DeepSeek-R1 - SGLang Documentation

This guide describes the best practice data for DeepSeek-R1 on the Ascend NPU.

Low Latency

Model	Hardware	Cards	Deploy Mode	Dataset	TPOT	Quantization	Configuration
DeepSeek-R1	Atlas 800I A3	32	PD Disaggregation	3.5K+1.5K	18.9ms	W8A8 INT8	Optimal Configuration
DeepSeek-R1	Atlas 800I A3	32	PD Disaggregation	3.5K+1K	19.0ms	W8A8 INT8	Optimal Configuration
DeepSeek-R1	Atlas 800I A3	32	PD Disaggregation	3.9K+1K	19.0ms	W8A8 INT8	Optimal Configuration
DeepSeek-R1	Atlas 800I A3	32	PD Disaggregation	6K+1.6K	20.5ms	W8A8 INT8	Optimal Configuration

High Throughput

Model	Hardware	Cards	Deploy Mode	Dataset	TPOT	Quantization	Configuration
DeepSeek-R1	Atlas 800I A3	16	PD Disaggregation	3.5K+1.5K	41ms	W4A8 INT8	Optimal Configuration
DeepSeek-R1	Atlas 800I A3	8	PD Mixed	3.5K+1.5K	50.36ms	W4A8 INT8	Optimal Configuration
DeepSeek-R1	Atlas 800I A3	32	PD Disaggregation	3.5K+1.5K	50ms	W8A8 INT8	Optimal Configuration

Optimal Configuration

DeepSeek-R1 W4A8 1P1D 16P IN3K5 OUT1K5 41ms

Model: DeepSeek-R1 Hardware: Atlas 800I A3 Cards: 16 Deploy Mode: PD Disaggregation Quantization: W4A8 INT8 Dataset: 3.5K+1.5K TPOT: 41ms

Model Deployment

Command

# ============================================================
# Before running, update the following variables:
#   P_IP: prefill node IP address
#   D_IP: decode node IP address
#   ASCEND_MF_STORE_URL: prefill node IP with port
#   MODEL_PATH: path to the model weights directory
#   HCCL_SOCKET_IFNAME: network interface name for HCCL
#   GLOO_SOCKET_IFNAME: network interface name for Gloo
# ============================================================


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

P_IP=('<your prefill ip>')
D_IP=('<your decode ip>')

export ASCEND_MF_STORE_URL="tcp://<your prefill ip>:24670"

MODEL_PATH=/path/to/model-weights

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 DEEP_NORMAL_MODE_USE_INT8_QUANT=1
        export GLOO_SOCKET_IFNAME=<network-interface>
        export HCCL_BUFFSIZE=3500
        export HCCL_SOCKET_IFNAME=<network-interface>
        export TASK_QUEUE_ENABLE=2

        python3 -m sglang.launch_server \
        --model-path ${MODEL_PATH} \
        --disaggregation-mode prefill \
        --host ${P_IP[$i]} \
        --port 8000 \
        --disaggregation-bootstrap-port 8998 \
        --node-rank 0 \
        --nnodes 1 \
        --tp-size 16 \
        --mem-fraction-static 0.62 \
        --quantization modelslim \
        --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 GLOO_SOCKET_IFNAME=<network-interface>
        export HCCL_BUFFSIZE=800
        export HCCL_SOCKET_IFNAME=<network-interface>
        export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=78
        export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
        export SGLANG_ENABLE_SPEC_V2=1
        export TASK_QUEUE_ENABLE=1

        python3 -m sglang.launch_server \
        --model-path ${MODEL_PATH} \
        --disaggregation-mode decode \
        --host ${D_IP[$i]} \
        --port 8001 \
        --nnodes 1 \
        --tp-size 16 \
        --dp-size 16 \
        --mem-fraction-static 0.805 \
        --max-running-requests 416 \
        --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 \
        --watchdog-timeout 9000 \
        --context-length 8192 \
        --speculative-algorithm NEXTN \
        --speculative-num-steps 2 \
        --speculative-eagle-topk 1 \
        --speculative-num-draft-tokens 3 \
        --prefill-round-robin-balance \
        --disable-shared-experts-fusion \
        --dtype bfloat16 \
        --tokenizer-worker-num 4 \
        --load-balance-method round_robin
        NODE_RANK=$i
        break
    fi
done

