Kimi-K2.6

이 페이지는 Ascend NPU에서 Kimi-K2.6의 최적 구성과 벤치마크 결과에 집중해요. 환경 설정, 모델 가중치 다운로드, 기능 구성, 배포 지침 등은 Kimi-K2.6 모델 튜토리얼을 참고하세요.

A3 시리즈에서는 각 카드에 2개의 die가 있어서 --tp-size가 카드 수의 2배예요. 자세한 내용은 Ascend NPU 참조를 확인하세요.

출처: 문서

본문

저지연 (Low Latency)

모델	하드웨어	카드	배포 모드	데이터셋	TPOT	TTFT	양자화	구성
Kimi-K2.6	Ascend A3 Series Products	8	PD Mixed	3.5k+1.5k	20ms	-	W4A8 INT8	최적 구성

고처리량 (High Throughput)

모델	하드웨어	카드	배포 모드	데이터셋	TPOT	TTFT	양자화	구성
Kimi-K2.6	Ascend A3 Series Products	16	PD Mixed	64k+1k	100ms	-	W4A8 INT8	최적 구성
Kimi-K2.6	Ascend A3 Series Products	16	PD Disaggregation	128k+1k	100ms	-	W4A8 INT8	최적 구성
Kimi-K2.6	Ascend A3 Series Products	16	PD Disaggregation	128k+1k (90% prefix cache hit rate)	100ms	5s	W4A8 INT8	최적 구성
Kimi-K2.6	Ascend A3 Series Products	16	PD Disaggregation	64k+1.5k	100ms	-	W4A8 INT8	최적 구성
Kimi-K2.6	Ascend A3 Series Products	16	PD Disaggregation	64k+1.5k (90% prefix cache hit rate)	100ms	3s	W4A8 INT8	최적 구성
Kimi-K2.6	Ascend A3 Series Products	8	PD Mixed	1024x1024 (30)+1024	50ms	-	W4A8 INT8	최적 구성
Kimi-K2.6	Ascend A3 Series Products	8	PD Mixed	1080p_30+256	50ms	-	W4A8 INT8	최적 구성
Kimi-K2.6	Ascend A3 Series Products	8	PD Mixed	3.5k+1.5k	50ms	-	W4A8 INT8	최적 구성

최적 구성 (Optimal Configuration)

Kimi-K2.6 W4A8 16P IN64K OUT1K 100ms

모델: Kimi-K2.6

하드웨어: Ascend A3 Series Products

카드: 16

배포 모드: PD Mixed

양자화: W4A8 INT8

데이터셋: 64k+1k

TPOT: 100ms

모델 배포 (Model Deployment)

# ============================================================
# Before running, update the following variables:
#   MODEL_PATH: path to the model weights directory
#   DRAFT_MODEL_PATH: path to the draft model weights directory
#   NODE_IPS: IP addresses of each node in the cluster
#   HCCL_SOCKET_IFNAME: network interface name for HCCL
#   GLOO_SOCKET_IFNAME: network interface name for Gloo
# ============================================================

MODEL_PATH=/path/to/model-weights
DRAFT_MODEL_PATH=/path/to/draft-model-weights

NODE_IPS=('<your node1 ip>' '<your node2 ip>')

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_HCCL_BUFFSIZE=4400
export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
export GLOO_SOCKET_IFNAME=<network-interface>
export HCCL_SOCKET_IFNAME=<network-interface>
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=64
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_SET_CPU_AFFINITY=1
export STREAMS_PER_DEVICE=32

