Qwen3-VL-235B-A22B-Instruct

Qwen3-VL-235B-A22B-Instruct#

1 Introduction#

Qwen3-VL-235B-A22B-Instruct is a large-scale sparse MoE vision-language model in the Qwen3-VL family. It is designed for multimodal chat, image understanding, multi-image reasoning, OCR-like visual question answering, and long-context generation.

This document describes the main validation steps for the model, including supported features, prerequisites, installation, single-node online deployment, multi-node deployment, Prefill-Decode (PD) disaggregation, functional verification, accuracy and performance evaluation, performance tuning, and FAQs.

The Qwen3-VL-235B-A22B-Instruct tutorial was introduced in the vLLM-Ascend validation cycle around v0.12.0. Use the current vllm-ascend documentation image placeholder or a later release for the examples below.

2 Supported Features#

Refer to Supported Features List to get the model’s supported feature matrix.

Refer to Feature Guide to get the feature’s configuration.

3 Prerequisites#

3.1 Model Weight#

  • Qwen3-VL-235B-A22B-Instruct (BF16 version): requires 1 Atlas 800 A3 (64G x 16) node or 2 Atlas 800 A2 (64G x 8) nodes. Model Weight.

  • Qwen3-VL-235B-A22B-Instruct-w8a8-QuaRot (quantized version used by single-node validation): requires 1 Atlas 800 A3 (64G x 16) node. Model Weight.

It is recommended to download the model weight to a shared directory across multiple nodes.

3.2 Verify Multi-node Communication (Optional)#

If you want to deploy the model in a multi-node environment, verify the communication environment according to verify multi-node communication environment.

4 Installation#

4.1 Docker Image Installation#

Select an image based on your machine type and start the docker image on your node, refer to using docker.

Start the docker image on each node.

export IMAGE=quay.io/ascend/vllm-ascend:v0.22.1rc1-a3
docker run --rm \
    --name vllm-ascend \
    --shm-size=512g \
    --net=host \
    --device /dev/davinci0 \
    --device /dev/davinci1 \
    --device /dev/davinci2 \
    --device /dev/davinci3 \
    --device /dev/davinci4 \
    --device /dev/davinci5 \
    --device /dev/davinci6 \
    --device /dev/davinci7 \
    --device /dev/davinci8 \
    --device /dev/davinci9 \
    --device /dev/davinci10 \
    --device /dev/davinci11 \
    --device /dev/davinci12 \
    --device /dev/davinci13 \
    --device /dev/davinci14 \
    --device /dev/davinci15 \
    --device /dev/davinci_manager \
    --device /dev/devmm_svm \
    --device /dev/hisi_hdc \
    -v /usr/local/dcmi:/usr/local/dcmi \
    -v /usr/local/Ascend/driver/tools/hccn_tool:/usr/local/Ascend/driver/tools/hccn_tool \
    -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
    -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
    -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
    -v /etc/ascend_install.info:/etc/ascend_install.info \
    -v /etc/hccn.conf:/etc/hccn.conf \
    -it $IMAGE bash

Start the docker image on each node.

export IMAGE=quay.io/ascend/vllm-ascend:v0.22.1rc1
docker run --rm \
    --name vllm-ascend \
    --shm-size=512g \
    --net=host \
    --device /dev/davinci0 \
    --device /dev/davinci1 \
    --device /dev/davinci2 \
    --device /dev/davinci3 \
    --device /dev/davinci4 \
    --device /dev/davinci5 \
    --device /dev/davinci6 \
    --device /dev/davinci7 \
    --device /dev/davinci_manager \
    --device /dev/devmm_svm \
    --device /dev/hisi_hdc \
    -v /usr/local/dcmi:/usr/local/dcmi \
    -v /usr/local/Ascend/driver/tools/hccn_tool:/usr/local/Ascend/driver/tools/hccn_tool \
    -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
    -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
    -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
    -v /etc/ascend_install.info:/etc/ascend_install.info \
    -v /etc/hccn.conf:/etc/hccn.conf \
    -it $IMAGE bash

After a successful docker run, you can verify the running container service by executing the docker ps command.