Command

# ============================================================
# Before running, replace the following placeholders:
#   <your prefill ip>: prefill node IP address
#   <your decode ip>: decode node IP address
# ============================================================

python -m sglang_router.launch_router \
    --pd-disaggregation \
    --policy cache_aware \
    --prefill http://<your prefill ip>:8000 8998 \
    --decode http://<your decode ip>:8001 \
    --host 127.0.0.1 \
    --port 6688 \

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 3584 \
    --random-output-len 1536 \
    --num-prompts 1664 \
    --random-range-ratio 1 \
    --request-rate 24

DeepSeek-R1 W4A8 8P IN3K5 OUT1K5 50.36ms

Model: DeepSeek-R1 Hardware: Atlas 800I A3 Cards: 8 Deploy Mode: PD Mixed Quantization: W4A8 INT8 Dataset: 3.5K+1.5K TPOT: 50.36ms

Model Deployment

Command

# ============================================================
# Before running, update the following variables:
#   MODEL_PATH: path to the model weights directory
#   HCCL_SOCKET_IFNAME: network interface name for HCCL
#   GLOO_SOCKET_IFNAME: network interface name for Gloo
# ============================================================

MODEL_PATH=/path/to/model-weights
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 DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS=512
export DEEPEP_NORMAL_LONG_SEQ_ROUND=10
export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
export GLOO_SOCKET_IFNAME=<network-interface>
export HCCL_BUFFSIZE=1200
export HCCL_SOCKET_IFNAME=<network-interface>
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=56
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_SPEC_V2=1
export SGLANG_NPU_USE_MLAPO=1
export SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES=200
export SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE=1
export SGLANG_USE_FIA_NZ=1
export STREAMS_PER_DEVICE=32

python3 -m sglang.launch_server \
    --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 6688 \
    --tp-size 16 \
    --trust-remote-code \
    --attention-backend ascend \
    --device npu \
    --quantization modelslim \
    --watchdog-timeout 9000 \
    --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 6688 \
    --max-concurrency 224 \
    --random-input-len 3500 \
    --random-output-len 1500 \
    --num-prompts 896 \
    --random-range-ratio 1

DeepSeek-R1 W8A8 2P1D 32P IN3K5 OUT1K5 18.9ms

Model: DeepSeek-R1 Hardware: Atlas 800I A3 Cards: 32 Deploy Mode: PD Disaggregation Quantization: W8A8 INT8 Dataset: 3.5K+1.5K TPOT: 18.9ms

Model Deployment

Command

# ============================================================
# Before running, update the following variables:
#   P_IP: prefill node IP address
#   D_IP: decode node IP address
#   ASCEND_MF_STORE_URL: prefill node IP with port
#   MODEL_PATH: path to the model weights directory
#   HCCL_SOCKET_IFNAME: network interface name for HCCL
#   GLOO_SOCKET_IFNAME: network interface name for Gloo
# ============================================================


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

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

export ASCEND_MF_STORE_URL="tcp://<your prefill ip1>:24670"