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 "${!NODE_IPS[@]}";
do
    if [[ "$LOCAL_HOST1" == "${NODE_IPS[$i]}" || "$LOCAL_HOST2" == "${NODE_IPS[$i]}" ]];
    then
        echo "${NODE_IPS[$i]}"
        python3 -m sglang.launch_server \
        --model-path $MODEL_PATH \
        --host ${NODE_IPS[$i]} --port 6688 \
        --nnodes 2 \
        --dist-init-addr ${NODE_IPS[0]}:5000 \
        --node-rank $i \
        --trust-remote-code \
        --attention-backend ascend \
        --device npu \
        --quantization modelslim \
        --dtype bfloat16 \
        --tp-size 32 \
        --mem-fraction-static 0.662 \
        --max-running-requests 32 \
        --chunked-prefill-size 262144 \
        --context-length 75000 \
        --enable-multimodal \
        --mm-attention-backend ascend_attn \
        --sampling-backend ascend \
        --enable-dp-attention \
        --dp-size 32 \
        --moe-a2a-backend deepep \
        --deepep-mode auto \
        --cuda-graph-bs-decode 1 \
        --disable-radix-cache \
        --speculative-algorithm EAGLE3 \
        --speculative-draft-model-path $DRAFT_MODEL_PATH \
        --speculative-num-steps 3 \
        --speculative-eagle-topk 1 \
        --speculative-num-draft-tokens 4 \
        --speculative-draft-model-quantization unquant \
        --reasoning-parser kimi_k2 \
        --tool-call-parser kimi_k2
        break
    fi
done

벤치마크 (Benchmark)

RANDOM 데이터셋을 기반으로 테스트했어요.

python -m sglang.bench_serving \
    --dataset-name random \
    --backend sglang \
    --host 127.0.0.1 \
    --port 6688 \
    --max-concurrency 32 \
    --num-prompts 32 \
    --random-input-len 64000 \
    --random-output-len 1000 \
    --random-range-ratio 1 \
    --seed 1

Kimi-K2.6 W4A8 1P1D 16P IN128K OUT1K 100ms

모델: Kimi-K2.6

하드웨어: Ascend A3 Series Products

카드: 16

배포 모드: PD Disaggregation

양자화: W4A8 INT8

데이터셋: 128k+1k

TPOT: 100ms

모델 배포 (Model Deployment)

# ============================================================
# 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
#   DRAFT_MODEL_PATH: path to the draft 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 DEEP_NORMAL_MODE_USE_INT8_QUANT=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=60
export SGLANG_SET_CPU_AFFINITY=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
DRAFT_MODEL_PATH=/path/to/draft-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 GLOO_SOCKET_IFNAME=<network-interface>
        export HCCL_BUFFSIZE=8
        export HCCL_SOCKET_IFNAME=<network-interface>
        export SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK=0
        export SGLANG_ZBAL_BOOTSTRAP_URL=tcp://127.0.0.1:24699
        export SGLANG_ZBAL_LOCAL_MEM_SIZE=61184
        export ZBAL_ENABLE_GRAPH=1
        export ZBAL_NPU_ALLOC_CONF=use_vmm_for_static_memory:True

        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 \
        --quantization modelslim \
        --dtype bfloat16 \
        --disaggregation-transfer-backend ascend \
        --nnodes 1 \
        --trust-remote-code \
        --attention-backend ascend \
        --device npu \
        --tp-size 16 \
        --disable-radix-cache \
        --mem-fraction-static 0.78 \
        --max-running-requests 2 \
        --moe-a2a-backend deepep \
        --deepep-mode auto \
        --chunked-prefill-size 16384 \
        --prefill-max-requests 2 \
        --max-prefill-tokens 65536 \
        --enable-multimodal \
        --mm-attention-backend ascend_attn \
        --sampling-backend ascend
        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 DEEPEP_HCCL_BUFFSIZE=1200
        export GLOO_SOCKET_IFNAME=<network-interface>
        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_NPU_USE_MLAPO=1
        export SGLANG_NPU_USE_MULTI_STREAM=1

        python3 -m sglang.launch_server \
        --model-path ${MODEL_PATH} \
        --disaggregation-mode decode \
        --host ${D_IP[$i]} \
        --port 8001 \
        --quantization modelslim \
        --dtype bfloat16 \
        --disaggregation-transfer-backend ascend \
        --nnodes 1 \
        --trust-remote-code \
        --attention-backend ascend \
        --device npu \
        --tp-size 16 \
        --mem-fraction-static 0.82 \
        --max-running-requests 2 \
        --enable-dp-attention \
        --dp-size 1 \
        --enable-dp-lm-head \
        --disable-radix-cache \
        --enable-multimodal \
        --mm-attention-backend ascend_attn \
        --sampling-backend ascend \
        --moe-a2a-backend deepep \
        --deepep-mode auto \
        --cuda-graph-bs-decode 1 2 4 6 8 16 \
        --speculative-algorithm EAGLE3 \
        --speculative-draft-model-path $DRAFT_MODEL_PATH \
        --speculative-num-steps 3 \
        --speculative-eagle-topk 1 \
        --speculative-num-draft-tokens 4 \
        --speculative-draft-model-quantization unquant \
        --reasoning-parser kimi_k2 \
        --tool-call-parser kimi_k2
        break
    fi
done

