Qwen3.5 四款小尺寸模型开源:昇腾适配已到位,AtomGit AI 开放体验
近日,千问(Qwen)正式开源 Qwen3.5 小尺寸模型系列,包括:Qwen3.5-0.8B、Qwen3.5-2B、Qwen3.5-4B、Qwen3.5-9B。接下来我们以 Qwen3.5-2B 为例,带大家一步步完成在 Ascend 上的基于 vLLM 和SGLang 的部署流程,其他几款模型的部署方式基本一致,可类比操作。在中小参数规模下实现较高能力上限,综合性能表现接近更大规模模型,媲美
近日,千问(Qwen)正式开源 Qwen3.5 小尺寸模型系列,包括:Qwen3.5-0.8B、Qwen3.5-2B、Qwen3.5-4B、Qwen3.5-9B。目前,昇腾生态已完成对 Qwen3.5 小尺寸系列四款模型的适配支持,相关模型文件与权重已同步上线 AtomGit AI,开发者们可直接获取并进行部署测试啦~
🔗 vLLM Ascend 部署:
0.8B ➡️ https://ai.atomgit.com/vLLM_Ascend/Qwen3.5-0.8B
2B ➡️ https://ai.atomgit.com/vLLM_Ascend/Qwen3.5-2B
4B ➡️ https://ai.atomgit.com/vLLM_Ascend/Qwen3.5-4B
9B ➡️ https://ai.atomgit.com/vLLM_Ascend/Qwen3.5-9B
🔗 SGLang 部署:
0.8B ➡️ https://ai.atomgit.com/SGLangAscend/Qwen3.5-0.8B
2B ➡️ https://ai.atomgit.com/SGLangAscend/Qwen3.5-2B
4B ➡️ https://ai.atomgit.com/SGLangAscend/Qwen3.5-4B
9B ➡️ https://ai.atomgit.com/SGLangAscend/Qwen3.5-9B
小尺寸,不等于低能力
Qwen3.5 小模型系列继承了家族的统一架构与训练体系,采用原生多模态训练与最新模型结构优化,在轻量体积下依然保持了不错的综合能力。
📌 0.8B / 2B:轻量化优先,面向端侧部署
特点:参数规模小,占用资源低,推理延迟控制表现突出,适合算力与显存受限环境。
适用场景:移动端设备、IoT 边缘节点、本地嵌入式场景,以及对实时响应要求较高的低时延交互应用。
📌 4B:轻量级 Agent 与多模态应用的平衡选择
特点:在保持较低资源消耗的同时,具备较完整的多模态理解与推理能力,适合作为中等复杂度任务的基础模型。
适用场景:轻量级智能体(Agent)构建、多模态交互应用、对推理能力与算力成本需平衡控制的业务场景。
📌 9B:紧凑规模下的高能力密度模型
特点:在中小参数规模下实现较高能力上限,综合性能表现接近更大规模模型,媲美 gpt-oss-120B,具备较强的泛化与复杂任务处理能力。
适用场景:适合需要较高智力水平但受限显存资源的服务器端部署,是性价比极高的通用模型选择。
接下来我们以 Qwen3.5-2B 为例,带大家一步步完成在 Ascend 上的基于 vLLM 和SGLang 的部署流程,其他几款模型的部署方式基本一致,可类比操作。
Qwen3.5-2B 基于 vLLM 部署流程
环境准备
模型权重
-
Qwen3.5-2B(BF16 版本):https://ai.atomgit.com/hf_mirrors/Qwen/Qwen3.5-2B
注: 建议将模型权重下载至多节点共享目录,例如 /root/.cache/。
安装
1️⃣ 官方 Docker 镜像
您可以通过镜像链接下载镜像压缩包来进行部署,具体流程如下:
# 使用docker加载下载的镜像压缩包
# 根据您的环境更新要加载的vllm-ascend镜像压缩包名称,以下以A3 arm为例:
docker load -i Vllm-ascend-Qwen3_5-A3-Ubuntu-v0.tar
# 根据您的设备更新 --device(Atlas A3:/dev/davinci[0-15])。
# 注意:您需要提前将权重下载至 /root/.cache。
# 更新 vllm-ascend 镜像,并配置对应的Image名
export IMAGE=vllm-ascend:qwen3_5-v0-a3
export NAME=vllm-ascend
# 使用定义的变量运行容器
# 注意:若使用 Docker 桥接网络,请提前开放可供多节点通信的端口
docker run --rm \
--name $NAME \
--net=host \
--shm-size=100g \
--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 /root/.cache:/root/.cache \
-it $IMAGE bash
2️⃣ 源码构建
如果您不希望使用上述 Docker 镜像,也可通过源码完整构建:
-
保证你的环境成功安装了 CANN 8.5.0
-
从源码安装 vllm-ascend ,请参考安装指南。
从源码安装 vllm-ascend 后,您需要将 vllm、vllm-ascend、transformers 升级至主分支:
# 升级 vllm
git clone https://github.com/vllm-project/vllm.git
cd vllm
git checkout a75a5b54c7f76bc2e15d3025d6
git fetch origin pull/34521/head:pr-34521
git merge pr-34521
VLLM_TARGET_DEVICE=empty pip install -v .
