Qwen3.6-27B - SGLang Documentation

This page focuses on optimal configuration and benchmark results for Qwen3.6-27B on the Ascend NPU. For environment setup, model weight download, feature configuration, and deployment instructions, etc., see the Qwen3.6-27B Model Tutorial.On A3 each card has 2 dies, so --tp-size is twice the card count; see Ascend NPU Reference for details.

High Throughput

Model	Hardware	Cards	Deploy Mode	Dataset	TPOT	Quantization	Configuration
Qwen3.6-27B	Atlas 800I A3	1	PD Mixed	1024x1024 (30)+1024	50ms	BF16	Optimal Configuration
Qwen3.6-27B	Atlas 800I A3	1	PD Mixed	1080p_30+256	50ms	BF16	Optimal Configuration
Qwen3.6-27B	Atlas 800I A3	1	PD Mixed	64k+1k (90% prefix cache hit rate)	50ms	BF16	Optimal Configuration
Qwen3.6-27B	Atlas 800I A3	1	PD Mixed	3.5k+1.5k	50ms	W8A8 INT8	Optimal Configuration
Qwen3.6-27B	Atlas 800I A3	1	PD Mixed	64k+1k	50ms	W8A8 INT8	Optimal Configuration
Qwen3.6-27B	Atlas 800I A3	2	PD Mixed	128k+1k	50ms	W8A8 INT8	Optimal Configuration
Qwen3.6-27B	Atlas 800I A3	2	PD Mixed	16k+1k	50ms	W8A8 INT8	Optimal Configuration

Optimal Configuration

Qwen3.6-27B 1P IN1024X1024 30 OUT1024 50ms

Model: Qwen3.6-27B Hardware: Atlas 800I A3 Cards: 1 Deploy Mode: PD Mixed Quantization: BF16 Dataset: 1024x1024 (30)+1024 Format: resolution (input tokens) + output tokens TPOT: 50ms

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 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=150
export SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE=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 52000 \
    --disable-radix-cache \
    --trust-remote-code \
    --max-running-requests 50 \
    --max-mamba-cache-size 60 \
    --mem-fraction-static 0.76 \
    --cuda-graph-bs 2 4 8 16 24 32 40 42 45 50 \
    --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 \
    --mm-enable-dp-encoder \
    --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 48 \
    --random-input-len 30 \
    --random-output-len 1024 \
    --num-prompts 48 \
    --random-range-ratio 1

Qwen3.6-27B 1P IN1080P 30 OUT256 50ms

Model: Qwen3.6-27B Hardware: Atlas 800I A3 Cards: 1 Deploy Mode: PD Mixed Quantization: BF16 Dataset: 1080p_30+256 TPOT: 50ms

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 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=150
export SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE=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 48000 \
    --disable-radix-cache \
    --trust-remote-code \
    --max-running-requests 30 \
    --max-mamba-cache-size 40 \
    --mem-fraction-static 0.76 \
    --cuda-graph-bs 2 4 8 16 24 28 30 \
    --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 \
    --mm-enable-dp-encoder \
    --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 30 \
    --random-input-len 30 \
    --random-output-len 256 \
    --num-prompts 120 \
    --random-range-ratio 1

Qwen3.6-27B 1P IN64K OUT1K PREFIX90 50ms

Model: Qwen3.6-27B Hardware: Atlas 800I A3 Cards: 1 Deploy Mode: PD Mixed Quantization: BF16 Dataset: 64k+1k (90% prefix cache hit rate) TPOT: 50ms

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 ASCEND_USE_FIA=1
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_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 32768 \
    --max-prefill-tokens 32768 \
    --mamba-radix-cache-strategy extra_buffer \
    --trust-remote-code \
    --max-running-requests 20 \
    --max-mamba-cache-size 160 \
    --mem-fraction-static 0.82 \
    --cuda-graph-bs 1 2 5 10 15 17 19 20 \
    --enable-prefill-delayer \
    --prefill-delayer-queue-min-ratio 0.7 \
    --prefill-delayer-max-delay-ms 20000 \
    --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

Benchmark

We tested it based on the generated-shared-prefix dataset with 90% cache hit (repeat_rate = 0.9): --gsp-system-prompt-len 57600 = int(64000 * 0.9) is the shared prefix portion. --gsp-question-len 6400 = int(64000 * (1 - 0.9)) is the unique per-request suffix. --gsp-num-groups 1 keeps all requests in one prefix group for maximum cache reuse.

