✨ feat: add Xinference provider support (#7081)

Dockerfile

@@ -227,6 +227,8 @@ ENV \
     WENXIN_API_KEY="" WENXIN_MODEL_LIST="" \
     # xAI
     XAI_API_KEY="" XAI_MODEL_LIST="" XAI_PROXY_URL="" \
+    # Xinference
+    XINFERENCE_API_KEY="" XINFERENCE_MODEL_LIST="" XINFERENCE_PROXY_URL="" \
     # 01.AI
     ZEROONE_API_KEY="" ZEROONE_MODEL_LIST="" \
     # Zhipu

Dockerfile.database

@@ -270,6 +270,8 @@ ENV \
     WENXIN_API_KEY="" WENXIN_MODEL_LIST="" \
     # xAI
     XAI_API_KEY="" XAI_MODEL_LIST="" XAI_PROXY_URL="" \
+    # Xinference
+    XINFERENCE_API_KEY="" XINFERENCE_MODEL_LIST="" XINFERENCE_PROXY_URL="" \
     # 01.AI
     ZEROONE_API_KEY="" ZEROONE_MODEL_LIST="" \
     # Zhipu

Dockerfile.pglite

@@ -224,6 +224,8 @@ ENV \
     WENXIN_API_KEY="" WENXIN_MODEL_LIST="" \
     # xAI
     XAI_API_KEY="" XAI_MODEL_LIST="" XAI_PROXY_URL="" \
+    # Xinference
+    XINFERENCE_API_KEY="" XINFERENCE_MODEL_LIST="" XINFERENCE_PROXY_URL="" \
     # 01.AI
     ZEROONE_API_KEY="" ZEROONE_MODEL_LIST="" \
     # Zhipu

src/app/[variants]/(main)/settings/llm/ProviderList/providers.tsx

@@ -37,6 +37,7 @@ import {
   VLLMProviderCard,
   WenxinProviderCard,
   XAIProviderCard,
+  XinferenceProviderCard,
   ZeroOneProviderCard,
   ZhiPuProviderCard,
 } from '@/config/modelProviders';
@@ -65,6 +66,7 @@ export const useProviderList = (): ProviderItem[] => {
       AzureProvider,
       OllamaProvider,
       VLLMProviderCard,
+      XinferenceProviderCard,
       AnthropicProviderCard,
       BedrockProvider,
       GoogleProviderCard,

src/config/aiModels/index.ts

@@ -49,6 +49,7 @@ import { default as vllm } from './vllm';
 import { default as volcengine } from './volcengine';
 import { default as wenxin } from './wenxin';
 import { default as xai } from './xai';
+import { default as xinference } from './xinference';
 import { default as zeroone } from './zeroone';
 import { default as zhipu } from './zhipu';
 
@@ -121,6 +122,7 @@ export const LOBE_DEFAULT_MODEL_LIST = buildDefaultModelList({
   volcengine,
   wenxin,
   xai,
+  xinference,
   zeroone,
   zhipu,
 });
@@ -174,5 +176,6 @@ export { default as vllm } from './vllm';
 export { default as volcengine } from './volcengine';
 export { default as wenxin } from './wenxin';
 export { default as xai } from './xai';
+export { default as xinference } from './xinference';
 export { default as zeroone } from './zeroone';
 export { default as zhipu } from './zhipu';

