Qwen3-8B - SGLang Documentation

This page focuses on optimal configuration and benchmark results for Qwen3-8B on the Ascend NPU. For environment setup, model weight download, feature configuration, and deployment instructions, etc., see the Qwen3-8B 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
Qwen3-8B	Atlas 800I A3	1	PD Mixed	3.5k+1.5k	5ms	W8A8 INT8	Optimal Configuration
Qwen3-8B	Atlas 800I A3	1	PD Mixed	6k+1.5k	11.79ms	W8A8 INT8	Optimal Configuration

High Throughput

Model	Hardware	Cards	Deploy Mode	Dataset	TPOT	Quantization	Configuration
Qwen3-8B	Atlas 800I A3	1	PD Mixed	3.5k+1.5k	37ms	W8A8 INT8	Optimal Configuration

Optimal Configuration

Qwen3-8B W8A8 1P IN3K5 OUT1K5 37ms

Model: Qwen3-8B Hardware: Atlas 800I A3 Cards: 1 Deploy Mode: PD Mixed Quantization: W8A8 INT8 Dataset: 3.5k+1.5k TPOT: 37ms

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 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_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES=50
export SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE=1

python3 -m sglang.launch_server \
    --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 6688 \
    --trust-remote-code \
    --nnodes 1 \
    --node-rank 0 \
    --attention-backend ascend \
    --device npu \
    --quantization modelslim \
    --max-running-requests 70 \
    --max-prefill-tokens 16384 \
    --disable-radix-cache \
    --chunked-prefill-size 16384 \
    --tp-size 1 \
    --mem-fraction-static 0.85 \
    --cuda-graph-bs 8 12 24 36 48 51 55 60 63 64 66 68 70 \
    --dtype bfloat16 \
    --speculative-draft-model-quantization unquant \
    --speculative-algorithm EAGLE3 \
    --speculative-draft-model-path $DRAFT_MODEL_PATH \
    --speculative-num-steps 3 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 4 \
    --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 64 \
    --random-input-len 3500 \
    --random-output-len 1500 \
    --num-prompts 256 \
    --random-range-ratio 1

Qwen3-8B W8A8 1P IN3K5 OUT1K5 5ms

Model: Qwen3-8B Hardware: Atlas 800I A3 Cards: 1 Deploy Mode: PD Mixed Quantization: W8A8 INT8 Dataset: 3.5k+1.5k TPOT: 5ms

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 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_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1

python3 -m sglang.launch_server \
    --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 6688 \
    --trust-remote-code \
    --nnodes 1 \
    --node-rank 0 \
    --attention-backend ascend \
    --device npu \
    --quantization modelslim \
    --max-running-requests 1 \
    --max-prefill-tokens 16384 \
    --disable-radix-cache \
    --chunked-prefill-size -1 \
    --tp-size 2 \
    --mem-fraction-static 0.894 \
    --cuda-graph-bs 1 \
    --dtype bfloat16 \
    --speculative-draft-model-quantization unquant \
    --speculative-algorithm EAGLE3 \
    --speculative-draft-model-path $DRAFT_MODEL_PATH \
    --speculative-num-steps 4 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 5 \
    --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 4 \
    --random-range-ratio 1

Qwen3-8B W8A8 1P IN6K OUT1K5 BS16

Model: Qwen3-8B Hardware: Atlas 800I A3 Cards: 1 Deploy Mode: PD Mixed Quantization: W8A8 INT8 Dataset: 6k+1.5k TPOT: 11.79ms

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 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_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1

python3 -m sglang.launch_server \
    --model-path $MODEL_PATH \
    --host 127.0.0.1 --port 6688 \
    --trust-remote-code \
    --nnodes 1 \
    --node-rank 0 \
    --attention-backend ascend \
    --device npu \
    --quantization modelslim \
    --max-running-requests 16 \
    --max-prefill-tokens 16384 \
    --disable-radix-cache \
    --chunked-prefill-size -1 \
    --tp-size 2 \
    --mem-fraction-static 0.894 \
    --cuda-graph-bs 1 5 15 16 \
    --dtype bfloat16 \
    --speculative-draft-model-quantization unquant \
    --speculative-algorithm EAGLE3 \
    --speculative-draft-model-path $DRAFT_MODEL_PATH \
    --speculative-num-steps 4 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 5 \
    --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 16 \
    --random-input-len 6144 \
    --random-output-len 1500 \
    --num-prompts 16 \
    --random-range-ratio 1

​Low Latency

​High Throughput

​Optimal Configuration

​Qwen3-8B W8A8 1P IN3K5 OUT1K5 37ms

​Model Deployment

​Benchmark

​Qwen3-8B W8A8 1P IN3K5 OUT1K5 5ms

​Model Deployment

​Benchmark

​Qwen3-8B W8A8 1P IN6K OUT1K5 BS16

​Model Deployment

​Benchmark

Low Latency

High Throughput

Optimal Configuration

Qwen3-8B W8A8 1P IN3K5 OUT1K5 37ms

Model Deployment

Benchmark

Qwen3-8B W8A8 1P IN3K5 OUT1K5 5ms

Model Deployment

Benchmark

Qwen3-8B W8A8 1P IN6K OUT1K5 BS16

Model Deployment

Benchmark