跳到主内容
@wquguru
精选88Microsoft Foundry 博客(RSS)产品发布/更新

Microsoft Foundry 推出 Multi-LoRA 共享部署功能

From fine-tuning to shared serving: Multi-LoRA deployment on Foundry Managed Compute

原文
发到 X
推荐理由

Foundry 推出的 Multi-LoRA 共享部署方案直击多适配器运维痛点,提供完整的技术实现细节与成本优势,适合正在构建垂直领域 Agent 的工程团队参考落地。

We're introducing Multi-LoRA adapter deployment on Managed Compute in Microsoft Foundry, now in preview. Multi-LoRA lets you deploy multiple compatible fine-tuned LoRA adapters on a shared base-model deployment, so you can serve specialized model experiences without provisioning separate compute for each adapter.

我们宣布在 Microsoft Foundry 的托管计算(Managed Compute)中引入 Multi-LoRA 适配器部署功能,目前处于预览阶段。Multi-LoRA 允许您在共享的基础模型部署上部署多个兼容的微调 LoRA 适配器,从而无需为每个适配器单独配置计算资源即可提供专业化的模型体验。

The preview initially supports Muse-Glimmer-30B, Qwen3.8-27B, Qwen3.6-35B-A3B, and GPT-OSS-120B.

预览版最初支持 Muse-Glimmer-30B、Qwen3.8-27B、Qwen3.6-35B-A3B 和 GPT-OSS-120B。

Managed Compute is Microsoft Foundry’s managed platform for deploying and serving open models on dedicated GPU capacity. Foundry manages the underlying infrastructure and serving runtime, so teams can focus on their models and applications rather than operating GPU infrastructure. Multi-LoRA builds on this foundation, allowing compatible adapters for supported base models to share a base-model deployment. Learn more in the Managed Compute announcement and Managed Compute overview.

托管计算是 Microsoft Foundry 用于在专用 GPU 容量上部署和服务开源模型的托管平台。Foundry 管理底层基础设施和服务运行时,使团队能够专注于模型和应用,而非操作 GPU 基础设施。Multi-LoRA 在此基础上构建,允许受支持基础模型的兼容适配器共享基础模型部署。更多信息请参阅托管计算公告和托管计算概述。

This brings the output of Foundry fine-tuning closer to the applications that use it. Whether you're building a support assistant or adapting a model to financial terminology and reporting formats, you can move from a registered adapter to a callable deployment while reusing the serving foundation already in place.

这使得 Foundry 微调的输出更接近使用它的应用程序。无论您是在构建支持助手,还是将模型适应于金融术语和报告格式,都可以从注册的适配器过渡到可调用的部署,同时复用已就位的服务基础架构。

Specialization without a separate serving stack

无需单独的推理栈即可实现专业化

Fine-tuning gives teams a way to shape a model for different tasks. The operational challenge comes afterward: serving each specialization without treating it as another infrastructure project. A separate deployment of every model variant can mean repeating the common base model's serving setup, even when the variants share the same foundation.

微调为团队提供了一种塑造模型以适应不同任务的方法。随之而来的运营挑战在于:在不将其视为另一个基础设施项目的前提下服务每种专业化能力。如果为每个模型变体进行单独部署,即使这些变体共享相同的基础,也可能需要重复设置公共基础模型的服务环境。

Multi-LoRA separates the identity of a specialization from ownership of the compute underneath it. Platform teams manage the shared target deployment; application teams select the adapter that fits their task. The benefit is a simpler serving arrangement for compatible specializations, not a new infrastructure stack for each one.

Multi-LoRA 将专业化的身份与其底层计算的所有权分离开来。平台团队管理共享的目标部署;应用团队选择适合其任务的适配器。其优势在于为兼容的专业化能力提供更简单的服务安排,而不是为每种能力创建新的基础设施栈。

  • Keep specialized behaviors distinct. Each adapter deployment has its own name, making the application's choice explicit.
  • Reuse the serving foundation. Compatible adapters share the target's base model and compute rather than provisioning accelerators of their own.
  • Keep application calls familiar. Select an adapter through the Foundry account's chat-completions endpoint using its deployment name.
  • 保持专业化行为的独立性。每个适配器部署都有各自的名称,从而使应用程序的选择明确可见。
  • 复用服务基础架构。兼容的适配器共享目标的基础模型和计算资源,而无需自行配置加速器。
  • 保持应用程序调用的熟悉度。通过 Foundry 账户的 chat-completions 端点,使用部署名称来选择适配器。

One base model, multiple adapter deployments

一个基础模型,多个适配器部署

A LoRA adapter contains a lightweight set of learned weight updates that specializes a compatible base model. The adapter doesn't run independently. A Managed Compute target hosts the base model and serving capacity, while individual adapter deployments connect registered fine-tuning outputs to that target.