4.2 Source Code Installation#

If you prefer not to use the Docker image, you can build from source. Install vLLM from source first:

  1. Clone and install vLLM:

    git clone https://github.com/vllm-project/vllm.git
cd vllm
pip install -e .

  2. Clone and install the vLLM-Ascend repository:

    git clone https://github.com/vllm-project/vllm-ascend.git
cd vllm-ascend
pip install -e .

Installation Verification:

pip show vllm vllm-ascend

Expected result: The version information for both packages is displayed, confirming a successful installation.

Note

If deploying a multi-node environment, set up the environment on each node.

For more details, please refer to the Installation Guide.

5 Online Service Deployment#

5.1 Single-Node Online Deployment#

Single-node deployment runs both Prefill and Decode on the same node. The following W8A8 example is suitable for functional validation and image-only online serving on 1 Atlas 800 A3 (64G x 16) node. The W8A8 version needs --quantization ascend.

Run the following script to start online serving on one A3 node:

#!/bin/sh

# Load model from ModelScope to speed up download.
export VLLM_USE_MODELSCOPE=True

# Reduce memory fragmentation and avoid out-of-memory errors.
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True

export HCCL_OP_EXPANSION_MODE="AIV"
export HCCL_BUFFSIZE=1536
export OMP_NUM_THREADS=1
export OMP_PROC_BIND=false
export TASK_QUEUE_ENABLE=1
export VLLM_ASCEND_ENABLE_FLASHCOMM1=1
export VLLM_ASCEND_ENABLE_FUSED_MC2=1
export VLLM_ASCEND_BALANCE_SCHEDULING=1

vllm serve Eco-Tech/Qwen3-VL-235B-A22B-Instruct-w8a8-QuaRot \
  --host 0.0.0.0 \
  --port 8000 \
  --served-model-name qwen3-vl-235b \
  --quantization ascend \
  --data-parallel-size 4 \
  --tensor-parallel-size 4 \
  --enable-expert-parallel \
  --seed 1024 \
  --max-num-seqs 32 \
  --max-model-len 32768 \
  --max-num-batched-tokens 16384 \
  --trust-remote-code \
  --gpu-memory-utilization 0.92 \
  --no-enable-prefix-caching \
  --mm-processor-cache-gb 0 \
  --limit-mm-per-prompt.image 1 \
  --limit-mm-per-prompt.video 0 \
  --compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY","cudagraph_capture_sizes":[1,2,4,8,16,24,32]}'

Common Issues Tip: If you encounter issues, please refer to the Public FAQ for troubleshooting.

Key parameters:

  • --data-parallel-size 4 and --tensor-parallel-size 4 map the 16 NPUs on one A3 node into four DP groups, each with TP4.

  • --enable-expert-parallel enables expert parallelism for MoE layers. Do not mix MoE tensor parallelism and expert parallelism in the same MoE layer.

  • --max-model-len is the maximum input plus output length for a single request. Multimodal inputs consume text tokens and visual tokens, so increase it only when enough KV cache is available.

  • --max-num-seqs is the maximum number of active requests scheduled by each DP group. For performance tests, set --max-num-seqs * --data-parallel-size greater than or equal to the test concurrency.

  • --max-num-batched-tokens is the maximum number of tokens processed in one scheduler step. A larger value can improve prefill efficiency but consumes more activation memory.

  • --gpu-memory-utilization controls how much HBM vLLM can use to calculate KV cache capacity. A higher value increases KV cache size but can trigger OOM if runtime memory is higher than the profile run.

  • --quantization ascend enables Ascend quantization for the W8A8 model. Remove this option when deploying the BF16 model.

  • --limit-mm-per-prompt.image 1 and --limit-mm-per-prompt.video 0 reserve multimodal capacity for one image per request and disable video inputs to save memory.