MODEL_PATH=/path/to/model-weights

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 DEEP_NORMAL_MODE_USE_INT8_QUANT=1
        export GLOO_SOCKET_IFNAME=<network-interface>
        export HCCL_BUFFSIZE=1536
        export HCCL_SOCKET_IFNAME=<network-interface>
        export TASK_QUEUE_ENABLE=2

        python3 -m sglang.launch_server \
        --model-path ${MODEL_PATH} \
        --disaggregation-mode prefill \
        --host ${P_IP[$i]} \
        --port 8000 \
        --disaggregation-bootstrap-port $((8998 + $i)) \
        --node-rank 0 \
        --nnodes 1 \
        --tp-size 16 \
        --mem-fraction-static 0.81 \
        --quantization modelslim \
        --max-running-requests 4 \
        --context-length 8192 \
        --disable-radix-cache \
        --chunked-prefill-size -1 \
        --max-prefill-tokens 28680 \
        --moe-a2a-backend deepep \
        --deepep-mode normal \
        --speculative-algorithm NEXTN \
        --speculative-num-steps 1 \
        --speculative-eagle-topk 1 \
        --speculative-num-draft-tokens 2 \
        --dp-size 2 \
        --enable-dp-attention \
        --disable-shared-experts-fusion \
        --dtype bfloat16 \
        --enable-attn-tp-input-scattered
        NODE_RANK=$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 GLOO_SOCKET_IFNAME=<network-interface>
        export HCCL_BUFFSIZE=650
        export HCCL_SOCKET_IFNAME=<network-interface>
        export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=12
        export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
        export SGLANG_ENABLE_SPEC_V2=1
        export SGLANG_SCHEDULER_SKIP_ALL_GATHER=1
        export TASK_QUEUE_ENABLE=1

        python3 -m sglang.launch_server \
        --model-path ${MODEL_PATH} \
        --disaggregation-mode decode \
        --host ${D_IP[$i]} \
        --port 8001 \
        --dist-init-addr ${D_IP[0]}:5000 \
        --node-rank $i \
        --nnodes 2 \
        --tp-size 32 \
        --dp-size 16 \
        --mem-fraction-static 0.75 \
        --max-running-requests 32 \
        --quantization modelslim \
        --moe-a2a-backend deepep \
        --enable-dp-attention \
        --deepep-mode low_latency \
        --enable-dp-lm-head \
        --moe-dense-tp 1 \
        --cuda-graph-bs 2 4 6 \
        --watchdog-timeout 9000 \
        --context-length 8192 \
        --speculative-algorithm NEXTN \
        --speculative-num-steps 3 \
        --speculative-eagle-topk 1 \
        --speculative-num-draft-tokens 4 \
        --tokenizer-worker-num 4 \
        --prefill-round-robin-balance \
        --disable-shared-experts-fusion \
        --dtype bfloat16 \
        --load-balance-method round_robin
        NODE_RANK=$i
        break
    fi
done

Command

# ============================================================
# Before running, replace the following placeholders:
#   <your prefill ip1>, <your prefill ip2>: prefill node IP addresses
#   <your decode ip1>: first decode node IP address (decode may have distributed nodes)
# ============================================================

export SGLANG_DP_ROUND_ROBIN=1
python -m sglang_router.launch_router \
    --pd-disaggregation \
    --policy cache_aware \
    --prefill http://<your prefill ip1>:8000 8998 \
    --prefill http://<your prefill ip2>:8000 8999 \
    --decode http://<your decode 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 32 \
    --random-input-len 3500 \
    --random-output-len 1500 \
    --num-prompts 32 \
    --random-range-ratio 1 \
    --request-rate 16

DeepSeek-R1 W8A8 2P1D 32P IN3K5 OUT1K5 50ms

Model: DeepSeek-R1 Hardware: Atlas 800I A3 Cards: 32 Deploy Mode: PD Disaggregation Quantization: W8A8 INT8 Dataset: 3.5K+1.5K TPOT: 50ms

Model Deployment

Command

# ============================================================
# Before running, update the following variables:
#   P_IP: prefill node IP address
#   D_IP: decode node IP address
#   ASCEND_MF_STORE_URL: prefill node IP with port
#   MODEL_PATH: path to the model weights directory
#   HCCL_SOCKET_IFNAME: network interface name for HCCL
#   GLOO_SOCKET_IFNAME: network interface name for Gloo
# ============================================================