# 升级 vllm-ascend
pip uninstall vllm-ascend -y
git clone https://github.com/vllm-project/vllm-ascend.git
cd vllm-ascend
git checkout c63b7a11888e9e1caeeff8
git fetch origin pull/6742/head:pr-6742
git merge pr-6742
pip install -v .
# 重新安装 transformers
git clone https://github.com/huggingface/transformers.git
cd transformers
git reset --hard fc9137225880a9d03f130634c20f9dbe36a7b8bf
pip install .
如需部署多节点环境,您需要在每个节点上分别完成环境配置。
部署
单节点部署
A3 系列
执行以下脚本进行在线推理。
export PYTORCH_NPU_ALLOC_CONF="expandable_segments:True"
export HCCL_OP_EXPANSION_MODE="AIV"
export HCCL_BUFFSIZE=1024
export OMP_NUM_THREADS=1
export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD
export TASK_QUEUE_ENABLE=1
vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/Qwen3.5-2B/ \
--served-model-name "qwen3.5" \
--host 0.0.0.0 \
--port 8010 \
--data-parallel-size 1 \
--tensor-parallel-size 4 \
--max-model-len 5000 \
--max-num-batched-tokens 16384 \
--max-num-seqs 128 \
--gpu-memory-utilization 0.94 \
--trust-remote-code \
--async-scheduling \
--allowed-local-media-path / \
--mm-processor-cache-gb 0 \
--enforce-eager \
--additional-config '{"enable_cpu_binding":true, "multistream_overlap_shared_expert": true}'
执行以下脚本向模型发送一条请求:
curl http://localhost:8000/v1/completions \
-H "Content-Type: application/json" \
-d '{
"prompt": "The future of AI is",
"path": "/path/to/model/Qwen3.5-2B/",
"max_tokens": 100,
"temperature": 0
}'
执行结束后,您可以看到模型回答如下:
Prompt: 'The future of AI is', Generated text: ' not just about building smarter machines, but about creating systems that can collaborate with humans in meaningful, ethical, and sustainable ways. As AI continues to evolve, it will increasingly shape how we live, work, and interact — and the decisions we make today will determine whether this future is one of shared prosperity or deepening inequality.\n\nThe rise of generative AI, for example, has already begun to transform creative industries, education, and scientific research. Tools like ChatGPT, Midjourney, and'
也可执行以下脚本向模型发送一条多模态请求:
curl http://localhost:8000/v1/completions \
-H "Content-Type: application/json" \
-d '{
"model": "qwen3.5",
"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 illustrate?"}
]}
]
}'
执行结束后,您可以看到模型回答如下:
{"id":"chatcmpl-9dab99d55addd8c0","object":"chat.completion","created":1771060145,"model":"qwen3.5","choices":[{"index":0,"message":{"role":"assistant","content":"TONGYI Qwen","refusal":null,"annotations":null,"audio":null,"function_call":null,"tool_calls":[],"reasoning":null},"logprobs":null,"finish_reason":"stop","stop_reason":null,"token_ids":null}],"service_tier":null,"system_fingerprint":null,"usage":{"prompt_tokens":112,"total_tokens":119,"completion_tokens":7,"prompt_tokens_details":null},"prompt_logprobs":null,"prompt_token_ids":null,"kv_transfer_params":null}
Qwen3.5-2B 基于 SGLang 部署流程
环境准备
安装
NPU 运行时环境所需的依赖已集成到 Docker 镜像中,并上传至 quay.io 平台,用户可直接拉取该镜像。
#Atlas
800 A3
swr.cn-southwest-2.myhuaweicloud.com/base_image/dockerhub/lmsysorg/sglang:main-cann8.5.0-a3