Command

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 80 \
    --gsp-system-prompt-len 57600 \
    --gsp-question-len 6400 \
    --gsp-output-len 1000 \
    --max-concurrency 20 \
    --num-prompts 80 \
    --request-rate inf

Qwen3.6-27B W8A8 1P IN3K5 OUT1K5 50ms

Model: Qwen3.6-27B Hardware: Atlas 800I A3 Cards: 1 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
#   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=0
export SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES=130
export SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE=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-prefill-tokens 60000 \
    --disable-radix-cache \
    --trust-remote-code \
    --max-running-requests 64 \
    --max-mamba-cache-size 74 \
    --mem-fraction-static 0.7 \
    --cuda-graph-bs 2 8 16 32 40 45 50 54 \
    --enable-multimodal \
    --quantization modelslim \
    --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

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 54 \
    --random-input-len 3500 \
    --random-output-len 1500 \
    --num-prompts 216 \
    --random-range-ratio 1

Qwen3.6-27B W8A8 1P IN64K OUT1K 50ms

Model: Qwen3.6-27B Hardware: Atlas 800I A3 Cards: 1 Deploy Mode: PD Mixed Quantization: W8A8 INT8 Dataset: 64k+1k TPOT: 50ms

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 GLOO_SOCKET_IFNAME=<network-interface>
export HCCL_OP_EXPANSION_MODE=AIV
export HCCL_SOCKET_IFNAME=<network-interface>
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 -1 \
    --max-prefill-tokens 48000 \
    --disable-radix-cache \
    --trust-remote-code \
    --max-running-requests 6 \
    --max-mamba-cache-size 16 \
    --mem-fraction-static 0.6 \
    --cuda-graph-bs 1 2 4 5 6 \
    --quantization modelslim \
    --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

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 6 \
    --random-input-len 64000 \
    --random-output-len 1000 \
    --num-prompts 12 \
    --random-range-ratio 1

Qwen3.6-27B W8A8 2P IN128K OUT1K 50ms

Model: Qwen3.6-27B Hardware: Atlas 800I A3 Cards: 2 Deploy Mode: PD Mixed Quantization: W8A8 INT8 Dataset: 128k+1k TPOT: 50ms

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 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=20
export SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE=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 -1 \
    --max-prefill-tokens 74000 \
    --disable-radix-cache \
    --trust-remote-code \
    --max-running-requests 6 \
    --max-mamba-cache-size 7 \
    --mem-fraction-static 0.63 \
    --cuda-graph-bs 1 2 4 5 6 \
    --enable-multimodal \
    --quantization modelslim \
    --mm-attention-backend ascend_attn \
    --dtype bfloat16 \
    --mamba-ssm-dtype bfloat16 \
    --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 4 \
    --random-input-len 128000 \
    --random-output-len 1000 \
    --num-prompts 16 \
    --random-range-ratio 1

Qwen3.6-27B W8A8 2P IN16K OUT1K 50ms

Model: Qwen3.6-27B Hardware: Atlas 800I A3 Cards: 2 Deploy Mode: PD Mixed Quantization: W8A8 INT8 Dataset: 16k+1k TPOT: 50ms

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 ASCEND_USE_FIA=1
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 SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES=100
export SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE=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 -1 \
    --max-prefill-tokens 58000 \
    --disable-radix-cache \
    --trust-remote-code \
    --max-running-requests 29 \
    --max-mamba-cache-size 58 \
    --mem-fraction-static 0.68 \
    --cuda-graph-bs 1 2 8 12 16 20 24 26 28 29 \
    --quantization modelslim \
    --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

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 29 \
    --random-input-len 16000 \
    --random-output-len 1000 \
    --num-prompts 116 \
    --random-range-ratio 1

​High Throughput

​Optimal Configuration

​Qwen3.6-27B 1P IN1024X1024 30 OUT1024 50ms

​Model Deployment

​Benchmark

​Qwen3.6-27B 1P IN1080P 30 OUT256 50ms

​Model Deployment

​Benchmark

​Qwen3.6-27B 1P IN64K OUT1K PREFIX90 50ms

​Model Deployment

​Benchmark

​Qwen3.6-27B W8A8 1P IN3K5 OUT1K5 50ms

​Model Deployment

​Benchmark

​Qwen3.6-27B W8A8 1P IN64K OUT1K 50ms

​Model Deployment

​Benchmark

​Qwen3.6-27B W8A8 2P IN128K OUT1K 50ms

​Model Deployment

​Benchmark

​Qwen3.6-27B W8A8 2P IN16K OUT1K 50ms

​Model Deployment

​Benchmark

High Throughput

Optimal Configuration

Qwen3.6-27B 1P IN1024X1024 30 OUT1024 50ms

Model Deployment

Benchmark

Qwen3.6-27B 1P IN1080P 30 OUT256 50ms

Model Deployment

Benchmark

Qwen3.6-27B 1P IN64K OUT1K PREFIX90 50ms

Model Deployment

Benchmark

Qwen3.6-27B W8A8 1P IN3K5 OUT1K5 50ms

Model Deployment

Benchmark

Qwen3.6-27B W8A8 1P IN64K OUT1K 50ms

Model Deployment

Benchmark

Qwen3.6-27B W8A8 2P IN128K OUT1K 50ms

Model Deployment

Benchmark

Qwen3.6-27B W8A8 2P IN16K OUT1K 50ms

Model Deployment

Benchmark