# ============================================================
# 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 \
    --prefill http://<your prefill ip>:8000 8998 \
    --decode http://<your decode ip>:8001 \
    --host 127.0.0.1 \
    --port 6688 \
    --policy cache_aware

벤치마크 (Benchmark)

RANDOM 데이터셋을 기반으로 테스트했어요.

python -m sglang.bench_serving \
    --dataset-name random \
    --backend sglang \
    --host 127.0.0.1 \
    --port 6688 \
    --max-concurrency 1 \
    --num-prompts 1 \
    --request-rate inf \
    --random-input-len 128000 \
    --random-output-len 1000 \
    --random-range-ratio 1 \
    --seed 1

Kimi-K2.6 W4A8 1P1D 16P IN128K OUT1K PREFIX90 100ms

모델: Kimi-K2.6

하드웨어: Ascend A3 Series Products

카드: 16

배포 모드: PD Disaggregation

양자화: W4A8 INT8

데이터셋: 128k+1k (90% prefix cache hit rate)

TPOT: 100ms

TTFT: 5s

모델 배포 (Model Deployment)

# ============================================================
# 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
#   DRAFT_MODEL_PATH: path to the draft 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 DEEP_NORMAL_MODE_USE_INT8_QUANT=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=60
export SGLANG_SET_CPU_AFFINITY=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
DRAFT_MODEL_PATH=/path/to/draft-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 GLOO_SOCKET_IFNAME=<network-interface>
        export HCCL_BUFFSIZE=8
        export HCCL_SOCKET_IFNAME=<network-interface>
        export SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK=0
        export SGLANG_ZBAL_BOOTSTRAP_URL=tcp://127.0.0.1:24699
        export SGLANG_ZBAL_LOCAL_MEM_SIZE=61184
        export ZBAL_ENABLE_GRAPH=1
        export ZBAL_NPU_ALLOC_CONF=use_vmm_for_static_memory:True

        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 \
        --quantization modelslim \
        --dtype bfloat16 \
        --disaggregation-transfer-backend ascend \
        --nnodes 1 \
        --trust-remote-code \
        --attention-backend ascend \
        --device npu \
        --tp-size 16 \
        --mem-fraction-static 0.78 \
        --max-running-requests 2 \
        --moe-a2a-backend deepep \
        --deepep-mode auto \
        --chunked-prefill-size 16384 \
        --prefill-max-requests 2 \
        --max-prefill-tokens 65536 \
        --enable-multimodal \
        --mm-attention-backend ascend_attn \
        --sampling-backend ascend \
        --reasoning-parser kimi_k2 \
        --tool-call-parser kimi_k2
        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 DEEPEP_HCCL_BUFFSIZE=1200
        export GLOO_SOCKET_IFNAME=<network-interface>
        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_NPU_USE_MLAPO=1
        export SGLANG_NPU_USE_MULTI_STREAM=1

        python3 -m sglang.launch_server \
        --model-path ${MODEL_PATH} \
        --disaggregation-mode decode \
        --host ${D_IP[$i]} \
        --port 8001 \
        --quantization modelslim \
        --dtype bfloat16 \
        --disaggregation-transfer-backend ascend \
        --nnodes 1 \
        --trust-remote-code \
        --attention-backend ascend \
        --device npu \
        --tp-size 16 \
        --mem-fraction-static 0.82 \
        --max-running-requests 2 \
        --enable-dp-attention \
        --dp-size 1 \
        --enable-dp-lm-head \
        --disable-radix-cache \
        --enable-multimodal \
        --mm-attention-backend ascend_attn \
        --sampling-backend ascend \
        --moe-a2a-backend deepep \
        --deepep-mode auto \
        --cuda-graph-bs-decode 1 2 4 6 8 16 \
        --speculative-algorithm EAGLE3 \
        --speculative-draft-model-path $DRAFT_MODEL_PATH \
        --speculative-num-steps 3 \
        --speculative-eagle-topk 1 \
        --speculative-num-draft-tokens 4 \
        --speculative-draft-model-quantization unquant \
        --reasoning-parser kimi_k2 \
        --tool-call-parser kimi_k2
        break
    fi
done