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

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

export ASCEND_MF_STORE_URL="tcp://<your prefill ip1>:24670"

MODEL_PATH=/path/to/model-weights

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 DEEP_NORMAL_MODE_USE_INT8_QUANT=1
        export GLOO_SOCKET_IFNAME=<network-interface>
        export HCCL_BUFFSIZE=800
        export HCCL_SOCKET_IFNAME=<network-interface>
        export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=131072
        export SGLANG_NPU_FUSED_MOE_MODE=2
        export SGLANG_USE_AG_AFTER_QLORA=1
        export TASK_QUEUE_ENABLE=2

        python3 -m sglang.launch_server \
        --model-path ${MODEL_PATH} \
        --disaggregation-mode prefill \
        --host ${P_IP[$i]} \
        --port 8000 \
        --disaggregation-bootstrap-port $((8998 + $i)) \
        --node-rank 0 \
        --nnodes 1 \
        --tp-size 16 \
        --mem-fraction-static 0.778 \
        --quantization modelslim \
        --max-running-requests 16 \
        --disable-radix-cache \
        --chunked-prefill-size -1 \
        --max-prefill-tokens 60000 \
        --moe-a2a-backend ascend_fuseep \
        --deepep-mode normal \
        --speculative-algorithm NEXTN \
        --speculative-num-steps 1 \
        --speculative-eagle-topk 1 \
        --speculative-num-draft-tokens 2 \
        --dp-size 4 \
        --enable-dp-attention \
        --disable-shared-experts-fusion \
        --dtype bfloat16 \
        --enable-attn-tp-input-scattered
        NODE_RANK=$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 GLOO_SOCKET_IFNAME=<network-interface>
        export HCCL_BUFFSIZE=600
        export HCCL_SOCKET_IFNAME=<network-interface>
        export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=64
        export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
        export SGLANG_ENABLE_SPEC_V2=1
        export SGLANG_LM_HEAD_TP=8
        export SGLANG_NPU_FUSED_MOE_MODE=1
        export TASK_QUEUE_ENABLE=1

        python3 -m sglang.launch_server \
        --model-path ${MODEL_PATH} \
        --disaggregation-mode decode \
        --host ${D_IP[$i]} \
        --port 8001 \
        --dist-init-addr ${D_IP[0]}:5000 \
        --node-rank $i \
        --nnodes 2 \
        --tp-size 32 \
        --dp-size 32 \
        --mem-fraction-static 0.82 \
        --max-running-requests 1024 \
        --quantization modelslim \
        --moe-a2a-backend ascend_fuseep \
        --enable-dp-attention \
        --deepep-mode low_latency \
        --moe-dense-tp 1 \
        --cuda-graph-bs 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 \
        --watchdog-timeout 9000 \
        --context-length 8192 \
        --speculative-algorithm NEXTN \
        --speculative-num-steps 1 \
        --speculative-eagle-topk 1 \
        --speculative-num-draft-tokens 2 \
        --tokenizer-worker-num 4 \
        --prefill-round-robin-balance \
        --disable-shared-experts-fusion \
        --dtype bfloat16 \
        --load-balance-method round_robin
        NODE_RANK=$i
        break
    fi
done

Command

# ============================================================
# Before running, replace the following placeholders:
#   <your prefill ip1>, <your prefill ip2>: prefill node IP addresses
#   <your decode ip1>: first decode node IP address (decode may have distributed nodes)
# ============================================================

export SGLANG_DP_ROUND_ROBIN=1
python -m sglang_router.launch_router \
    --pd-disaggregation \
    --policy cache_aware \
    --prefill http://<your prefill ip1>:8000 8998 \
    --prefill http://<your prefill ip2>:8000 8999 \
    --decode http://<your decode 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 1024 \
    --random-input-len 3584 \
    --random-output-len 1536 \
    --num-prompts 7168 \
    --random-range-ratio 1 \
    --request-rate 40

