Qwen3-Next-80B-A3B-Instruct

This page focuses on optimal configuration and benchmark results for Qwen3-Next-80B-A3B-Instruct on the Ascend NPU. For environment setup, model weight download, feature configuration, and deployment instructions, etc., see the Qwen3-Next-80B-A3B-Instruct Model Tutorial.

On A3 each card has 2 dies, so --tp-size is twice the card count; see Ascend NPU Reference for details.

Low Latency

Model	Hardware	Cards	Deploy Mode	Dataset	TPOT	Quantization	Configuration
Qwen3-Next-80B-A3B-Instruct	Atlas 800I A3	1	PD Mixed	3.5k+1.5k	20ms	W8A8 INT8	Optimal Configuration
Qwen3-Next-80B-A3B-Instruct	Atlas 800I A3	2	PD Mixed	6k+1.5k	15.62ms	W8A8 INT8	Optimal Configuration

High Throughput

Model	Hardware	Cards	Deploy Mode	Dataset	TPOT	Quantization	Configuration
Qwen3-Next-80B-A3B-Instruct	Atlas 800I A3	2	PD Mixed	3.5k+1.5k	50ms	W8A8 INT8	Optimal Configuration

Optimal Configuration

Qwen3-Next-80B-A3B-Instruct W8A8 1P IN3K5 OUT1K5 20ms

Model: Qwen3-Next-80B-A3B-Instruct

Hardware: Atlas 800I A3

Cards: 1

Deploy Mode: PD Mixed

Quantization: W8A8 INT8

Dataset: 3.5k+1.5k

TPOT: 20ms

Model Deployment

Command

# ============================================================
# 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 ASCEND_USE_FIA=1
export DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS=2048
export DEEPEP_NORMAL_LONG_SEQ_ROUND=10
export FORCE_DRAFT_MODEL_NON_QUANT=1
export GLOO_SOCKET_IFNAME=<network-interface>
export HCCL_BUFFSIZE=2000
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=400
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK=0
export SGLANG_NPU_USE_MULTI_STREAM=0
export SGLANG_SET_CPU_AFFINITY=1
export SGLANG_WARMUP_TIMEOUT=3600
export STREAMS_PER_DEVICE=32
export TASK_QUEUE_ENABLE=1
export ZBCCL_BOOTSTRAP_URL=tcp://127.0.0.1:24669
export ZBCCL_ENABLE_GRAPH=1
export ZBCCL_LOCAL_MEM_SIZE=60416
export ZBCCL_NPU_ALLOC_CONF=use_vmm_for_static_memory:True

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 \
    --page-size 128 \
    --tp-size 2 \
    --watchdog-timeout 9000 \
    --mem-fraction-static 0.85 \
    --disable-radix-cache \
    --max-prefill-tokens 28672 \
    --context-length 26384 \
    --max-total-tokens 122304 \
    --speculative-algorithm NEXTN \
    --speculative-num-steps 3 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 4 \
    --speculative-draft-model-quantization unquant \
    --chunked-prefill-size -1 \
    --max-running-requests 2 \
    --cuda-graph-bs 2 \
    --mamba-ssm-dtype bfloat16 \
    --speculative-draft-model-path $DRAFT_MODEL_PATH \
    --reasoning-parser qwen3 \
    --tool-call-parser qwen

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 1 \
    --random-input-len 3500 \
    --random-output-len 1500 \
    --num-prompts 1 \
    --random-range-ratio 1