# ============================================================
# 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 \
    --prefill http://<your prefill ip>:8000 8998 \
    --decode http://<your decode ip>:8001 \
    --host 127.0.0.1 \
    --port 6688 \
    --policy cache_aware

벤치마크 (Benchmark)

90% 캐시 히트(repeat_rate = 0.9)를 갖는 generated-shared-prefix 데이터셋을 기반으로 테스트했어요: --gsp-system-prompt-len 115200 = round(128000 * 0.9)가 공유 프리픽스 부분이에요. --gsp-question-len 12800 = round(128000 * (1 - 0.9))가 요청별 고유 접미사예요. --gsp-num-groups 1은 최대 캐시 재사용을 위해 모든 요청을 하나의 프리픽스 그룹으로 유지해요.

python -m sglang.bench_serving \
    --dataset-name generated-shared-prefix \
    --backend sglang \
    --host 127.0.0.1 \
    --port 6688 \
    --gsp-num-groups 1 \
    --gsp-prompts-per-group 4 \
    --gsp-system-prompt-len 115200 \
    --gsp-question-len 12800 \
    --gsp-output-len 1000 \
    --max-concurrency 1 \
    --num-prompts 4 \
    --request-rate inf

Kimi-K2.6 W4A8 1P1D 16P IN64K OUT1K5 100ms

모델: Kimi-K2.6

하드웨어: Ascend A3 Series Products

카드: 16

배포 모드: PD Disaggregation

양자화: W4A8 INT8

데이터셋: 64k+1.5k

TPOT: 100ms

모델 배포 (Model Deployment)

# ============================================================
# 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
#   DRAFT_MODEL_PATH: path to the draft 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 DEEP_NORMAL_MODE_USE_INT8_QUANT=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=60
export SGLANG_SET_CPU_AFFINITY=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
DRAFT_MODEL_PATH=/path/to/draft-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 DEEPEP_HCCL_BUFFSIZE=1800
        export GLOO_SOCKET_IFNAME=<network-interface>
        export HCCL_SOCKET_IFNAME=<network-interface>

        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 \
        --quantization modelslim \
        --dtype bfloat16 \
        --disaggregation-transfer-backend ascend \
        --nnodes 1 \
        --trust-remote-code \
        --attention-backend ascend \
        --device npu \
        --tp-size 16 \
        --disable-radix-cache \
        --disable-cuda-graph \
        --mem-fraction-static 0.78 \
        --max-running-requests 1 \
        --moe-a2a-backend deepep \
        --deepep-mode auto \
        --chunked-prefill-size 16384 \
        --prefill-max-requests 1 \
        --max-prefill-tokens 65536 \
        --enable-multimodal \
        --mm-attention-backend ascend_attn \
        --sampling-backend ascend \
        --reasoning-parser kimi_k2 \
        --tool-call-parser kimi_k2
        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 DEEPEP_HCCL_BUFFSIZE=1200
        export GLOO_SOCKET_IFNAME=<network-interface>
        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_NPU_USE_MLAPO=1
        export SGLANG_NPU_USE_MULTI_STREAM=1