DeepSeek-R1 W8A8 2P1D 32P IN3K5 OUT1K 19.0ms

Model: DeepSeek-R1 Hardware: Atlas 800I A3 Cards: 32 Deploy Mode: PD Disaggregation Quantization: W8A8 INT8 Dataset: 3.5K+1K TPOT: 19.0ms

Model Deployment

Command

# ============================================================
# Before running, update the following variables:
#   P_IP: prefill node IP address
#   D_IP: decode node IP address
#   ASCEND_MF_STORE_URL: prefill node IP with port
#   MODEL_PATH: path to the model weights directory
#   HCCL_SOCKET_IFNAME: network interface name for HCCL
#   GLOO_SOCKET_IFNAME: network interface name for Gloo
# ============================================================


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

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

export ASCEND_MF_STORE_URL="tcp://<your prefill ip1>:24670"

MODEL_PATH=/path/to/model-weights

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 DEEP_NORMAL_MODE_USE_INT8_QUANT=1
        export GLOO_SOCKET_IFNAME=<network-interface>
        export HCCL_BUFFSIZE=1536
        export HCCL_SOCKET_IFNAME=<network-interface>
        export TASK_QUEUE_ENABLE=2

        python3 -m sglang.launch_server \
        --model-path ${MODEL_PATH} \
        --disaggregation-mode prefill \
        --host ${P_IP[$i]} \
        --port 8000 \
        --disaggregation-bootstrap-port $((8998 + $i)) \
        --node-rank 0 \
        --nnodes 1 \
        --tp-size 16 \
        --mem-fraction-static 0.81 \
        --quantization modelslim \
        --max-running-requests 4 \
        --context-length 8192 \
        --disable-radix-cache \
        --chunked-prefill-size -1 \
        --max-prefill-tokens 28680 \
        --moe-a2a-backend deepep \
        --deepep-mode normal \
        --speculative-algorithm NEXTN \
        --speculative-num-steps 1 \
        --speculative-eagle-topk 1 \
        --speculative-num-draft-tokens 2 \
        --dp-size 2 \
        --enable-dp-attention \
        --disable-shared-experts-fusion \
        --dtype bfloat16 \
        --enable-attn-tp-input-scattered
        NODE_RANK=$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 GLOO_SOCKET_IFNAME=<network-interface>
        export HCCL_BUFFSIZE=650
        export HCCL_SOCKET_IFNAME=<network-interface>
        export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=12
        export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
        export SGLANG_ENABLE_SPEC_V2=1
        export SGLANG_SCHEDULER_SKIP_ALL_GATHER=1
        export TASK_QUEUE_ENABLE=1

        python3 -m sglang.launch_server \
        --model-path ${MODEL_PATH} \
        --disaggregation-mode decode \
        --host ${D_IP[$i]} \
        --port 8001 \
        --dist-init-addr ${D_IP[0]}:5000 \
        --node-rank $i \
        --nnodes 2 \
        --tp-size 32 \
        --dp-size 16 \
        --mem-fraction-static 0.75 \
        --max-running-requests 32 \
        --quantization modelslim \
        --moe-a2a-backend deepep \
        --enable-dp-attention \
        --deepep-mode low_latency \
        --enable-dp-lm-head \
        --moe-dense-tp 1 \
        --cuda-graph-bs 2 4 6 \
        --watchdog-timeout 9000 \
        --context-length 8192 \
        --speculative-algorithm NEXTN \
        --speculative-num-steps 3 \
        --speculative-eagle-topk 1 \
        --speculative-num-draft-tokens 4 \
        --tokenizer-worker-num 4 \
        --prefill-round-robin-balance \
        --disable-shared-experts-fusion \
        --dtype bfloat16 \
        --load-balance-method round_robin
        NODE_RANK=$i
        break
    fi
done