#Atlas
800 A2
swr.cn-southwest-2.myhuaweicloud.com/base_image/dockerhub/lmsysorg/sglang:main-cann8.5.0-910b
#start
container
docker run -itd --shm-size=16g --privileged=true --name ${NAME} \
--privileged=true --net=host \
-v /var/queue_schedule:/var/queue_schedule \
-v /etc/ascend_install.info:/etc/ascend_install.info \
-v /usr/local/sbin:/usr/local/sbin \
-v /usr/local/Ascend/driver:/usr/local/Ascend/driver \
-v /usr/local/Ascend/firmware:/usr/local/Ascend/firmware \
--device=/dev/davinci0:/dev/davinci0 \
--device=/dev/davinci1:/dev/davinci1 \
--device=/dev/davinci2:/dev/davinci2 \
--device=/dev/davinci3:/dev/davinci3 \
--device=/dev/davinci4:/dev/davinci4 \
--device=/dev/davinci5:/dev/davinci5 \
--device=/dev/davinci6:/dev/davinci6 \
--device=/dev/davinci7:/dev/davinci7 \
--device=/dev/davinci8:/dev/davinci8 \
--device=/dev/davinci9:/dev/davinci9 \
--device=/dev/davinci10:/dev/davinci10 \
--device=/dev/davinci11:/dev/davinci11 \
--device=/dev/davinci12:/dev/davinci12 \
--device=/dev/davinci13:/dev/davinci13 \
--device=/dev/davinci14:/dev/davinci14 \
--device=/dev/davinci15:/dev/davinci15 \
--device=/dev/davinci_manager:/dev/davinci_manager \
--device=/dev/hisi_hdc:/dev/hisi_hdc \
--entrypoint=bash \
quay.io/ascend/sglang:${tag}
权重下载
🔗 Qwen2.5-2B:https://ai.atomgit.com/hf_mirrors/Qwen/Qwen3.5-2B
部署
单节点部署
执行以下脚本进行在线推理。
# 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 HCCL_BUFFSIZE=1000
export HCCL_OP_EXPANSION_MODE=AIV
export HCCL_SOCKET_IFNAME=lo
export GLOO_SOCKET_IFNAME=lo
python3 -m sglang.launch_server \
--model-path $MODEL_PATH \
--attention-backend ascend \
--device npu \
--tp-size 1 \
--chunked-prefill-size -1 --max-prefill-tokens 120000 \
--disable-radix-cache \
--trust-remote-code \
--host 127.0.0.1 \
--mem-fraction-static 0.8 \
--port 8000 \
--cuda-graph-bs 16 \
--enable-multimodal \
--mm-attention-backend ascend_attn
发送请求测试
curl --location http://127.0.0.1:8000/v1/chat/completions --header 'Content-Type: application/json' --data '{
"model": "qwen3.5",
"messages": [
{
"role": "user",
"content": [
{
"type": "image_url",
"image_url": {"url": "/image_path/qwen.png"}
},
{"type": "text", "text": "What is the text in the illustrate?"}
]
}
]
}'
结果返回如下
{"id":"cdcd6d14645846e69cc486554f198154","object":"chat.completion","created":1772098465,"model":"qwen3.5","choices":[{"index":0,"message":{"role":"assistant","content":"The user is asking about the text present in the image. I will analyze the image to identify the text.\n</think>\n\nThe text in the image is \"TONGyi Qwen\".","reasoning_content":null,"tool_calls":null},"logprobs":null,"finish_reason":"stop","matched_stop":248044}],"usage":{"prompt_tokens":98,"total_tokens":138,"completion_tokens":40,"prompt_tokens_details":null,"reasoning_tokens":0},"metadata":{"weight_version":"default"}}
声明
当前为尝鲜版本,我们还在持续优化性能,给大家带来更好的体验。
以上内容及代码仓中提到的数据集和模型仅作示例使用,仅供非商业用途学习与参考。如果您基于示例使用这些数据集和模型,请注意遵守对应的开源协议(License),避免产生相关纠纷。
如果您在使用过程中遇到任何问题(包括功能、合规等),欢迎在代码仓提交 Issue,我们会及时查看并回复~
更多推荐





所有评论(0)