GLM-5.2

Introduction

GLM-5.2 is a large language model in the GLM (General Language Model) series, jointly developed by the KEG Laboratory of Tsinghua University and Zhipu AI. GLM-5.2 adopts the DeepSeek-V3/V3.2 architecture, including DeepSeek Sparse Attention (DSA) and multi-token prediction (MTP), and supports high-throughput inference with SGLang on Ascend NPUs.

This document demonstrates the deployment of GLM-5.2 on Ascend NPUs using SGLang, including single-node deployment, multi-node deployment, prefill-decode disaggregation, feature configuration, and performance optimization.

GLM-5.2 currently uses a daily build image because 0Day support was released before the related code was merged into the SGLang main branch. The deployment commands and image tags in this document will be updated to a stable release once the support lands in a stable SGLang release.

Supported features

Feature	Example usage
Tensor Parallelism	--tp-size 16
Data Parallelism	--dp-size 16
Expert Parallelism	--ep-size 16 \ --moe-a2a-backend deepep \ --deepep-mode auto
PD Disaggregation	--disaggregation-mode prefill \ --disaggregation-transfer-backend ascend
Quantization	--quantization modelslim
Chunked Prefill	auto based on device memory, or set explicit value; disable with --chunked-prefill-size -1; e.g., --chunked-prefill-size 16384
NPU Graph	enabled by default; disable with --disable-cuda-graph; control range via --cuda-graph-bs or --cuda-graph-max-bs; e.g., --cuda-graph-bs 16
Speculative Decoding	--speculative-algorithm NEXTN \ --speculative-num-steps 3 \ --speculative-eagle-topk 1 \ --speculative-num-draft-tokens 4 \ --speculative-draft-model-quantization unquant
Overlap Schedule	export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
DP LM Head	--enable-dp-lm-head

The values in the Example usage column are for illustration only. Adjust them according to your hardware, deployment mode, and workload. For parameter details, see Feature descriptions; for recommended configurations for each deployment scenario, see Best practices.

For feature compatibility and conflict information between features, see Feature Compatibility.

Prerequisites

Environment

Before following this tutorial, complete the environment setup in the documents below:

• Ascend NPU Quickstart — the fastest way to get started. It walks you through launching the official container image, starting the SGLang server, and sending a test request. Recommended if you are new to SGLang on Ascend.
• SGLang Installation with NPU Support — the full installation guide. It covers the component version mapping (CANN, TorchNPU, Triton, kernels, etc.), building from source or from a Dockerfile, and recommended system settings (CPU power scheme, NUMA, swap). Use it when you need to install or customize the environment instead of using the official image.

Model weights

Before downloading model weights, check the model size to reserve enough disk space. For multi-node deployment, download the weights to a shared directory accessible to all nodes.

• GLM-5.2 (BF16, 1.51TB)
• GLM-5.2-w8a8 (Quantized version without MTP, 774.08GB)
• You can use msmodelslim to quantize the model naively.

We recommend deploying the W8A8 variant for reduced resource usage and higher throughput. It (774.08GB) can be deployed on 16 × 64GB of device memory (--tp-size 16), which corresponds to one full A3 node (8 cards, 16 dies) or two A2 nodes.

This is the minimum recommended configuration. For optimized configurations, see Best practices, which may require additional cards or nodes.

For the hardware specifications (memory per die, dies per card, and the difference between A2 and A3), see Ascend NPU Reference — Hardware.

Installation

Ensure sufficient disk space before pulling images. The Docker image requires at least 30GB of free space.

The dependencies required for the NPU runtime environment have been integrated into a Docker image and uploaded to the online platform. You can directly pull it.

The GLM-5.2 images below use daily build tags because 0Day support was released before the related code was merged into the main branch. These tags will be switched to stable release images after the support lands in a stable release.

Atlas 800I A3

Command

docker pull swr.cn-southwest-2.myhuaweicloud.com/base_image/dockerhub/lmsysorg/sglang:cann9.0.0-a3-glm5.2-20260615