src/config/aiModels/xinference.ts

@@ -0,0 +1,171 @@
+import { AIChatModelCard } from '@/types/aiModel';
+
+const xinferenceChatModels: AIChatModelCard[] = [
+  {
+    abilities: {
+      functionCall: true,
+    },
+    contextWindowTokens: 163_840,
+    description:
+      'DeepSeek-V3 是一个强大的专家混合（MoE）语言模型，拥有总计 6710 亿参数，每个 token 激活 370 亿参数。',
+    displayName: 'DeepSeek V3',
+    enabled: true,
+    id: 'deepseek-v3',
+    type: 'chat'
+  },
+  {
+    abilities: {
+      reasoning: true,
+    },
+    contextWindowTokens: 163_840,
+    description:
+      'DeepSeek-R1 在强化学习（RL）之前引入了冷启动数据，在数学、代码和推理任务上表现可与 OpenAI-o1 相媲美。',
+    displayName: 'DeepSeek R1',
+    enabled: true,
+    id: 'deepseek-r1',
+    type: 'chat'
+  },
+  {
+    abilities: {
+      reasoning: true,
+    },
+    contextWindowTokens: 131_072,
+    description:
+      'deepseek-r1-distill-llama 是基于 Llama 从 DeepSeek-R1 蒸馏而来的模型。',
+    displayName: 'DeepSeek R1 Distill Llama',
+    enabled: true,
+    id: 'deepseek-r1-distill-llama',
+    type: 'chat'
+  },
+  {
+    abilities: {
+      reasoning: true,
+    },
+    contextWindowTokens: 131_072,
+    description:
+      'deepseek-r1-distill-qwen 是基于 Qwen 从 DeepSeek-R1 蒸馏而来的模型。',
+    displayName: 'DeepSeek R1 Distill Qwen',
+    enabled: true,
+    id: 'deepseek-r1-distill-qwen',
+    type: 'chat'
+  },
+  {
+    abilities: {
+      reasoning: true,
+    },
+    contextWindowTokens: 32_768,
+    description:
+      'QwQ 是 Qwen 系列的推理模型。与传统的指令微调模型相比，QwQ 具备思考和推理能力，在下游任务中，尤其是复杂问题上，能够实现显著增强的性能。QwQ-32B 是一款中型推理模型，其性能可与最先进的推理模型（如 DeepSeek-R1、o1-mini）相媲美。',
+    displayName: 'QwQ 32B',
+    enabled: true,
+    id: 'qwq-32b',
+    type: 'chat'
+  },
+  {
+    abilities: {
+      reasoning: true,
+      vision: true,
+    },
+    contextWindowTokens: 32_768,
+    description:
+      'QVQ-72B-Preview 是由 Qwen 团队开发的实验性研究模型，专注于提升视觉推理能力。',
+    displayName: 'QVQ 72B Preview',
+    enabled: true,
+    id: 'qvq-72b-preview',
+    type: 'chat'
+  },
+  {
+    abilities: {
+      functionCall: true,
+    },
+    contextWindowTokens: 32_768,
+    description:
+      'Qwen2.5 是 Qwen 大型语言模型的最新系列。对于 Qwen2.5，我们发布了多个基础语言模型和指令微调语言模型，参数范围从 5 亿到 72 亿不等。',
+    displayName: 'Qwen2.5 Instruct',
+    enabled: true,
+    id: 'qwen2.5-instruct',
+    type: 'chat'
+  },
+  {
+    abilities: {
+      functionCall: true,
+    },
+    contextWindowTokens: 32_768,
+    description:
+      'Qwen2.5-Coder 是 Qwen 系列中最新的代码专用大型语言模型（前身为 CodeQwen）。',
+    displayName: 'Qwen2.5 Coder Instruct',
+    enabled: true,
+    id: 'qwen2.5-coder-instruct',
+    type: 'chat'
+  },
+  {
+    abilities: {
+      vision: true,
+    },
+    contextWindowTokens: 128_000,
+    description:
+      'Qwen2.5-VL 是 Qwen 模型家族中视觉语言模型的最新版本。',
+    displayName: 'Qwen2.5 VL Instruct',
+    enabled: true,
+    id: 'qwen2.5-vl-instruct',
+    type: 'chat'
+  },
+  {
+    contextWindowTokens: 1_024_000,
+    description:
+      'Mistral-Nemo-Instruct-2407 大型语言模型（LLM）是 Mistral-Nemo-Base-2407 的指令微调版本。',
+    displayName: 'Mistral Nemo Instruct',
+    enabled: true,
+    id: 'mistral-nemo-instruct',
+    type: 'chat'
+  },
+  {
+    contextWindowTokens: 131_072,
+    description:
+      'Mistral-Large-Instruct-2407 是一款先进的稠密大型语言模型（LLM），拥有 1230 亿参数，具备最先进的推理、知识和编码能力。',
+    displayName: 'Mistral Large Instruct',
+    enabled: true,
+    id: 'mistral-large-instruct',
+    type: 'chat'
+  },
+  {
+    abilities: {
+      functionCall: true,
+    },
+    contextWindowTokens: 131_072,
+    description:
+      'Llama 3.3 指令微调模型针对对话场景进行了优化，在常见的行业基准测试中，超越了许多现有的开源聊天模型。',
+    displayName: 'Llama 3.3 Instruct',
+    enabled: true,
+    id: 'llama-3.3-instruct',
+    type: 'chat'
+  },
+  {
+    abilities: {
+      vision: true,
+    },
+    contextWindowTokens: 163_840,
+    description:
+      'Llama 3.2-Vision 指令微调模型针对视觉识别、图像推理、图像描述和回答与图像相关的常规问题进行了优化。',
+    displayName: 'Llama 3.2 Vision Instruct',
+    enabled: true,
+    id: 'llama-3.2-vision-instruct',
+    type: 'chat'
+  },
+  {
+    abilities: {
+      functionCall: true,
+    },
+    contextWindowTokens: 131_072,
+    description:
+      'Llama 3.1 指令微调模型针对对话场景进行了优化，在常见的行业基准测试中，超越了许多现有的开源聊天模型。',
+    displayName: 'Llama 3.1 Instruct',
+    enabled: true,
+    id: 'llama-3.1-instruct',
+    type: 'chat'
+  },
+]
+
+export const allModels = [...xinferenceChatModels];
+
+export default allModels;