Command

# ============================================================
# Before running, replace the following placeholders:
#   <your prefill ip1>, <your prefill ip2>: prefill node IP addresses
#   <your decode ip1>: first decode node IP address (decode may have distributed nodes)
# ============================================================

export SGLANG_DP_ROUND_ROBIN=1
python -m sglang_router.launch_router \
    --pd-disaggregation \
    --policy cache_aware \
    --prefill http://<your prefill ip1>:8000 8998 \
    --prefill http://<your prefill ip2>:8000 8999 \
    --decode http://<your decode 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 32 \
    --random-input-len 3500 \
    --random-output-len 1024 \
    --num-prompts 32 \
    --random-range-ratio 1 \
    --request-rate 16

DeepSeek-R1 W8A8 2P1D 32P IN3K9 OUT1K 19.0ms

Model: DeepSeek-R1 Hardware: Atlas 800I A3 Cards: 32 Deploy Mode: PD Disaggregation Quantization: W8A8 INT8 Dataset: 3.9K+1K TPOT: 19.0ms

Model Deployment

Command

# ============================================================
# Before running, update the following variables:
#   P_IP: prefill node IP address
#   D_IP: decode node IP address
#   ASCEND_MF_STORE_URL: prefill node IP with port
#   MODEL_PATH: path to the model weights directory
#   HCCL_SOCKET_IFNAME: network interface name for HCCL
#   GLOO_SOCKET_IFNAME: network interface name for Gloo
# ============================================================


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

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

export ASCEND_MF_STORE_URL="tcp://<your prefill ip1>:24670"

MODEL_PATH=/path/to/model-weights

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 DEEP_NORMAL_MODE_USE_INT8_QUANT=1
        export GLOO_SOCKET_IFNAME=<network-interface>
        export HCCL_BUFFSIZE=1536
        export HCCL_SOCKET_IFNAME=<network-interface>
        export TASK_QUEUE_ENABLE=2

        python3 -m sglang.launch_server \
        --model-path ${MODEL_PATH} \
        --disaggregation-mode prefill \
        --host ${P_IP[$i]} \
        --port 8000 \
        --disaggregation-bootstrap-port $((8998 + $i)) \
        --node-rank 0 \
        --nnodes 1 \
        --tp-size 16 \
        --mem-fraction-static 0.81 \
        --quantization modelslim \
        --max-running-requests 4 \
        --context-length 8192 \
        --disable-radix-cache \
        --chunked-prefill-size -1 \
        --max-prefill-tokens 28680 \
        --moe-a2a-backend deepep \
        --deepep-mode normal \
        --speculative-algorithm NEXTN \
        --speculative-num-steps 1 \
        --speculative-eagle-topk 1 \
        --speculative-num-draft-tokens 2 \
        --dp-size 2 \
        --enable-dp-attention \
        --disable-shared-experts-fusion \
        --dtype bfloat16 \
        --enable-attn-tp-input-scattered
        NODE_RANK=$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 GLOO_SOCKET_IFNAME=<network-interface>
        export HCCL_BUFFSIZE=650
        export HCCL_SOCKET_IFNAME=<network-interface>
        export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=12
        export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
        export SGLANG_ENABLE_SPEC_V2=1
        export SGLANG_SCHEDULER_SKIP_ALL_GATHER=1
        export TASK_QUEUE_ENABLE=1