docker run -itd --shm-size=16g --name ${NAME} \
--privileged=true --net=host \
-v /var/queue_schedule:/var/queue_schedule \
-v /etc/ascend_install.info:/etc/ascend_install.info \
-v /usr/local/sbin:/usr/local/sbin \
-v /usr/local/Ascend/driver:/usr/local/Ascend/driver \
-v /usr/local/Ascend/firmware:/usr/local/Ascend/firmware \
--device=/dev/davinci0:/dev/davinci0  \
--device=/dev/davinci1:/dev/davinci1  \
--device=/dev/davinci2:/dev/davinci2  \
--device=/dev/davinci3:/dev/davinci3  \
--device=/dev/davinci4:/dev/davinci4  \
--device=/dev/davinci5:/dev/davinci5  \
--device=/dev/davinci6:/dev/davinci6  \
--device=/dev/davinci7:/dev/davinci7  \
--device=/dev/davinci8:/dev/davinci8  \
--device=/dev/davinci9:/dev/davinci9  \
--device=/dev/davinci10:/dev/davinci10  \
--device=/dev/davinci11:/dev/davinci11  \
--device=/dev/davinci12:/dev/davinci12  \
--device=/dev/davinci13:/dev/davinci13  \
--device=/dev/davinci14:/dev/davinci14  \
--device=/dev/davinci15:/dev/davinci15  \
--device=/dev/davinci_manager:/dev/davinci_manager \
--device=/dev/hisi_hdc:/dev/hisi_hdc \
--entrypoint=bash \
swr.cn-southwest-2.myhuaweicloud.com/base_image/dockerhub/lmsysorg/sglang:cann9.0.0-a3-glm5.2-20260615

Atlas 800I A2

Command

docker pull swr.cn-southwest-2.myhuaweicloud.com/base_image/dockerhub/lmsysorg/sglang:cann9.0.0-910b-glm5.2-20260615

docker run -itd --shm-size=16g --name ${NAME} \
--privileged=true --net=host \
-v /var/queue_schedule:/var/queue_schedule \
-v /etc/ascend_install.info:/etc/ascend_install.info \
-v /usr/local/sbin:/usr/local/sbin \
-v /usr/local/Ascend/driver:/usr/local/Ascend/driver \
-v /usr/local/Ascend/firmware:/usr/local/Ascend/firmware \
--device=/dev/davinci0:/dev/davinci0  \
--device=/dev/davinci1:/dev/davinci1  \
--device=/dev/davinci2:/dev/davinci2  \
--device=/dev/davinci3:/dev/davinci3  \
--device=/dev/davinci4:/dev/davinci4  \
--device=/dev/davinci5:/dev/davinci5  \
--device=/dev/davinci6:/dev/davinci6  \
--device=/dev/davinci7:/dev/davinci7  \
--device=/dev/davinci8:/dev/davinci8  \
--device=/dev/davinci9:/dev/davinci9  \
--device=/dev/davinci10:/dev/davinci10  \
--device=/dev/davinci11:/dev/davinci11  \
--device=/dev/davinci12:/dev/davinci12  \
--device=/dev/davinci13:/dev/davinci13  \
--device=/dev/davinci14:/dev/davinci14  \
--device=/dev/davinci15:/dev/davinci15  \
--device=/dev/davinci_manager:/dev/davinci_manager \
--device=/dev/hisi_hdc:/dev/hisi_hdc \
--entrypoint=bash \
swr.cn-southwest-2.myhuaweicloud.com/base_image/dockerhub/lmsysorg/sglang:cann9.0.0-910b-glm5.2-20260615

• If the model weights have already been downloaded to a shared directory, use -v to mount the model path into the container, for example: -v /path/to/models:/models.
• Replace ${NAME} with your own container name or remove --name to use default name.

Online service deployment

Single-node deployment

Quantized model GLM-5.2-w8a8 can be deployed on one Atlas 800I A3 node.

Run the following script to execute online inference.

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

# high performance cpu
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
# bind cpu
export SGLANG_SET_CPU_AFFINITY=1

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
# cann
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh

export STREAMS_PER_DEVICE=32
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600
# MTP OVERLAP
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
export SGLANG_NPU_USE_MULTI_STREAM=1

export HCCL_BUFFSIZE=1000
export HCCL_OP_EXPANSION_MODE=AIV
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
# DEEPEP
export DEEPEP_NORMAL_LONG_SEQ_ROUND=72
export DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS=1024
export DEEPEP_NORMAL_COMBINE_ENABLE_LONG_SEQ=1
export DEEP_NORMAL_MODE_USE_INT8_QUANT=1

MODEL_PATH=/path/to/model-weights

python3 -m sglang.launch_server \
        --model-path $MODEL_PATH \
        --attention-backend ascend \
        --device npu \
        --tp-size 16 --nnodes 1 --node-rank 0 \
        --chunked-prefill-size 16384 --max-prefill-tokens 280000 \
        --trust-remote-code \
        --host 127.0.0.1 \
        --mem-fraction-static 0.7 \
        --port 8000 \
        --served-model-name glm-5 \
        --cuda-graph-bs 16 \
        --quantization modelslim \
        --speculative-draft-model-quantization unquant \
        --speculative-algorithm NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4  \
        --moe-a2a-backend deepep --deepep-mode auto

Multi-node deployment

Quantized model GLM-5.2-w8a8 can be deployed on two Atlas 800I A3 nodes.

