Qwen3.6-35B-A3B

Qwen3.6-35B-A3B

이 페이지는 Ascend NPU에서 Qwen3.6-35B-A3B의 최적 구성과 벤치마크 결과에 집중해요. 환경 설정, 모델 가중치 다운로드, 기능 구성, 배포 지침 등은 Qwen3.6-35B-A3B 모델 튜토리얼을 참고하세요.

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

출처: 문서

본문

저지연 (Low Latency)

모델 (Model) 하드웨어 (Hardware) 카드 (Cards) 배포 모드 (Deploy Mode) 데이터셋 (Dataset) TPOT 양자화 (Quantization) 구성 (Configuration)
Qwen3.6-35B-A3B Ascend A3 Series Products 1 PD Mixed 254k+1k 16.1ms BF16 최적 구성

고처리량 (High Throughput)

모델 (Model) 하드웨어 (Hardware) 카드 (Cards) 배포 모드 (Deploy Mode) 데이터셋 (Dataset) TPOT 양자화 (Quantization) 구성 (Configuration)
Qwen3.6-35B-A3B Ascend A3 Series Products 1 PD Mixed 1024x1024 (30)+1024 50ms BF16 최적 구성
Qwen3.6-35B-A3B Ascend A3 Series Products 1 PD Mixed 1080p\_30+256 50ms BF16 최적 구성
Qwen3.6-35B-A3B Ascend A3 Series Products 1 PD Mixed 128k+1k 50ms BF16 최적 구성
Qwen3.6-35B-A3B Ascend A3 Series Products 1 PD Mixed 128k+1k (90% prefix cache hit rate) 50ms BF16 최적 구성
Qwen3.6-35B-A3B Ascend A3 Series Products 1 PD Mixed 3.5k+1.5k 50ms BF16 최적 구성
Qwen3.6-35B-A3B Ascend A3 Series Products 1 PD Mixed 64k+1k 50ms BF16 최적 구성
Qwen3.6-35B-A3B Ascend A3 Series Products 1 PD Mixed 64k+1k (90% prefix cache hit rate) 50ms BF16 최적 구성
Qwen3.6-35B-A3B Ascend A3 Series Products 2 PD Mixed 984k+1k 40.91ms BF16 최적 구성

최적 구성 (Optimal Configuration)

Qwen3.6-35B-A3B 1P IN1024X1024 30 OUT1024 50ms

모델 (Model): Qwen3.6-35B-A3B

하드웨어 (Hardware): Ascend A3 Series Products

카드 (Cards): 1

배포 모드 (Deploy Mode): PD Mixed

양자화 (Quantization): BF16

데이터셋 (Dataset): 1024x1024 (30)+1024

형식 (Format): 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
#   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 ASCEND_USE_FIA=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_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_SET_CPU_AFFINITY=1
export SGLANG_VIT_ENABLE_CUDA_GRAPH=1
export STREAMS_PER_DEVICE=32

python3 -m sglang.launch_server \
    --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 6688 \
    --tp-size 2 \
    --nnodes 1 \
    --attention-backend ascend \
    --device npu \
    --chunked-prefill-size -1 \
    --max-prefill-tokens 9999999 \
    --max-total-tokens 365000 \
    --prefill-max-requests 30 \
    --disable-radix-cache \
    --trust-remote-code \
    --max-running-requests 120 \
    --max-mamba-cache-size 120 \
    --mem-fraction-static 0.85 \
    --cuda-graph-bs-decode 4 16 32 48 64 110 165 \
    --enable-multimodal \
    --mm-attention-backend ascend_attn \
    --dtype bfloat16 \
    --mamba-ssm-dtype bfloat16 \
    --speculative-algorithm NEXTN \
    --speculative-num-steps 3 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 4 \
    --reasoning-parser qwen3 \
    --tool-call-parser qwen3_coder

벤치마크 (Benchmark)

IMAGE 데이터셋을 1024x1024 해상도로 테스트했어요.