        python3 -m sglang.launch_server \
        --model-path ${MODEL_PATH} \
        --disaggregation-mode decode \
        --host ${D_IP[$i]} \
        --port 8001 \
        --dist-init-addr ${D_IP[0]}:5000 \
        --node-rank $i \
        --nnodes 2 \
        --tp-size 32 \
        --dp-size 16 \
        --mem-fraction-static 0.75 \
        --max-running-requests 32 \
        --quantization modelslim \
        --moe-a2a-backend deepep \
        --enable-dp-attention \
        --deepep-mode low_latency \
        --enable-dp-lm-head \
        --moe-dense-tp 1 \
        --cuda-graph-bs 2 4 6 \
        --watchdog-timeout 9000 \
        --context-length 8192 \
        --speculative-algorithm NEXTN \
        --speculative-num-steps 3 \
        --speculative-eagle-topk 1 \
        --speculative-num-draft-tokens 4 \
        --tokenizer-worker-num 4 \
        --prefill-round-robin-balance \
        --disable-shared-experts-fusion \
        --dtype bfloat16 \
        --load-balance-method round_robin
        NODE_RANK=$i
        break
    fi
done

Command

# ============================================================
# Before running, replace the following placeholders:
#   <your prefill ip1>, <your prefill ip2>: prefill node IP addresses
#   <your decode ip1>: first decode node IP address (decode may have distributed nodes)
# ============================================================

export SGLANG_DP_ROUND_ROBIN=1
python -m sglang_router.launch_router \
    --pd-disaggregation \
    --policy cache_aware \
    --prefill http://<your prefill ip1>:8000 8998 \
    --prefill http://<your prefill ip2>:8000 8999 \
    --decode http://<your decode 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 32 \
    --random-input-len 3900 \
    --random-output-len 1024 \
    --num-prompts 32 \
    --random-range-ratio 1 \
    --request-rate 16

DeepSeek-R1 W8A8 2P1D 32P IN6K OUT1K6 20.5ms

Model: DeepSeek-R1 Hardware: Atlas 800I A3 Cards: 32 Deploy Mode: PD Disaggregation Quantization: W8A8 INT8 Dataset: 6K+1.6K TPOT: 20.5ms

Model Deployment

Command

# ============================================================
# Before running, update the following variables:
#   P_IP: prefill node IP address
#   D_IP: decode node IP address
#   ASCEND_MF_STORE_URL: prefill node IP with port
#   MODEL_PATH: path to the model weights directory
#   HCCL_SOCKET_IFNAME: network interface name for HCCL
#   GLOO_SOCKET_IFNAME: network interface name for Gloo
# ============================================================


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

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

export ASCEND_MF_STORE_URL="tcp://<your prefill ip1>:24670"

MODEL_PATH=/path/to/model-weights

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 DEEP_NORMAL_MODE_USE_INT8_QUANT=1
        export GLOO_SOCKET_IFNAME=<network-interface>
        export HCCL_BUFFSIZE=1536
        export HCCL_SOCKET_IFNAME=<network-interface>
        export TASK_QUEUE_ENABLE=2

        python3 -m sglang.launch_server \
        --model-path ${MODEL_PATH} \
        --disaggregation-mode prefill \
        --host ${P_IP[$i]} \
        --port 8000 \
        --disaggregation-bootstrap-port $((8998 + $i)) \
        --node-rank 0 \
        --nnodes 1 \
        --tp-size 16 \
        --mem-fraction-static 0.81 \
        --quantization modelslim \
        --max-running-requests 4 \
        --disable-radix-cache \
        --chunked-prefill-size -1 \
        --max-prefill-tokens 28680 \
        --moe-a2a-backend deepep \
        --deepep-mode normal \
        --speculative-algorithm NEXTN \
        --speculative-num-steps 1 \
        --speculative-eagle-topk 1 \
        --speculative-num-draft-tokens 2 \
        --dp-size 2 \
        --enable-dp-attention \
        --disable-shared-experts-fusion \
        --dtype bfloat16 \
        --enable-attn-tp-input-scattered
        NODE_RANK=$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 GLOO_SOCKET_IFNAME=<network-interface>
        export HCCL_BUFFSIZE=650
        export HCCL_SOCKET_IFNAME=<network-interface>
        export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=16
        export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
        export SGLANG_ENABLE_SPEC_V2=1
        export SGLANG_SCHEDULER_SKIP_ALL_GATHER=1
        export TASK_QUEUE_ENABLE=1