Modify the IP addresses of the two nodes, then run the same script on both nodes.

# ============================================================
# Before running, update the following variables:
#   IPS: IP addresses of each node in the cluster
#   IP_MASTER: rank 0 node IP address with port
#   MODEL_PATH: path to the 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
# bind cpu
export SGLANG_SET_CPU_AFFINITY=1

unset https_proxy
unset http_proxy
unset HTTPS_PROXY
unset HTTP_PROXY
unset ASCEND_LAUNCH_BLOCKING
# cann
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh

export STREAMS_PER_DEVICE=32
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600
# MTP OVERLAP
export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1

export SGLANG_NPU_USE_MULTI_STREAM=1
export HCCL_BUFFSIZE=1000
export HCCL_OP_EXPANSION_MODE=AIV

# Run command ifconfig on two nodes, find out which inet addr has same IP with your node IP. That is your public interface, which should be added here
export HCCL_SOCKET_IFNAME=<network-interface>
export GLOO_SOCKET_IFNAME=<network-interface>

# DEEPEP
export DEEPEP_NORMAL_LONG_SEQ_ROUND=72
export DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS=1024
export DEEPEP_NORMAL_COMBINE_ENABLE_LONG_SEQ=1
export DEEP_NORMAL_MODE_USE_INT8_QUANT=1


IPS=('<your node1 ip>' '<your node2 ip>')
IP_MASTER="${IPS[0]}:5000"

MODEL_PATH=/path/to/model-weights

LOCAL_HOST1=`hostname -I|awk -F " " '{print$1}'`
LOCAL_HOST2=`hostname -I|awk -F " " '{print$2}'`
for i in "${!IPS[@]}";
do
    if [[ "$LOCAL_HOST1" == "${IPS[$i]}" || "$LOCAL_HOST2" == "${IPS[$i]}" ]];
    then
        echo "${IPS[$i]}"
        python3 -m sglang.launch_server \
        --model-path $MODEL_PATH \
        --attention-backend ascend \
        --device npu \
        --tp-size 32 --nnodes 2 --node-rank $i --dist-init-addr $IP_MASTER \
        --chunked-prefill-size 16384 --max-prefill-tokens 131072 \
        --trust-remote-code \
        --host 127.0.0.1 \
        --mem-fraction-static 0.8 \
        --port 8000 \
        --served-model-name glm-5 \
        --cuda-graph-max-bs 32 \
        --moe-a2a-backend deepep \
        --deepep-mode auto \
        --speculative-draft-model-quantization unquant \
        --speculative-algorithm NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4  \
        --disable-radix-cache
        NODE_RANK=$i
        break
    fi
done

Prefill-decode disaggregation deployment

PD disaggregation splits the prefill and decode stages onto separate nodes, reducing interference and improving throughput for high-concurrency scenarios.

# ============================================================
# Before running, update the following variables:
#   ASCEND_MF_STORE_URL: prefill master IP address with port
#   P_IP: prefill node IP address
#   D_IP: decode node IP address
#   MODEL_PATH: path to the 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

export SGLANG_SET_CPU_AFFINITY=1

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 PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export STREAMS_PER_DEVICE=32
# pd transfer, prefill master IP
export ASCEND_MF_STORE_URL="tcp://<your prefill ip>:24707"
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600

P_IP=('<your prefill ip>')
D_IP=('<your decode ip>')

MODEL_PATH=/path/to/model-weights

export TRANSFORMERS_VERBOSITY=error

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_NORMAL_LONG_SEQ_ROUND=72
        export DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS=1024
        export DEEPEP_NORMAL_COMBINE_ENABLE_LONG_SEQ=1
        export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
        export TASK_QUEUE_ENABLE=2
        export HCCL_SOCKET_IFNAME=<network-interface>
        export GLOO_SOCKET_IFNAME=<network-interface>