python -m sglang.bench_serving \
    --dataset-name image \
    --backend sglang-oai-chat \
    --host 127.0.0.1 \
    --port 6688 \
    --max-concurrency 120 \
    --num-prompts 120 \
    --warmup-requests 120 \
    --random-input-len 30 \
    --random-output-len 1024 \
    --random-range-ratio 1 \
    --image-resolution 1024x1024 \
    --image-count 1 \
    --seed 1 \
    --request-rate inf

Qwen3.6-35B-A3B 1P IN1080P 30 OUT256 50ms

모델 (Model): Qwen3.6-35B-A3B

하드웨어 (Hardware): Ascend A3 Series Products

카드 (Cards): 1

배포 모드 (Deploy Mode): PD Mixed

양자화 (Quantization): BF16

데이터셋 (Dataset): 1080p\_30+256

TPOT: 50ms

모델 배포 (Model Deployment)

# ============================================================
# 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 ASCEND_USE_FIA=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_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_SET_CPU_AFFINITY=1
export SGLANG_VIT_ENABLE_CUDA_GRAPH=1
export STREAMS_PER_DEVICE=32

python3 -m sglang.launch_server \
    --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 6688 \
    --tp-size 2 \
    --nnodes 1 \
    --attention-backend ascend \
    --device npu \
    --chunked-prefill-size -1 \
    --max-prefill-tokens 150000 \
    --max-total-tokens 200000 \
    --disable-radix-cache \
    --trust-remote-code \
    --max-running-requests 42 \
    --max-mamba-cache-size 42 \
    --mem-fraction-static 0.75 \
    --cuda-graph-bs-decode 4 8 16 24 48 64 80 \
    --enable-multimodal \
    --mm-attention-backend ascend_attn \
    --dtype bfloat16 \
    --mamba-ssm-dtype bfloat16 \
    --speculative-algorithm NEXTN \
    --speculative-num-steps 3 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 4 \
    --reasoning-parser qwen3 \
    --tool-call-parser qwen3_coder

벤치마크 (Benchmark)

IMAGE 데이터셋을 1920x1080 해상도로 테스트했어요.

python -m sglang.bench_serving \
    --dataset-name image \
    --backend sglang-oai-chat \
    --host 127.0.0.1 \
    --port 6688 \
    --max-concurrency 42 \
    --num-prompts 42 \
    --warmup-requests 42 \
    --random-input-len 30 \
    --random-output-len 256 \
    --random-range-ratio 1 \
    --image-resolution 1920x1080 \
    --image-count 1 \
    --seed 1 \
    --request-rate inf

Qwen3.6-35B-A3B 1P IN128K OUT1K 50ms

모델 (Model): Qwen3.6-35B-A3B

하드웨어 (Hardware): Ascend A3 Series Products

카드 (Cards): 1

배포 모드 (Deploy Mode): PD Mixed

양자화 (Quantization): BF16

데이터셋 (Dataset): 128k+1k

TPOT: 50ms

모델 배포 (Model Deployment)

# ============================================================
# 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 ASCEND_USE_FIA=1
export DEEPEP_HCCL_BUFFSIZE=1600
export GDN_ATTN_BACKEND_TRITON=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_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES=20
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 \
    --tp-size 2 \
    --nnodes 1 \
    --attention-backend ascend \
    --device npu \
    --chunked-prefill-size -1 \
    --max-total-tokens 420000 \
    --max-prefill-tokens 128000 \
    --disable-radix-cache \
    --trust-remote-code \
    --max-running-requests 3 \
    --max-mamba-cache-size 3 \
    --mem-fraction-static 0.9 \
    --cuda-graph-bs-decode 1 2 3 \
    --enable-multimodal \
    --mm-attention-backend ascend_attn \
    --dtype bfloat16 \
    --mamba-ssm-dtype bfloat16 \
    --speculative-algorithm NEXTN \
    --speculative-num-steps 3 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 4 \
    --reasoning-parser qwen3 \
    --tool-call-parser qwen3_coder

벤치마크 (Benchmark)

RANDOM 데이터셋으로 테스트했어요.

