Qwen3-235B-A22B

Introduction

Qwen3-235B-A22B is a Mixture-of-Experts (MoE) large language model developed by Alibaba, featuring 235B total parameters with 22B active parameters. It employs Grouped-Query Attention (GQA) and Qwen3MoE architecture, with support for EAGLE3 speculative decoding for accelerated inference. The model excels at instruction following, logical reasoning, text comprehension, mathematics, science, coding, and tool usage, available in both standard and thinking/reasoning-enhanced editions.

This document demonstrates the deployment of Qwen3-235B-A22B on Ascend NPUs using SGLang, including single-node PD mixed mode, multi-node PD disaggregation mode, 256k long-sequence inference, Prefill Context Parallel, feature configuration, and performance optimization.

This document is validated and written based on SGLang v0.5.13. The current model (Qwen3-235B-A22B) is fully supported in this version. To use the latest features (e.g., PD disaggregation, speculative decoding), it is recommended to use v0.5.13 or a later version.

Supported features

Feature	Example usage
Tensor Parallelism	--tp-size 16
Data Parallelism	--dp-size 16
Expert Parallelism	--ep-size 16 \ --moe-a2a-backend ascend_fuseep
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 94208
NPU Graph	enabled by default; disable with --disable-cuda-graph; control range via --cuda-graph-bs or --cuda-graph-max-bs-decode; e.g., --cuda-graph-bs 1 2 4 8 16 20 24 26 27
Speculative Decoding	--speculative-algorithm EAGLE3 \ --speculative-draft-model-path /path/to/draft-model-weights \ --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
Context Parallelism	--enable-prefill-context-parallel \ --attn-cp-size 2 \ --moe-dp-size 2

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.

• Qwen3-235B-A22B-Instruct-2507 (BF16, 470.21GB)
• Qwen3-235B-A22B-W8A8 (W8A8 quantized version, 236.80GB)
• Qwen3-235B-A22B-Eagle3 (EAGLE3 draft model for speculative decoding)

The BF16 variant (470.21GB) 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.

The W8A8 variant (236.80GB) can be deployed on 8 × 64GB (--tp-size 8), which corresponds to one full A2 node or 8 dies on A3 (4 cards).

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

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

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.

Both stable releases and daily builds are available. The following command is based on the stable release tag. For details, see Docker image versions.

Atlas 800I A3

Command

docker pull quay.io/ascend/sglang:v0.5.13.post1-cann9.0.0-a3

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 \
quay.io/ascend/sglang:v0.5.13.post1-cann9.0.0-a3

Atlas 800I A2

Command

docker pull quay.io/ascend/sglang:v0.5.13.post1-cann9.0.0-910b

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/davinci_manager:/dev/davinci_manager \
--device=/dev/hisi_hdc:/dev/hisi_hdc \
--entrypoint=bash \
quay.io/ascend/sglang:v0.5.13.post1-cann9.0.0-910b

• 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 online deployment

Single-node deployment completes both prefill and decode within the same node (PD mixed mode), suitable for scenarios with limited hardware resources. This scenario is already covered in the best practice. For the complete, optimized deployment commands and benchmark data, see Qwen3-235B-A22B Best Practice — PD Mixed On A3.

Multi-node PD disaggregation deployment

256K long-sequence PD disaggregation on 2 x Atlas 800I A3 (without CP)

This configuration uses PD disaggregation for 256K long-sequence inference on 2 x Atlas 800I A3 with context parallel disabled. The following command is based on the W8A8 quantized model.

1. Set the shared environment variables on both prefill and decode nodes:

Shared environment

#============================================================
# Before running, update the following variables:
#   MODEL_PATH: path to the model weights directory
#   PREFILL_HOST_IP: prefill node IP address
#   NETWORK_IFACE: network interface name (use ifconfig to find)
#============================================================

export ASCEND_USE_FIA=1
export SGLANG_SET_CPU_AFFINITY=1
export ASCEND_MF_STORE_URL="tcp://<PREFILL_HOST_IP>:12345"
export HCCL_SOCKET_IFNAME=<NETWORK_IFACE>
export GLOO_SOCKET_IFNAME=<NETWORK_IFACE>

MODEL_PATH=/path/to/model-weights

2. Run on the prefill node:

Prefill node

#============================================================
# Before running, update the following variable:
#   PREFILL_HOST_IP: prefill node IP address
#============================================================

export ASCEND_LAUNCH_BLOCKING=1
export HCCL_BUFFSIZE=1500
export DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS=1024
export DEEPEP_NORMAL_LONG_SEQ_ROUND=128
export DEEPEP_NORMAL_COMBINE_ENABLE_LONG_SEQ=1
export DEEP_NORMAL_MODE_USE_INT8_QUANT=1

python3 -m sglang.launch_server \
    --model-path ${MODEL_PATH} \
    --disaggregation-mode prefill \
    --disaggregation-transfer-backend ascend \
    --disaggregation-bootstrap-port 8995 \
    --attention-backend ascend \
    --disable-radix-cache \
    --chunked-prefill-size -1 \
    --skip-server-warmup \
    --device npu \
    --quantization modelslim \
    --tp-size 16 \
    --mem-fraction-static 0.45 \
    --max-running-requests 1 \
    --host <PREFILL_HOST_IP> \
    --port 8000 \
    --dist-init-addr <PREFILL_HOST_IP>:5000 \
    --nnodes 1 \
    --node-rank 0 \
    --moe-a2a-backend deepep \
    --deepep-mode normal