LoRA 适配器包含一组轻量级的已学习权重更新,用于使兼容的基础模型专业化。该适配器不能独立运行。托管计算目标托管基础模型和服务能力,而各个适配器的部署将注册的微调输出连接到该目标。

For example, suppose you fine-tune the same Qwen base model and version for two scenarios: customer support and finance.

例如,假设您针对两个场景对同一 Qwen 基础模型和版本进行微调:客户支持和金融。

You could deploy both adapters against a shared target deployment qwen-prod:

您可以将这两个适配器部署到共享的目标部署 qwen-prod 上:

  • qwen-prod: the shared Managed Compute deployment hosting the base model and LoRA-capable serving runtime.
  • qwen-support: an adapter deployment referencing the registered support fine-tuning output and targeting qwen-prod.
  • qwen-finance: an adapter deployment referencing the registered finance fine-tuning output and targeting qwen-prod.
  • qwen-prod:托管基础模型和支持 LoRA 的服务运行时的共享托管计算部署。
  • qwen-support:引用注册的支持微调输出并指向 qwen-prod 的适配器部署。
  • qwen-finance:引用的注册金融微调输出并指向 qwen-prod 的适配器部署。

Applications invoke qwen-support or qwen-finance through the respective adapter deployment’s inference route. The service resolves the shared parent internally; applications do not select an adapter by changing requests to the existing qwen-prod base-model route.

应用程序通过各自适配器部署的推理路由调用 qwen-support 或 qwen-finance。服务在内部解析共享父级;应用程序不会通过将请求更改为现有的 qwen-prod 基础模型路由来选择适配器。

Each adapter deployment can be replaced with a newer registered output or deleted independently. This independence does not mean dedicated compute or isolated serving capacity: both adapters use qwen-prod’s shared resources.

每个适配器部署都可以替换为更新的注册输出或删除。这种独立性并不意味着专用的计算或隔离的服务能力:两个适配器都使用 qwen-prod 的共享资源。

The shared deployment must use the compatible base-model version and a deployment template and runtime that support LoRA. Matching the base-model name alone is insufficient—the service also validates adapter compatibility and available adapter capacity.

共享部署必须使用兼容的基础模型版本以及支持 LoRA 的部署模板和运行时。仅匹配基础模型名称是不够的——服务还会验证适配器兼容性和可用的适配器容量。

From fine-tuning to deployment in the portal

从微调到门户中的部署

The Foundry portal connects a registered fine-tuning output to the compute that will serve it. You can start from Build > Models or a completed fine-tuning job, without leaving the adapter deployment flow to set up a target separately.

Foundry 门户将注册微调输出与将为其提供服务的计算连接起来。您可以从“构建”>“模型”或完成的微调作业开始,无需离开适配器部署流程即可单独设置目标。

  • Select your adapter: Start with the registered output from a completed Foundry fine-tuning job and begin a LoRA-enabled Managed Compute deployment.
  • Choose compatible target deployment. Reuse an existing LoRA-enabled target deployment or create a new LoRA-enabled target deployment in the same flow.
  • Name and deploy the adapter. Once it's ready, open its details to test the specialized behavior.
  • 选择您的适配器:从完成的 Foundry 微调作业的注册输出开始,并开始启用 LoRA 的托管计算部署。
  • 选择兼容的目标部署。重用现有的启用 LoRA 的目标部署,或者在同一流程中创建新的启用 LoRA 的目标部署。
  • 命名并部署适配器。一旦准备就绪,打开其详细信息以测试专门的行为。

To add the finance specialization, deploy its compatible adapter against the same qwen-prod target using the name qwen-finance. The target's details bring the attached adapter deployments together, making the shared-serving arrangement easy to inspect.

要添加金融专业领域适配器,请使用名称 qwen-finance 将其兼容的适配器部署到相同的 qwen-prod 目标。目标的详细信息会将附加的适配器部署整合在一起,使共享服务配置易于检查。

Open an adapter in the playground and try a prompt relevant to its task. Select the adapter deployment, rather than only the base-model target, to evaluate the fine-tuned behavior.