python -m sglang.bench_serving \
    --dataset-name random \
    --backend sglang \
    --host 127.0.0.1 \
    --port 6688 \
    --max-concurrency 3 \
    --warmup-requests 3 \
    --num-prompts 3 \
    --random-input-len 128000 \
    --random-output-len 1000 \
    --random-range-ratio 1 \
    --seed 1

Qwen3.6-35B-A3B 1P IN128K OUT1K PREFIX90 50ms

모델 (Model): Qwen3.6-35B-A3B

하드웨어 (Hardware): Ascend A3 Series Products

카드 (Cards): 1

배포 모드 (Deploy Mode): PD Mixed

양자화 (Quantization): BF16

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

TPOT: 50ms

모델 배포 (Model Deployment)

# ============================================================
# 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 ASCEND_USE_FIA=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_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES=30
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 \
    --tp-size 2 \
    --nnodes 1 \
    --attention-backend ascend \
    --device npu \
    --chunked-prefill-size 16384 \
    --max-prefill-tokens 65536 \
    --trust-remote-code \
    --enable-prefill-delayer \
    --mamba-radix-cache-strategy extra_buffer \
    --max-running-requests 103 \
    --max-mamba-cache-size 85 \
    --mem-fraction-static 0.85 \
    --cuda-graph-bs-decode 2 4 8 16 32 48 64 80 96 103 \
    --enable-multimodal \
    --mm-attention-backend ascend_attn \
    --dtype bfloat16 \
    --mamba-ssm-dtype bfloat16 \
    --speculative-algorithm NEXTN \
    --speculative-num-steps 3 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 4 \
    --reasoning-parser qwen3 \
    --tool-call-parser qwen3_coder

벤치마크 (Benchmark)

90% 캐시 히트(repeat_rate = 0.9)를 가진 generated-shared-prefix 데이터셋으로 테스트했어요: --gsp-system-prompt-len 115200 = round(128000 * 0.9)은 공유 프리픽스(shared prefix) 부분이에요. --gsp-question-len 12800 = round(128000 * (1 - 0.9))은 요청별 고유한 접미사(suffix) 부분이에요. --gsp-num-groups 1은 캐시 재사용을 최대화하도록 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 412 \
    --gsp-system-prompt-len 115200 \
    --gsp-question-len 12800 \
    --gsp-output-len 1000 \
    --max-concurrency 103 \
    --num-prompts 412 \
    --request-rate inf

Qwen3.6-35B-A3B 1P IN254K OUT1K

모델 (Model): Qwen3.6-35B-A3B

하드웨어 (Hardware): Ascend A3 Series Products

카드 (Cards): 1

배포 모드 (Deploy Mode): PD Mixed

양자화 (Quantization): BF16

데이터셋 (Dataset): 254k+1k

TPOT: 16.1ms

모델 배포 (Model Deployment)

# ============================================================
# 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 ASCEND_USE_FIA=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_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 \
    --tp-size 2 \
    --nnodes 1 \
    --attention-backend ascend \
    --device npu \
    --chunked-prefill-size 131072 \
    --max-prefill-tokens 254000 \
    --disable-radix-cache \
    --trust-remote-code \
    --max-running-requests 1 \
    --max-mamba-cache-size 6 \
    --mem-fraction-static 0.65 \
    --cuda-graph-bs-decode 1 \
    --enable-multimodal \
    --mm-attention-backend ascend_attn \
    --dtype bfloat16 \
    --mamba-ssm-dtype bfloat16 \
    --speculative-algorithm NEXTN \
    --speculative-num-steps 3 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 4 \
    --reasoning-parser qwen3 \
    --tool-call-parser qwen3_coder

벤치마크 (Benchmark)

RANDOM 데이터셋으로 테스트했어요.

python -m sglang.bench_serving \
    --dataset-name random \
    --backend sglang \
    --host 127.0.0.1 \
    --port 6688 \
    --max-concurrency 1 \
    --warmup-requests 1 \
    --num-prompts 1 \
    --random-input-len 254000 \
    --random-output-len 1000 \
    --random-range-ratio 1 \
    --seed 1