  • --mm-processor-cache-gb 0 disables the multimodal processor cache. Increase it only when your workload benefits from reused media preprocessing and you have enough host memory.

  • --compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY"}' enables full decode ACLGraph replay to reduce dispatch overhead.

5.3 Multi-Node PD Separation Deployment#

PD disaggregation separates Prefill and Decode into different service groups. Prefill nodes process large prompt chunks, Decode nodes serve token generation, and a proxy forwards requests between them. This mode is suitable for production serving scenarios where prefill and decode resource ratios need to be tuned separately.

We recommend using Mooncake for deployment. Refer to Mooncake for the general PD disaggregation workflow and request forwarding setup.

The following example matches the validated A3 two-node topology for Qwen3-VL-235B-A22B-Instruct-w8a8-QuaRot:

  • 1 Prefill node: 1 Atlas 800 A3 (64G x 16), DP2 + TP8 + EP.

  • 1 Decode node: 1 Atlas 800 A3 (64G x 16), DP4 + TP4 + EP + full decode ACLGraph.

5.3.1 Prefill Node#

Create run_p.sh on the prefill node.

#!/bin/bash

export VLLM_USE_MODELSCOPE=True
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export HCCL_BUFFSIZE=1024
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=1
export HCCL_OP_EXPANSION_MODE="AIV"
export TASK_QUEUE_ENABLE=1

vllm serve Eco-Tech/Qwen3-VL-235B-A22B-Instruct-w8a8-QuaRot \
  --host 0.0.0.0 \
  --port 8080 \
  --quantization ascend \
  --data-parallel-size 2 \
  --data-parallel-size-local 2 \
  --tensor-parallel-size 8 \
  --seed 1024 \
  --served-model-name qwen3-vl-235b \
  --enable-expert-parallel \
  --max-num-seqs 32 \
  --max-model-len 8192 \
  --max-num-batched-tokens 8192 \
  --trust-remote-code \
  --no-enable-prefix-caching \
  --gpu-memory-utilization 0.9 \
  --kv-transfer-config \
  '{"kv_connector":"MooncakeConnectorV1",
    "kv_role":"kv_producer",
    "kv_port":"30000",
    "kv_connector_extra_config":{
      "prefill":{"dp_size":2,"tp_size":8},
      "decode":{"dp_size":4,"tp_size":4}
    }
  }'

Common Issues Tip: If the prefill service is not ready for a long time, check whether the model path is shared, all 16 NPUs are visible, and the Mooncake kv_port is available.

5.3.2 Decode Node#

Create run_d.sh on the decode node.

#!/bin/bash

export VLLM_USE_MODELSCOPE=True
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export HCCL_BUFFSIZE=1024
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=1
export HCCL_OP_EXPANSION_MODE="AIV"
export TASK_QUEUE_ENABLE=1

vllm serve Eco-Tech/Qwen3-VL-235B-A22B-Instruct-w8a8-QuaRot \
  --host 0.0.0.0 \
  --port 8080 \
  --quantization ascend \
  --data-parallel-size 4 \
  --data-parallel-size-local 4 \
  --tensor-parallel-size 4 \
  --seed 1024 \
  --served-model-name qwen3-vl-235b \
  --enable-expert-parallel \
  --max-num-seqs 32 \
  --max-model-len 8192 \
  --max-num-batched-tokens 8192 \
  --trust-remote-code \
  --no-enable-prefix-caching \
  --gpu-memory-utilization 0.9 \
  --compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY"}' \
  --kv-transfer-config \
  '{"kv_connector":"MooncakeConnectorV1",
    "kv_role":"kv_consumer",
    "kv_port":"30200",
    "kv_connector_extra_config":{
      "prefill":{"dp_size":2,"tp_size":8},
      "decode":{"dp_size":4,"tp_size":4}
    }
  }'

Key parameters for PD disaggregation:

  • Prefill uses --data-parallel-size 2, --data-parallel-size-local 2, and --tensor-parallel-size 8.