        python3 -m sglang.launch_server \
        --model-path ${MODEL_PATH} \
        --disaggregation-mode decode \
        --host ${D_IP[$i]} \
        --port 8001 \
        --quantization modelslim \
        --dtype bfloat16 \
        --disaggregation-transfer-backend ascend \
        --nnodes 1 \
        --trust-remote-code \
        --attention-backend ascend \
        --device npu \
        --tp-size 16 \
        --mem-fraction-static 0.82 \
        --max-running-requests 16 \
        --enable-dp-attention \
        --dp-size 1 \
        --enable-dp-lm-head \
        --disable-radix-cache \
        --enable-multimodal \
        --mm-attention-backend ascend_attn \
        --sampling-backend ascend \
        --moe-a2a-backend deepep \
        --deepep-mode auto \
        --cuda-graph-bs-decode 16 \
        --reasoning-parser kimi_k2 \
        --tool-call-parser kimi_k2 \
        --speculative-algorithm EAGLE3 \
        --speculative-draft-model-path $DRAFT_MODEL_PATH \
        --speculative-num-steps 4 \
        --speculative-eagle-topk 1 \
        --speculative-num-draft-tokens 5 \
        --speculative-draft-model-quantization unquant
        break
    fi
done

# ============================================================
# 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 \
    --prefill http://<your prefill ip>:8000 8998 \
    --decode http://<your decode ip>:8001 \
    --host 127.0.0.1 \
    --port 6688 \
    --policy cache_aware

벤치마크 (Benchmark)

RANDOM 데이터셋을 기반으로 테스트했어요.

python -m sglang.bench_serving \
    --dataset-name random \
    --backend sglang \
    --host 127.0.0.1 \
    --port 6688 \
    --max-concurrency 1 \
    --num-prompts 1 \
    --request-rate inf \
    --random-input-len 64000 \
    --random-output-len 1500 \
    --random-range-ratio 1 \
    --seed 1

Kimi-K2.6 W4A8 1P1D 16P IN64K OUT1K5 PREFIX90 100ms

모델: Kimi-K2.6

하드웨어: Ascend A3 Series Products

카드: 16

배포 모드: PD Disaggregation

양자화: W4A8 INT8

데이터셋: 64k+1.5k (90% prefix cache hit rate)

TPOT: 100ms

TTFT: 3s

모델 배포 (Model Deployment)

# ============================================================
# 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
#   DRAFT_MODEL_PATH: path to the draft 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 DEEP_NORMAL_MODE_USE_INT8_QUANT=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=60
export SGLANG_SET_CPU_AFFINITY=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
DRAFT_MODEL_PATH=/path/to/draft-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 DEEPEP_HCCL_BUFFSIZE=1800
        export GLOO_SOCKET_IFNAME=<network-interface>
        export HCCL_SOCKET_IFNAME=<network-interface>

        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 \
        --quantization modelslim \
        --dtype bfloat16 \
        --disaggregation-transfer-backend ascend \
        --nnodes 1 \
        --trust-remote-code \
        --attention-backend ascend \
        --device npu \
        --tp-size 16 \
        --mem-fraction-static 0.78 \
        --max-running-requests 2 \
        --moe-a2a-backend deepep \
        --deepep-mode auto \
        --chunked-prefill-size 16384 \
        --prefill-max-requests 2 \
        --max-prefill-tokens 65536 \
        --enable-multimodal \
        --mm-attention-backend ascend_attn \
        --sampling-backend ascend \
        --reasoning-parser kimi_k2 \
        --tool-call-parser kimi_k2
        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 DEEPEP_HCCL_BUFFSIZE=1200
        export GLOO_SOCKET_IFNAME=<network-interface>
        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_NPU_USE_MLAPO=1
        export SGLANG_NPU_USE_MULTI_STREAM=1

        python3 -m sglang.launch_server \
        --model-path ${MODEL_PATH} \
        --disaggregation-mode decode \
        --host ${D_IP[$i]} \
        --port 8001 \
        --quantization modelslim \
        --dtype bfloat16 \
        --disaggregation-transfer-backend ascend \
        --nnodes 1 \
        --trust-remote-code \
        --attention-backend ascend \
        --device npu \
        --tp-size 16 \
        --mem-fraction-static 0.82 \
        --max-running-requests 2 \
        --enable-dp-attention \
        --dp-size 2 \
        --enable-dp-lm-head \
        --disable-radix-cache \
        --enable-multimodal \
        --mm-attention-backend ascend_attn \
        --sampling-backend ascend \
        --moe-a2a-backend deepep \
        --deepep-mode auto \
        --cuda-graph-bs-decode 1 2 4 6 8 \
        --reasoning-parser kimi_k2 \
        --tool-call-parser kimi_k2 \
        --speculative-algorithm EAGLE3 \
        --speculative-draft-model-path $DRAFT_MODEL_PATH \
        --speculative-num-steps 4 \
        --speculative-eagle-topk 1 \
        --speculative-num-draft-tokens 5 \
        --speculative-draft-model-quantization unquant
        break
    fi
done