3. Run on the decode node:

Decode node

#============================================================
# Before running, update the following variable:
#   DECODE_HOST_IP: decode node IP address
#============================================================

export HCCL_BUFFSIZE=4000
export DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS=4096
export DEEPEP_NORMAL_LONG_SEQ_ROUND=16

python3 -m sglang.launch_server \
    --model-path ${MODEL_PATH} \
    --disaggregation-mode decode \
    --disaggregation-transfer-backend ascend \
    --attention-backend ascend \
    --mem-fraction-static 0.8 \
    --disable-cuda-graph \
    --device npu \
    --quantization modelslim \
    --disable-radix-cache \
    --chunked-prefill-size 8192 \
    --skip-server-warmup \
    --tp-size 16 \
    --max-running-requests 1 \
    --host <DECODE_HOST_IP> \
    --port 8232 \
    --moe-a2a-backend deepep \
    --deepep-mode low_latency \
    --disable-overlap-schedule

4. Launch the SGLang Router (on any reachable node):

Router

#============================================================
# Before running, update the following variables:
#   PREFILL_HOST_IP: prefill node IP address
#   DECODE_HOST_IP: decode node IP address
#   ROUTER_HOST_IP: router node IP address
#============================================================

python3 -m sglang_router.launch_router \
    --pd-disaggregation \
    --policy cache_aware \
    --prefill http://<PREFILL_HOST_IP>:8000 8995 \
    --decode http://<DECODE_HOST_IP>:8232 \
    --host <ROUTER_HOST_IP> \
    --port 6689 \
    --prometheus-port 29010

Prefill Context Parallel (PCP) on 2 x Atlas 800I A3

This configuration enables Prefill Context Parallel (--enable-prefill-context-parallel) to split the context across CP ranks during prefill, reducing per-device memory pressure and improving TTFT for long sequences. PD disaggregation is required. The following command is based on the W8A8 quantized model.

Constraints:

• Prefill side must set --max-running-requests 1 (PCP only supports batch_size=1)
• --attn-cp-size must evenly divide --tp-size; each CP rank occupies tp_size / cp_size NPUs

1. Run on the prefill node:

Prefill node

#============================================================
# Before running, update the following variables:
#   MODEL_PATH: path to the model weights directory
#   PREFILL_HOST_IP: prefill node IP address
#============================================================

export SGLANG_SET_CPU_AFFINITY=1
export ASCEND_MF_STORE_URL="tcp://<PREFILL_HOST_IP>:23456"
export ASCEND_USE_FIA=True

python3 -m sglang.launch_server \
    --model-path ${MODEL_PATH} \
    --trust-remote-code \
    --disaggregation-mode prefill \
    --disaggregation-transfer-backend ascend \
    --disaggregation-bootstrap-port 8995 \
    --quantization modelslim \
    --attention-backend ascend \
    --skip-server-warmup \
    --mem-fraction-static 0.7 \
    --chunked-prefill-size 32768 \
    --device npu \
    --base-gpu-id 0 \
    --tp-size 16 \
    --enable-prefill-context-parallel \
    --attn-cp-size 2 \
    --moe-dp-size 2 \
    --max-running-requests 1 \
    --host <PREFILL_HOST_IP> \
    --port 8000 \
    --nnodes 1 \
    --node-rank 0 \
    --dist-init-addr <PREFILL_HOST_IP>:6688

Key parameters for PCP:

Parameter	Value	Description
--enable-prefill-context-parallel	flag	Enable PCP feature
--attn-cp-size	2	Split context across 2 CP ranks (each rank handles half the sequence)
--moe-dp-size	2	MoE DP size, should match --attn-cp-size
--max-running-requests	1	Required by PCP (batch_size=1 constraint)

2. Run on the decode node:

Decode node

#============================================================
# Before running, update the following variables:
#   MODEL_PATH: path to the model weights directory
#   DECODE_HOST_IP: decode node IP address
#   PREFILL_HOST_IP: prefill node IP address (for ASCEND_MF_STORE_URL)
#============================================================

export ASCEND_MF_STORE_URL="tcp://<PREFILL_HOST_IP>:23456"
export ASCEND_USE_FIA=True

python3 -m sglang.launch_server \
    --model-path ${MODEL_PATH} \
    --trust-remote-code \
    --disaggregation-mode decode \
    --disaggregation-transfer-backend ascend \
    --quantization modelslim \
    --attention-backend ascend \
    --disable-radix-cache \
    --disable-cuda-graph \
    --mem-fraction-static 0.7 \
    --chunked-prefill-size 32768 \
    --skip-server-warmup \
    --device npu \
    --base-gpu-id 0 \
    --tp-size 8 \
    --max-running-requests 32 \
    --host <DECODE_HOST_IP> \
    --port 8001 \
    --nnodes 1 \
    --node-rank 0 \
    --dist-init-addr <DECODE_HOST_IP>:6688

ASCEND_MF_STORE_URL on both nodes must point to the same KV store (typically the prefill node IP). ASCEND_USE_FIA=True enables fast interconnect aggregation for KV transfer. PCP is a prefill-only feature; the decode side needs no CP-related flags.

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., 6689)
# ============================================================

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

For complete optimal configurations with deployment scripts and benchmark commands, see the Qwen3-235B-A22B Best Practice page.

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.