        python3 -m sglang.launch_server \
        --model-path ${MODEL_PATH} \
        --disaggregation-mode decode \
        --host ${D_IP[$i]} \
        --port 8001 \
        --dist-init-addr ${D_IP[0]}:5000 \
        --node-rank $i \
        --nnodes 2 \
        --tp-size 32 \
        --dp-size 8 \
        --mem-fraction-static 0.75 \
        --max-running-requests 32 \
        --quantization modelslim \
        --moe-a2a-backend deepep \
        --enable-dp-attention \
        --deepep-mode low_latency \
        --enable-dp-lm-head \
        --moe-dense-tp 1 \
        --cuda-graph-bs 2 4 6 \
        --watchdog-timeout 9000 \
        --speculative-algorithm NEXTN \
        --speculative-num-steps 3 \
        --speculative-eagle-topk 1 \
        --speculative-num-draft-tokens 4 \
        --tokenizer-worker-num 4 \
        --prefill-round-robin-balance \
        --disable-shared-experts-fusion \
        --dtype bfloat16 \
        --load-balance-method round_robin
        NODE_RANK=$i
        break
    fi
done

Command

# ============================================================
# Before running, replace the following placeholders:
#   <your prefill ip1>, <your prefill ip2>: prefill node IP addresses
#   <your decode ip1>: first decode node IP address (decode may have distributed nodes)
# ============================================================

export SGLANG_DP_ROUND_ROBIN=1
python -m sglang_router.launch_router \
    --pd-disaggregation \
    --policy cache_aware \
    --prefill http://<your prefill ip1>:8000 8998 \
    --prefill http://<your prefill ip2>:8000 8999 \
    --decode http://<your decode 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 32 \
    --random-input-len 6000 \
    --random-output-len 1600 \
    --num-prompts 32 \
    --random-range-ratio 1 \
    --request-rate 16

​Low Latency

​High Throughput

​Optimal Configuration

​DeepSeek-R1 W4A8 1P1D 16P IN3K5 OUT1K5 41ms

​Model Deployment

​Benchmark

​DeepSeek-R1 W4A8 8P IN3K5 OUT1K5 50.36ms

​Model Deployment

​Benchmark

​DeepSeek-R1 W8A8 2P1D 32P IN3K5 OUT1K5 18.9ms

​Model Deployment

​Benchmark

​DeepSeek-R1 W8A8 2P1D 32P IN3K5 OUT1K5 50ms

​Model Deployment

​Benchmark

​DeepSeek-R1 W8A8 2P1D 32P IN3K5 OUT1K 19.0ms

​Model Deployment

​Benchmark

​DeepSeek-R1 W8A8 2P1D 32P IN3K9 OUT1K 19.0ms

​Model Deployment

​Benchmark

​DeepSeek-R1 W8A8 2P1D 32P IN6K OUT1K6 20.5ms

​Model Deployment

​Benchmark

Low Latency

High Throughput

Optimal Configuration

DeepSeek-R1 W4A8 1P1D 16P IN3K5 OUT1K5 41ms

Model Deployment

Benchmark

DeepSeek-R1 W4A8 8P IN3K5 OUT1K5 50.36ms

Model Deployment

Benchmark

DeepSeek-R1 W8A8 2P1D 32P IN3K5 OUT1K5 18.9ms

Model Deployment

Benchmark

DeepSeek-R1 W8A8 2P1D 32P IN3K5 OUT1K5 50ms

Model Deployment

Benchmark

DeepSeek-R1 W8A8 2P1D 32P IN3K5 OUT1K 19.0ms

Model Deployment

Benchmark

DeepSeek-R1 W8A8 2P1D 32P IN3K9 OUT1K 19.0ms

Model Deployment

Benchmark

DeepSeek-R1 W8A8 2P1D 32P IN6K OUT1K6 20.5ms

Model Deployment

Benchmark