# ============================================================
# 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 \
    --prefill http://<your prefill ip>:8000 8998 \
    --decode http://<your decode ip>:8001 \
    --host 127.0.0.1 \
    --port 6688 \
    --policy cache_aware

벤치마크 (Benchmark)

90% 캐시 히트(repeat_rate = 0.9)를 갖는 generated-shared-prefix 데이터셋을 기반으로 테스트했어요: --gsp-system-prompt-len 57600 = round(64000 * 0.9)가 공유 프리픽스 부분이에요. --gsp-question-len 6400 = round(64000 * (1 - 0.9))가 요청별 고유 접미사예요. --gsp-num-groups 1은 최대 캐시 재사용을 위해 모든 요청을 하나의 프리픽스 그룹으로 유지해요.

python -m sglang.bench_serving \
    --dataset-name generated-shared-prefix \
    --backend sglang \
    --host 127.0.0.1 \
    --port 6688 \
    --gsp-num-groups 1 \
    --gsp-prompts-per-group 8 \
    --gsp-system-prompt-len 57600 \
    --gsp-question-len 6400 \
    --gsp-output-len 1500 \
    --max-concurrency 2 \
    --num-prompts 8 \
    --request-rate inf

Kimi-K2.6 W4A8 8P IN1024X1024 30 OUT1024 50ms

모델: Kimi-K2.6

하드웨어: Ascend A3 Series Products

카드: 8

배포 모드: PD Mixed

양자화: W4A8 INT8

데이터셋: 1024x1024 (30)+1024

형식: 해상도 (입력 토큰) + 출력 토큰

TPOT: 50ms

모델 배포 (Model Deployment)

# ============================================================
# Before running, update the following variables:
#   MODEL_PATH: path to the model weights directory
#   DRAFT_MODEL_PATH: path to the draft model weights directory
# ============================================================

MODEL_PATH=/path/to/model-weights
DRAFT_MODEL_PATH=/path/to/draft-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_HCCL_BUFFSIZE=1280
export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
export HCCL_OP_EXPANSION_MODE=AIV
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=112
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_NPU_USE_MULTI_STREAM=1
export SGLANG_SET_CPU_AFFINITY=1
export STREAMS_PER_DEVICE=32

python3 -m sglang.launch_server \
    --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 6688 \
    --quantization modelslim \
    --dtype bfloat16 \
    --model-loader-extra-config '{"enable_multithread_load": true}' \
    --trust-remote-code \
    --device npu \
    --attention-backend ascend \
    --tp-size 16 \
    --mem-fraction-static 0.865 \
    --max-running-requests 176 \
    --chunked-prefill-size 32768 \
    --context-length 8192 \
    --max-prefill-tokens 16384 \
    --enable-multimodal \
    --mm-attention-backend ascend_attn \
    --sampling-backend ascend \
    --enable-dp-attention \
    --dp-size 16 \
    --moe-a2a-backend deepep \
    --deepep-mode auto \
    --cuda-graph-bs-decode 1 2 4 8 9 10 11 \
    --disable-radix-cache \
    --speculative-algorithm EAGLE3 \
    --speculative-draft-model-path $DRAFT_MODEL_PATH \
    --speculative-num-steps 2 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 3 \
    --speculative-draft-model-quantization unquant \
    --prefill-delayer-max-delay-passes 200 \
    --enable-prefill-delayer \
    --reasoning-parser kimi_k2 \
    --tool-call-parser kimi_k2

벤치마크 (Benchmark)

1024x1024 해상도의 IMAGE 데이터셋을 기반으로 테스트했어요.

python -m sglang.bench_serving \
    --dataset-name image \
    --backend sglang-oai-chat \
    --host 127.0.0.1 \
    --port 6688 \
    --image-resolution 1024x1024 \
    --image-count 1 \
    --max-concurrency 160 \
    --num-prompts 640 \
    --request-rate inf \
    --random-input-len 30 \
    --random-output-len 1024 \
    --random-range-ratio 1 \
    --warmup-requests 16 \
    --seed 1