        # prefill node
        python -m sglang.launch_server --model-path ${MODEL_PATH}  --disaggregation-mode prefill --host ${P_IP[$i]} \
        --port 8000 --disaggregation-bootstrap-port 8998 --trust-remote-code --nnodes 1 --node-rank $i \
        --tp-size 16 --mem-fraction-static 0.8 --attention-backend ascend --device npu --quantization modelslim \
        --disaggregation-transfer-backend ascend --max-running-requests 64 \
        --served-model-name glm-5 --chunked-prefill-size 524288 --max-prefill-tokens 180000 --moe-a2a-backend deepep --deepep-mode normal \
        --disable-shared-experts-fusion --disable-cuda-graph --dtype bfloat16 \
        --dp-size 4 --enable-dp-attention \
        --load-balance-method round_robin \
        --enable-dp-lm-head --moe-dense-tp 1 \
        --speculative-draft-model-quantization unquant \
        --speculative-algorithm NEXTN --speculative-num-steps 1 --speculative-eagle-topk 1 --speculative-num-draft-tokens 2

        # cp
        #--enable-nsa-prefill-context-parallel \
        #--nsa-prefill-cp-mode in-seq-split \
        #--attn-cp-size 4 \
        NODE_RANK=$i
        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 SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
        export HCCL_BUFFSIZE=650

        export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=32
        export TASK_QUEUE_ENABLE=0

        export HCCL_SOCKET_IFNAME=<network-interface>
        export GLOO_SOCKET_IFNAME=<network-interface>

        export SGLANG_NPU_USE_MULTI_STREAM=1

        python -m sglang.launch_server --model-path ${MODEL_PATH} --disaggregation-mode decode --host ${D_IP[$i]} \
        --port 8003 --trust-remote-code --nnodes 1 --node-rank $i --tp-size 16 --dp-size 16 --ep-size 16 \
        --mem-fraction-static 0.8 --max-running-requests 128 --attention-backend ascend --device npu --quantization modelslim \
        --served-model-name glm-5 --moe-a2a-backend deepep --enable-dp-attention --deepep-mode low_latency \
        --cuda-graph-max-bs 4 --disaggregation-transfer-backend ascend --watchdog-timeout 9000 --context-length 180000 \
        --tokenizer-worker-num 4 --prefill-round-robin-balance --disable-shared-experts-fusion --dtype bfloat16  --load-balance-method round_robin \
        --speculative-draft-model-quantization unquant \
        --speculative-algorithm NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4
        NODE_RANK=$i
        break
    fi
done

Launch the router after the prefill and decode services are ready.

# ============================================================
# Before running, update the following variables:
#   P_MASTER_IP: prefill master IP address
#   D_MASTER_IP: decode master IP address
#   ROUTER_HOST_IP: router node IP address
# ============================================================

P_MASTER_IP="<your prefill ip>"
D_MASTER_IP="<your decode ip>"
ROUTER_HOST_IP="<your router ip>"

python3 -m sglang_router.launch_router \
--pd-disaggregation \
--policy round_robin \
--prefill http://${P_MASTER_IP}:8000 8998 \
--decode http://${D_MASTER_IP}:8003 \
--host ${ROUTER_HOST_IP} \
--port 6688

Functional verification

After the service is started, you can invoke the model by sending a prompt:

# ============================================================
# Before running, update the following variables:
#   HOST: the server host address (e.g., localhost)
#   PORT: the server port number (e.g., 8000)
# ============================================================

curl http://${HOST}:${PORT}/generate \
    -H "Content-Type: application/json" \
    -d '{
        "text": "What is the capital of France?",
        "sampling_params": {
            "max_new_tokens": 64,
            "temperature": 0
        }
    }'

Expected result: an HTTP 200 response with the generated text containing “Paris”.

Once the server prints The server is fired up and ready to roll! in the logs, it is ready to accept requests. For more testing examples (Health Check, Generate, Chat Completions, and port usage guidance), see Testing the Service.

Accuracy evaluation

For accuracy evaluation methods and datasets, see Accuracy Evaluation on Ascend NPU.

Performance

For performance data and benchmark commands, see Performance Testing on Ascend NPU.

Best practices

Best practice configuration reference

GLM-5.2 has no standalone best practice page yet, as 0Day support is still on daily build images and tuning parameters are not finalized. Instead, the Online service deployment section above provides ready-to-use scripts for single-node, multi-node, and PD disaggregation, each embedding the recommended feature combinations and tuning parameters (e.g., DeepEP mode, speculative decoding, overlap schedule).

For the full catalog of optimization features and their parameter and compatibility details, see Optimization on Ascend NPU.

Performance tuning

For the full list of supported features, see Supported features. For detailed optimization guidance, see Optimization on Ascend NPU.

FAQ

For common environment, installation, and general parameter issues, please refer to the Ascend NPU FAQ.