Qwen3.6-35B-A3B 1P IN3K5 OUT1K5 50ms

모델 (Model): Qwen3.6-35B-A3B

하드웨어 (Hardware): Ascend A3 Series Products

카드 (Cards): 1

배포 모드 (Deploy Mode): PD Mixed

양자화 (Quantization): BF16

데이터셋 (Dataset): 3.5k+1.5k

TPOT: 50ms

모델 배포 (Model Deployment)

# ============================================================
# 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 ASCEND_USE_FIA=1
export GLOO_SOCKET_IFNAME=<network-interface>
export HCCL_BUFFSIZE=100
export HCCL_OP_EXPANSION_MODE=AIV
export HCCL_SOCKET_IFNAME=<network-interface>
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=0
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 \
    --tp-size 2 \
    --nnodes 1 \
    --attention-backend ascend \
    --device npu \
    --chunked-prefill-size -1 \
    --max-total-tokens 659840 \
    --max-prefill-tokens 43400 \
    --disable-radix-cache \
    --trust-remote-code \
    --prefill-max-requests 12 \
    --max-running-requests 122 \
    --max-mamba-cache-size 122 \
    --mem-fraction-static 0.9 \
    --cuda-graph-bs-decode 4 16 32 64 96 116 120 122 \
    --enable-multimodal \
    --mm-attention-backend ascend_attn \
    --dtype bfloat16 \
    --mamba-ssm-dtype bfloat16 \
    --speculative-algorithm NEXTN \
    --speculative-num-steps 3 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 4 \
    --reasoning-parser qwen3 \
    --tool-call-parser qwen3_coder

벤치마크 (Benchmark)

RANDOM 데이터셋으로 테스트했어요.

python -m sglang.bench_serving \
    --dataset-name random \
    --backend sglang \
    --host 127.0.0.1 \
    --port 6688 \
    --max-concurrency 122 \
    --warmup-requests 122 \
    --num-prompts 122 \
    --random-input-len 3500 \
    --random-output-len 1500 \
    --random-range-ratio 1 \
    --seed 1

Qwen3.6-35B-A3B 1P IN64K OUT1K 50ms

모델 (Model): Qwen3.6-35B-A3B

하드웨어 (Hardware): Ascend A3 Series Products

카드 (Cards): 1

배포 모드 (Deploy Mode): PD Mixed

양자화 (Quantization): BF16

데이터셋 (Dataset): 64k+1k

TPOT: 50ms

모델 배포 (Model Deployment)

# ============================================================
# 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 ASCEND_USE_FIA=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_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES=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 \
    --tp-size 2 \
    --nnodes 1 \
    --attention-backend ascend \
    --device npu \
    --chunked-prefill-size -1 \
    --max-total-tokens 600000 \
    --max-prefill-tokens 65536 \
    --disable-radix-cache \
    --trust-remote-code \
    --enable-prefill-delayer \
    --max-running-requests 10 \
    --max-mamba-cache-size 20 \
    --mem-fraction-static 0.65 \
    --cuda-graph-bs-decode 2 4 8 12 14 16 \
    --enable-multimodal \
    --mm-attention-backend ascend_attn \
    --dtype bfloat16 \
    --mamba-ssm-dtype bfloat16 \
    --speculative-algorithm NEXTN \
    --speculative-num-steps 3 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 4 \
    --reasoning-parser qwen3 \
    --tool-call-parser qwen3_coder

벤치마크 (Benchmark)

RANDOM 데이터셋으로 테스트했어요.

python -m sglang.bench_serving \
    --dataset-name random \
    --backend sglang \
    --host 127.0.0.1 \
    --port 6688 \
    --max-concurrency 10 \
    --warmup-requests 10 \
    --num-prompts 40 \
    --random-input-len 64000 \
    --random-output-len 1000 \
    --random-range-ratio 1 \
    --seed 1