Kimi-K2.6 W4A8 8P IN1080P 30 OUT256 50ms

모델: Kimi-K2.6

하드웨어: Ascend A3 Series Products

카드: 8

배포 모드: PD Mixed

양자화: W4A8 INT8

데이터셋: 1080p_30+256

TPOT: 50ms

모델 배포 (Model Deployment)

# ============================================================
# Before running, update the following variables:
#   MODEL_PATH: path to the model weights directory
#   DRAFT_MODEL_PATH: path to the draft model weights directory
# ============================================================

MODEL_PATH=/path/to/model-weights
DRAFT_MODEL_PATH=/path/to/draft-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_HCCL_BUFFSIZE=1024
export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
export HCCL_OP_EXPANSION_MODE=AIV
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=32
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_NPU_USE_MULTI_STREAM=1
export SGLANG_SET_CPU_AFFINITY=1
export STREAMS_PER_DEVICE=32

python3 -m sglang.launch_server \
    --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 6688 \
    --quantization modelslim \
    --dtype bfloat16 \
    --model-loader-extra-config '{"enable_multithread_load": true}' \
    --trust-remote-code \
    --device npu \
    --attention-backend ascend \
    --tp-size 16 \
    --mem-fraction-static 0.852 \
    --max-running-requests 64 \
    --chunked-prefill-size 16384 \
    --context-length 8192 \
    --max-prefill-tokens 16384 \
    --enable-multimodal \
    --mm-attention-backend ascend_attn \
    --sampling-backend ascend \
    --enable-dp-attention \
    --dp-size 16 \
    --moe-a2a-backend deepep \
    --deepep-mode auto \
    --cuda-graph-bs-decode 1 2 3 4 \
    --disable-radix-cache \
    --speculative-algorithm EAGLE3 \
    --speculative-draft-model-path $DRAFT_MODEL_PATH \
    --speculative-num-steps 2 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 3 \
    --speculative-draft-model-quantization unquant \
    --prefill-delayer-max-delay-passes 200 \
    --enable-prefill-delayer \
    --reasoning-parser kimi_k2 \
    --tool-call-parser kimi_k2

벤치마크 (Benchmark)

1920x1080 해상도의 IMAGE 데이터셋을 기반으로 테스트했어요.

python -m sglang.bench_serving \
    --dataset-name image \
    --backend sglang-oai-chat \
    --host 127.0.0.1 \
    --port 6688 \
    --image-resolution 1920x1080 \
    --image-count 1 \
    --max-concurrency 48 \
    --num-prompts 196 \
    --request-rate inf \
    --random-input-len 30 \
    --random-output-len 256 \
    --random-range-ratio 1 \
    --warmup-requests 16 \
    --seed 1

Kimi-K2.6 W4A8 8P IN3K5 OUT1K5 20ms

모델: Kimi-K2.6

하드웨어: Ascend A3 Series Products

카드: 8

배포 모드: PD Mixed

양자화: W4A8 INT8

데이터셋: 3.5k+1.5k

TPOT: 20ms

모델 배포 (Model Deployment)

# ============================================================
# Before running, update the following variables:
#   MODEL_PATH: path to the model weights directory
#   DRAFT_MODEL_PATH: path to the draft 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
DRAFT_MODEL_PATH=/path/to/draft-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_HCCL_BUFFSIZE=1200
export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
export GLOO_SOCKET_IFNAME=<network-interface>
export HCCL_OP_EXPANSION_MODE=AIV
export HCCL_SOCKET_IFNAME=<network-interface>
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=96
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_NPU_USE_MLAPO=1
export SGLANG_NPU_USE_MULTI_STREAM=1
export SGLANG_SET_CPU_AFFINITY=1
export STREAMS_PER_DEVICE=32