在 Playground 中打开一个适配器并尝试与其任务相关的提示词。选择适配器部署,而不仅仅是基础模型目标,以评估微调后的行为。

Deploy with SDK calls

使用 SDK 调用进行部署

For automated workflows, you can deploy a registered LoRA adapter to a compatible Managed Compute target using the Python management SDK. The azure-mgmt-cognitiveservices 15.0.0b6 prerelease includes adapter deployment operations.

对于自动化工作流,您可以使用 Python 管理 SDK 将注册的 LoRA 适配器部署到兼容的托管计算(Managed Compute)目标。azure-mgmt-cognitiveservices 15.0.0b6 预发布版本包含了适配器部署操作。

The example below assumes that you have an authenticated management client named `client`, the required resource identifiers, and an existing compatible target deployment named `qwen-prod`. For the complete deployment flow and prerequisites, see Deploy fine-tuned LoRA adapters to Managed Compute for more details.

以下示例假设您拥有一个经过身份验证的管理客户端 `client`、所需的资源标识符以及一个现有的兼容目标部署 `qwen-prod`。有关完整的部署流程和先决条件,请参阅“将微调后的 LoRA 适配器部署到托管计算”以获取更多详细信息。

Connect the support adapter to the shared target

将支持适配器连接到共享目标

Use the registered adapter model ID as the source and specify the Managed Compute LoRA-enabled deployment that will serve it as the target.

使用注册的适配器模型 ID 作为源,并指定将为其提供服务的托管计算 LoRA 启用部署作为目标。

from azure.mgmt.cognitiveservices.models import ( AdapterDeployment, AdapterDeploymentProperties, ) client.adapter_deployments.begin_create_or_update( resource_group_name=RESOURCE_GROUP, account_name=ACCOUNT_NAME, adapter_deployment_name="qwen-support", resource=AdapterDeployment( properties=AdapterDeploymentProperties( source_model_id=SUPPORT_ADAPTER_MODEL_ID, target_deployment_name="qwen-prod", ) ), ).result()

To deploy another compatible adapter, repeat the operation with its registered model ID and a unique adapter deployment name while keeping `qwen-prod` as the target. Multiple adapters can share the same target deployment and its underlying compute.

要部署另一个兼容的适配器,请重复该操作,使用其注册的模型 ID 和唯一的适配器部署名称,同时保持 `qwen-prod` 为目标。多个适配器可以共享同一个目标部署及其底层计算资源。

Choose the specialization in your application

在您的应用程序中选择专业领域

After the adapter deployment succeeds, use an authenticated OpenAI client configured for your Foundry account endpoint. Set model to the adapter deployment name, rather than the target deployment name, to select the adapter your application should use.

适配器部署成功后,使用为您的 Foundry 账户端点配置的经过身份验证的 OpenAI 客户端。将 model 设置为适配器部署名称,而不是目标部署名称,以选择应用程序应使用的适配器。

For example, to send a request to the `qwen-support` adapter:

例如,要向 `qwen-support` 适配器发送请求:

completion = openai_client.chat.completions.create( model="qwen-support", messages=[ {"role": "user", "content": "How do I reset my account password?"} ], )

To invoke another adapter hosted on the same target, simply change the `model` value to that adapter's deployment name. The application can therefore select between specialized adapters while they share the same Managed Compute target.

要调用托管在同一目标上的另一个适配器,只需将 `model` 值更改为该适配器的部署名称。因此,应用程序可以在共享同一 Managed Compute 目标的同时,在专用适配器之间进行选择。

Pricing and capacity considerations

定价与容量考量

Before deploying Multi-LoRA, here are a few things to keep in mind:

在部署 Multi-LoRA 之前,请注意以下几点:

  • Pricing: Adapter deployments have no separate pricing and don’t provision additional compute. Billing applies to the shared Managed Compute target deployment, which is billed hourly based on the accelerators used.
  • Capacity: A target deployment can host up to eight adapter deployments. Attached adapters share the target deployment’s serving capacity, and adding an adapter doesn’t create another model instance or increase the target’s capacity.
  • 定价:适配器部署没有单独的定价,也不会配置额外的计算资源。费用适用于共享的 Managed Compute 目标部署,该部署根据所使用的加速器按小时计费。
  • 容量:一个目标部署最多可托管八个适配器部署。附加的适配器共享目标部署的服务容量,添加适配器不会创建新的模型实例,也不会增加目标的容量。

更进一步:量化金融体系

看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力

进入量化体系 →

相似阅读

关联信息,但可能不是同一事件