Qwen3-Next-80B-A3B-Instruct W8A8 2P IN3K5 OUT1K5 50ms

Model: Qwen3-Next-80B-A3B-Instruct

Hardware: Atlas 800I A3

Cards: 2

Deploy Mode: PD Mixed

Quantization: W8A8 INT8

Dataset: 3.5k+1.5k

TPOT: 50ms

Model Deployment

Command

# ============================================================
# 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 ASCEND_USE_FIA=1
export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
export FORCE_DRAFT_MODEL_NON_QUANT=1
export GLOO_SOCKET_IFNAME=<network-interface>
export HCCL_BUFFSIZE=64
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=330
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK=0
export SGLANG_NPU_USE_MULTI_STREAM=0
export SGLANG_SET_CPU_AFFINITY=1
export SGLANG_WARMUP_TIMEOUT=3600
export SGLANG_ZBAL_BOOTSTRAP_URL=tcp://127.0.0.1:24669
export SGLANG_ZBAL_LOCAL_MEM_SIZE=59648
export STREAMS_PER_DEVICE=32
export ZBAL_ENABLE_GRAPH=1
export ZBAL_HCCL_OP=allreduce,_allgather_base,allgather,broadcast,scatter,reduce_scatter,_reduce_scatter_base,alltoall_base
export ZBAL_NPU_ALLOC_CONF=use_vmm_for_static_memory:True

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 \
    --page-size 128 \
    --tp-size 4 \
    --watchdog-timeout 9000 \
    --mem-fraction-static 0.75 \
    --disable-radix-cache \
    --max-prefill-tokens 14080 \
    --context-length 26384 \
    --chunked-prefill-size -1 \
    --max-running-requests 300 \
    --mamba-ssm-dtype bfloat16 \
    --speculative-algorithm NEXTN \
    --speculative-num-steps 3 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 4 \
    --speculative-draft-model-quantization unquant \
    --speculative-draft-model-path $DRAFT_MODEL_PATH \
    --dp-size 2 \
    --enable-dp-attention \
    --enable-dp-lm-head \
    --moe-a2a-backend deepep \
    --deepep-mode auto \
    --cuda-graph-bs 1 2 3 4 5 6 7 8 10 12 14 16 18 20 22 24 26 28 30 32 40 44 48 52 56 60 64 72 80 88 96 104 112 120 128 136 144 150 \
    --reasoning-parser qwen3 \
    --tool-call-parser qwen

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 300 \
    --random-input-len 3500 \
    --random-output-len 1500 \
    --num-prompts 300 \
    --random-range-ratio 1

Qwen3-Next-80B-A3B-Instruct W8A8 2P IN6K OUT1K5 BS16

Model: Qwen3-Next-80B-A3B-Instruct

Hardware: Atlas 800I A3

Cards: 2

Deploy Mode: PD Mixed

Quantization: W8A8 INT8

Dataset: 6k+1.5k

TPOT: 15.62ms

Model Deployment

Command

# ============================================================
# 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 ASCEND_USE_FIA=1
export DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS=2048
export DEEPEP_NORMAL_LONG_SEQ_ROUND=10
export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
export FORCE_DRAFT_MODEL_NON_QUANT=1
export GLOO_SOCKET_IFNAME=<network-interface>
export HCCL_BUFFSIZE=2000
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=400
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK=0
export SGLANG_NPU_USE_MULTI_STREAM=0
export SGLANG_WARMUP_TIMEOUT=3600
export STREAMS_PER_DEVICE=32
export TASK_QUEUE_ENABLE=1
export ZBCCL_BOOTSTRAP_URL=tcp://127.0.0.1:24669
export ZBCCL_ENABLE_GRAPH=1
export ZBCCL_LOCAL_MEM_SIZE=60416
export ZBCCL_NPU_ALLOC_CONF=use_vmm_for_static_memory:True

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 \
    --page-size 128 \
    --tp-size 4 \
    --watchdog-timeout 9000 \
    --mem-fraction-static 0.85 \
    --disable-radix-cache \
    --max-prefill-tokens 28672 \
    --context-length 81920 \
    --max-total-tokens 122304 \
    --dp-size 2 \
    --enable-dp-attention \
    --enable-dp-lm-head \
    --speculative-algorithm NEXTN \
    --speculative-num-steps 3 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 4 \
    --speculative-draft-model-quantization unquant \
    --chunked-prefill-size -1 \
    --max-running-requests 16 \
    --cuda-graph-bs 2 4 8 \
    --mamba-ssm-dtype bfloat16 \
    --speculative-draft-model-path $DRAFT_MODEL_PATH \
    --reasoning-parser qwen3 \
    --tool-call-parser qwen3_coder

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 16 \
    --random-input-len 6144 \
    --random-output-len 1500 \
    --num-prompts 16 \
    --random-range-ratio 1