python3 -m sglang.launch_server \
    --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 6688 \
    --trust-remote-code \
    --attention-backend ascend \
    --device npu \
    --quantization modelslim \
    --dtype bfloat16 \
    --tp-size 16 \
    --mem-fraction-static 0.865 \
    --max-running-requests 80 \
    --chunked-prefill-size 32768 \
    --context-length 6144 \
    --max-prefill-tokens 65536 \
    --enable-multimodal \
    --mm-attention-backend ascend_attn \
    --sampling-backend ascend \
    --enable-dp-attention \
    --dp-size 16 \
    --moe-a2a-backend deepep \
    --deepep-mode auto \
    --cuda-graph-bs-decode 1 2 3 4 5 \
    --disable-radix-cache \
    --model-loader-extra-config '{"enable_multithread_load": true}' \
    --speculative-algorithm EAGLE3 \
    --speculative-draft-model-path $DRAFT_MODEL_PATH \
    --speculative-num-steps 4 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 5 \
    --speculative-draft-model-quantization unquant \
    --prefill-delayer-max-delay-passes 200 \
    --enable-prefill-delayer \
    --reasoning-parser kimi_k2 \
    --tool-call-parser kimi_k2

벤치마크 (Benchmark)

RANDOM 데이터셋을 기반으로 테스트했어요.

python -m sglang.bench_serving \
    --dataset-name random \
    --backend sglang \
    --host 127.0.0.1 \
    --port 6688 \
    --max-attempts 5 \
    --max-concurrency 64 \
    --num-prompts 256 \
    --random-input-len 3500 \
    --random-output-len 1500 \
    --random-range-ratio 1 \
    --warmup-requests 0 \
    --seed 1

Kimi-K2.6 W4A8 8P IN3K5 OUT1K5 50ms

모델: Kimi-K2.6

하드웨어: Ascend A3 Series Products

카드: 8

배포 모드: PD Mixed

양자화: W4A8 INT8

데이터셋: 3.5k+1.5k

TPOT: 50ms

모델 배포 (Model Deployment)

# ============================================================
# Before running, update the following variables:
#   MODEL_PATH: path to the model weights directory
#   DRAFT_MODEL_PATH: path to the draft 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
DRAFT_MODEL_PATH=/path/to/draft-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_HCCL_BUFFSIZE=1200
export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
export GLOO_SOCKET_IFNAME=<network-interface>
export HCCL_BUFFSIZE=200
export HCCL_OP_EXPANSION_MODE=AIV
export HCCL_SOCKET_IFNAME=<network-interface>
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=96
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_SET_CPU_AFFINITY=1
export STREAMS_PER_DEVICE=32

python3 -m sglang.launch_server \
    --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 6688 \
    --trust-remote-code \
    --attention-backend ascend \
    --device npu \
    --quantization modelslim \
    --dtype bfloat16 \
    --tp-size 16 \
    --mem-fraction-static 0.895 \
    --max-running-requests 208 \
    --chunked-prefill-size 32768 \
    --context-length 6144 \
    --max-prefill-tokens 16384 \
    --enable-multimodal \
    --mm-attention-backend ascend_attn \
    --sampling-backend ascend \
    --enable-dp-attention \
    --dp-size 16 \
    --moe-a2a-backend deepep \
    --deepep-mode auto \
    --cuda-graph-bs-decode 1 2 4 8 11 12 13 \
    --disable-radix-cache \
    --model-loader-extra-config '{"enable_multithread_load": true}' \
    --speculative-algorithm EAGLE3 \
    --speculative-draft-model-path $DRAFT_MODEL_PATH \
    --speculative-num-steps 4 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 5 \
    --speculative-draft-model-quantization unquant \
    --prefill-delayer-max-delay-passes 50 \
    --enable-prefill-delayer \
    --reasoning-parser kimi_k2 \
    --tool-call-parser kimi_k2

벤치마크 (Benchmark)

RANDOM 데이터셋을 기반으로 테스트했어요.

python -m sglang.bench_serving \
    --dataset-name random \
    --backend sglang \
    --host 127.0.0.1 \
    --port 6688 \
    --max-concurrency 192 \
    --num-prompts 768 \
    --random-input-len 3500 \
    --random-output-len 1500 \
    --random-range-ratio 1 \
    --warmup-requests 0 \
    --seed 1

더 알아보기 (Learn more)

SGLang Kimi-K2.6 최적 구성 공식 문서