Qwen3.6-35B-A3B 1P IN64K OUT1K PREFIX90 50ms

모델 (Model): Qwen3.6-35B-A3B

하드웨어 (Hardware): Ascend A3 Series Products

카드 (Cards): 1

배포 모드 (Deploy Mode): PD Mixed

양자화 (Quantization): BF16

데이터셋 (Dataset): 64k+1k (90% prefix cache hit rate)

TPOT: 50ms

모델 배포 (Model Deployment)

# ============================================================
# 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 ASCEND_USE_FIA=1
export DEEPEP_HCCL_BUFFSIZE=300
export GDN_ATTN_BACKEND_TRITON=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_ENABLE_OVERLAP_PLAN_STREAM=0
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 \
    --tp-size 2 \
    --nnodes 1 \
    --attention-backend ascend \
    --device npu \
    --chunked-prefill-size -1 \
    --max-total-tokens 470784 \
    --max-prefill-tokens 65536 \
    --trust-remote-code \
    --mamba-radix-cache-strategy extra_buffer \
    --max-running-requests 40 \
    --max-mamba-cache-size 200 \
    --mem-fraction-static 0.9 \
    --cuda-graph-bs-decode 2 8 16 24 32 36 40 \
    --enable-multimodal \
    --mm-attention-backend ascend_attn \
    --dtype bfloat16 \
    --mamba-ssm-dtype bfloat16 \
    --speculative-algorithm NEXTN \
    --speculative-num-steps 3 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 4 \
    --reasoning-parser qwen3 \
    --tool-call-parser qwen3_coder

벤치마크 (Benchmark)

90% 캐시 히트(repeat_rate = 0.9)를 가진 generated-shared-prefix 데이터셋으로 테스트했어요: --gsp-system-prompt-len 58982 = round(65536 * 0.9)은 공유 프리픽스(shared prefix) 부분이에요. --gsp-question-len 6554 = round(65536 * (1 - 0.9))은 요청별 고유한 접미사(suffix) 부분이에요. --gsp-num-groups 1은 캐시 재사용을 최대화하도록 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 40 \
    --gsp-system-prompt-len 58982 \
    --gsp-question-len 6554 \
    --gsp-output-len 1024 \
    --max-concurrency 40 \
    --num-prompts 40 \
    --request-rate inf

Qwen3.6-35B-A3B 2P IN984K OUT1K

모델 (Model): Qwen3.6-35B-A3B

하드웨어 (Hardware): Ascend A3 Series Products

카드 (Cards): 2

배포 모드 (Deploy Mode): PD Mixed

양자화 (Quantization): BF16

데이터셋 (Dataset): 984k+1k

TPOT: 40.91ms

모델 배포 (Model Deployment)

# ============================================================
# 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 ASCEND_USE_FIA=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_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1
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 \
    --tp-size 4 \
    --nnodes 1 \
    --attention-backend ascend \
    --device npu \
    --chunked-prefill-size 131072 \
    --max-prefill-tokens 984000 \
    --disable-radix-cache \
    --trust-remote-code \
    --max-running-requests 1 \
    --max-mamba-cache-size 6 \
    --mem-fraction-static 0.68 \
    --cuda-graph-bs-decode 1 \
    --enable-multimodal \
    --mm-attention-backend ascend_attn \
    --dtype bfloat16 \
    --mamba-ssm-dtype bfloat16 \
    --speculative-algorithm NEXTN \
    --speculative-num-steps 3 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 4 \
    --context-length 1010000 \
    --reasoning-parser qwen3 \
    --tool-call-parser qwen3_coder

벤치마크 (Benchmark)

RANDOM 데이터셋으로 테스트했어요.

python -m sglang.bench_serving \
    --dataset-name random \
    --backend sglang \
    --host 127.0.0.1 \
    --port 6688 \
    --max-concurrency 1 \
    --warmup-requests 1 \
    --num-prompts 1 \
    --random-input-len 984000 \
    --random-output-len 1000 \
    --random-range-ratio 1 \
    --seed 1

더 알아보기 (Learn more)