  • Decode uses --data-parallel-size 4, --data-parallel-size-local 4, and --tensor-parallel-size 4.

  • --max-num-batched-tokens is set to 8192 on both sides in this validation topology. Increase the prefill value only if activation memory is sufficient.

  • --kv-transfer-config sets the Mooncake connector. kv_role is kv_producer on prefill and kv_consumer on decode.

  • kv_connector_extra_config.prefill.dp_size/tp_size and decode.dp_size/tp_size must match the actual global DP and TP layout.

  • --no-enable-prefix-caching disables prefix caching. For PD disaggregation, first validate the service without prefix caching before enabling additional cache features.

  • --compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY"}' is recommended on decode nodes to reduce decode dispatch overhead.

Common Issues Tip: If you encounter issues, please refer to the Public FAQ for troubleshooting.

Service Verification:

curl http://<server_ip>:<port>/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "qwen3-vl-235b",
        "messages": [
            {
                "role": "user",
                "content": "Who are you?"
            }
        ],
        "max_tokens": 256,
        "temperature": 0
    }'

Expected Result:

The service returns HTTP 200 OK with a JSON response containing the choices field.

6 Functional Verification#

After the server is started, send a request to verify basic multimodal functionality. For single-node and MP deployment, use the API endpoint on node 0. For PD disaggregation, use the proxy endpoint from the Mooncake deployment guide.

curl http://<server_ip>:<port>/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3-vl-235b",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": [
        {"type": "image_url", "image_url": {"url": "https://modelscope.oss-cn-beijing.aliyuncs.com/resource/qwen.png"}},
        {"type": "text", "text": "What is the text in the illustration?"}
      ]}
    ],
    "max_completion_tokens": 100,
    "temperature": 0
  }'

Expected result: the HTTP status is 200 and the JSON response contains a choices field with generated text, for example text similar to TONGYI Qwen.

7 Accuracy Evaluation#

Using AISBench#

  1. Refer to Using AISBench for details.

  2. After execution, you can get the result.

dataset

version

metric

mode

vllm-api-general-chat

textvqa-lite

-

accuracy

gen

83

aime2024

-

accuracy

gen

93

8 Performance Evaluation#

8.1 Using AISBench#

Refer to Using AISBench for performance evaluation for details. For multimodal performance, use a dataset with image payloads, such as TextVQA-style requests, instead of random text-only prompts.

8.2 Using vLLM Benchmark#

Run performance evaluation of Qwen3-VL-235B-A22B-Instruct as an example. Refer to vLLM benchmark for more details.

There are three vllm bench subcommands:

  • latency: benchmark the latency of a single batch of requests.

  • serve: benchmark online serving throughput.

  • throughput: benchmark offline inference throughput.

Take serve as an example:

export VLLM_USE_MODELSCOPE=True

vllm bench serve \
  --model Eco-Tech/Qwen3-VL-235B-A22B-Instruct-w8a8-QuaRot \
  --served-model-name qwen3-vl-235b \
  --dataset-name random \
  --random-input 200 \
  --num-prompts 200 \
  --request-rate 1 \
  --save-result \
  --result-dir ./

After several minutes, you can get the performance evaluation result. This random benchmark is useful for serving pipeline validation; use AISBench or a custom multimodal dataset for image-token performance.

9 Performance Tuning#

9.2 Tuning Guidelines#

9.2.1 General Tuning Reference#

Please refer to the Public Performance Tuning Documentation for tuning methods.

Please refer to the Feature Guide for detailed feature descriptions.

9.3 Model-Specific Optimizations#

Optimization

Enablement

Benefit

Notes

Multimodal prompt limits

--limit-mm-per-prompt.image, --limit-mm-per-prompt.video

Avoids reserving memory for unused media types.

Disable video for image-only serving.

Multimodal processor cache

--mm-processor-cache-gb

Caches processed media features when repeated media appears.

Set to 0 for memory-constrained validation.

