU_AFFINITY=1
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export HCCL_OP_EXPANSION_MODE="AIV"
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

python -m sglang.launch_server --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 7239 --trust-remote-code --nnodes 1 --node-rank 0 \
    --attention-backend ascend --device npu --quantization modelslim \
    --disable-radix-cache --mem-fraction-static 0.894 \
    --tp-size 2 \
    --max-running-requests 1 \
    --max-prefill-tokens 16384 \
    --served-model-name Qwen3-8B \
    --chunked-prefill-size -1 \
    --cuda-graph-bs 1 \
    --dtype bfloat16 \
    --speculative-draft-model-quantization unquant \
    --speculative-algorithm EAGLE3 --speculative-draft-model-path xxx \
    --speculative-num-steps 4 --speculative-eagle-topk 1 --speculative-num-draft-tokens 5

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 7239 --max-concurrency 1 --random-output-len 1500 --random-input-len 3500 --num-prompts 4 --random-range-ratio 1

​
Qwen3-Next 3_5K-1_5K 20ms on A3 2 Cards Mixed Mode

Model: Qwen3-Next-80B-A3B-Instruct Hardware: Atlas 800I A3 2Card DeployMode: PD Mixed Dataset: random Input Output Length: 3.5K+1.5K TPOT: 20ms

​
Model Deployment

Command
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
export PATH=/usr/local/Ascend/8.5.0/compiler/bishengir/bin:$PATH

export SGLANG_SET_CPU_AFFINITY=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export STREAMS_PER_DEVICE=32
export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=400
export DEEPEP_NORMAL_LONG_SEQ_ROUND=10
export DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS=2048
export HCCL_OP_EXPANSION_MODE="AIV"
export TASK_QUEUE_ENABLE=1
export ASCEND_USE_FIA=1
export SGLANG_NPU_USE_MULTI_STREAM=0
export SGLANG_WARMUP_TIMEOUT=3600
export SGLANG_ENABLE_SPEC_V2=1
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export FORCE_DRAFT_MODEL_NON_QUANT=1
export HCCL_BUFFSIZE=2000
export ZBCCL_LOCAL_MEM_SIZE=60416
export SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK=0

export ZBCCL_BOOTSTRAP_URL=tcp://127.0.0.1:24669
export ZBCCL_NPU_ALLOC_CONF=use_vmm_for_static_memory:True
export ZBCCL_ENABLE_GRAPH=1

export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo

MODEL_PATH=xxx

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`

echo "${LOCAL_HOST1}"
echo "${LOCAL_HOST2}"

python3 -m sglang.launch_server --model-path ${MODEL_PATH} \
    --page-size 128 \
    --tp-size 4 --dp-size 2 \
    --trust-remote-code \
    --attention-backend ascend \
    --device npu \
    --quantization modelslim \
    --watchdog-timeout 9000 \
    --host 127.0.0.1 --port 6699 \
    --mem-fraction-static 0.85 \
    --disable-radix-cache --max-prefill-tokens 28672 --context-length 26384 --max-total-tokens 122304 \
    --enable-dp-attention --enable-dp-lm-head \
    --speculative-algorithm NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 --speculative-draft-model-quantization unquant \
    --chunked-prefill-size -1 --max-running-requests 16 \
    --cuda-graph-bs 2 4 8 \
    --mamba-ssm-dtype bfloat16 \
    --speculative-draft-model-path /path/to/Qwen3-Next-80B-A3B-Instruct

​
Benchmark

We tested it based on the RANDOM dataset.
Command
python3 -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 6699 --random-range-ratio 1 --max-concurrency 1 --random-output-len 1500 --random-input-len 3500 --num-prompts 1