Full decode ACLGraph

--compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY"}'

Reduces operator dispatch overhead and stabilizes decode performance.

Recommended for decode-heavy serving.

FlashComm1

VLLM_ASCEND_ENABLE_FLASHCOMM1=1 or --additional-config '{"enable_flashcomm1":true}'

Reduces communication overhead in large TP and high-concurrency scenarios.

May not help low-concurrency workloads.

Fused MC2

VLLM_ASCEND_ENABLE_FUSED_MC2=1

Enables MoE fused operators to improve MoE efficiency.

Compare with disabled state if accuracy or performance regresses.

Prefix caching

--enable-prefix-caching

Improves repeated-prefix workloads.

Validate HBM usage first. For PD, start with prefix caching disabled.

Asynchronous scheduling

--async-scheduling

Can improve high-concurrency throughput.

Disable and compare for latency-sensitive workloads.

PD disaggregation

--kv-transfer-config

Separates prefill and decode resources.

Ensure producer/consumer DP and TP sizes match the actual topology.

10 FAQ#

For common environment, installation, and general parameter issues, refer to Public FAQs. This section only covers model-specific issues for Qwen3-VL-235B-A22B-Instruct.

Q1: Why does the service report OOM during startup or soon after accepting requests?#

Phenomenon: The service fails during profile run, or it starts successfully but reports OOM when real traffic arrives.

Cause: Qwen3-VL-235B-A22B-Instruct has high weight, KV cache, and multimodal preprocessing memory requirements. Large --max-model-len, large --max-num-seqs, large --max-num-batched-tokens, high image resolution, too many images per prompt, or high --gpu-memory-utilization can leave insufficient HBM headroom.

Solution: Use the W8A8 model with --quantization ascend when possible, lower --max-model-len, lower --max-num-seqs, lower --max-num-batched-tokens, lower image/video limits, or reduce --gpu-memory-utilization. Keep PYTORCH_NPU_ALLOC_CONF=expandable_segments:True.

Q2: Why does multi-node MP deployment hang during initialization?#

Phenomenon: One node waits for other ranks, HCCL initialization times out, or the headless node exits.

Cause: Network interface names, IP addresses, DP ranks, or RPC ports are inconsistent across nodes.

Solution: Verify multi-node communication first. Ensure HCCL_IF_IP, GLOO_SOCKET_IFNAME, TP_SOCKET_IFNAME, and HCCL_SOCKET_IFNAME match the selected NIC. Ensure all nodes use the same --data-parallel-rpc-port, non-master nodes use --headless, and --data-parallel-start-rank does not overlap.

Q3: Why is video disabled in the image-only examples?#

Phenomenon: The service reserves more memory than expected, or startup OOM occurs even when requests only contain images.

Cause: Allowing video inputs can reserve memory for long visual embeddings and preprocessing paths that are not needed by image-only workloads.

Solution: Use --limit-mm-per-prompt.video 0 for image-only serving. Enable video only when your workload needs it, and lower --max-model-len or request concurrency if needed.

Q4: Why does enabling prefix caching not improve performance?#

Phenomenon: Prefix caching is enabled, but throughput or latency does not improve.

Cause: Prefix caching only helps when requests share reusable prefixes. Random prompts, unique images, or low cache hit rates may add memory pressure without visible gains.

Solution: Enable prefix caching for repeated-prefix workloads. For random benchmark datasets, memory-constrained long-context workloads, or PD validation, compare with --no-enable-prefix-caching.

Q5: Why does PD disaggregation fail to transfer KV cache?#

Phenomenon: Requests reach the proxy or prefill service, but decode nodes do not produce output or report KV transfer errors.

Cause: The Mooncake connector ports, producer/consumer roles, or kv_connector_extra_config DP/TP sizes do not match the actual topology.

Solution: Check kv_role, kv_port, and the prefill/decode DP/TP sizes on all nodes. Start from the validated topology in Section 5.4, then change one dimension at a time.