src/config/llm.ts

@@ -75,6 +75,9 @@ export const getLLMConfig = () => {
       ENABLED_VLLM: z.boolean(),
       VLLM_API_KEY: z.string().optional(),
 
+      ENABLED_XINFERENCE: z.boolean(),
+      XINFERENCE_API_KEY: z.string().optional(),
+
       ENABLED_QWEN: z.boolean(),
       QWEN_API_KEY: z.string().optional(),
 
@@ -234,6 +237,9 @@ export const getLLMConfig = () => {
       ENABLED_VLLM: !!process.env.VLLM_API_KEY,
       VLLM_API_KEY: process.env.VLLM_API_KEY,
 
+      ENABLED_XINFERENCE: !!process.env.XINFERENCE_API_KEY,
+      XINFERENCE_API_KEY: process.env.XINFERENCE_API_KEY,
+
       ENABLED_QWEN: !!process.env.QWEN_API_KEY,
       QWEN_API_KEY: process.env.QWEN_API_KEY,
 

src/config/modelProviders/index.ts

@@ -49,6 +49,7 @@ import VLLMProvider from './vllm';
 import VolcengineProvider from './volcengine';
 import WenxinProvider from './wenxin';
 import XAIProvider from './xai';
+import XinferenceProvider from './xinference';
 import ZeroOneProvider from './zeroone';
 import ZhiPuProvider from './zhipu';
 
@@ -69,6 +70,7 @@ export const LOBE_DEFAULT_MODEL_LIST: ChatModelCard[] = [
   MoonshotProvider.chatModels,
   OllamaProvider.chatModels,
   VLLMProvider.chatModels,
+  XinferenceProvider.chatModels,
   OpenRouterProvider.chatModels,
   TogetherAIProvider.chatModels,
   FireworksAIProvider.chatModels,
@@ -108,6 +110,7 @@ export const DEFAULT_MODEL_PROVIDER_LIST = [
   AzureAIProvider,
   OllamaProvider,
   VLLMProvider,
+  XinferenceProvider,
   AnthropicProvider,
   BedrockProvider,
   GoogleProvider,
@@ -215,5 +218,6 @@ export { default as VLLMProviderCard } from './vllm';
 export { default as VolcengineProviderCard } from './volcengine';
 export { default as WenxinProviderCard } from './wenxin';
 export { default as XAIProviderCard } from './xai';
+export { default as XinferenceProviderCard } from './xinference';
 export { default as ZeroOneProviderCard } from './zeroone';
 export { default as ZhiPuProviderCard } from './zhipu';

src/config/modelProviders/xinference.ts

@@ -0,0 +1,18 @@
+import { ModelProviderCard } from '@/types/llm';
+
+const Xinference: ModelProviderCard = {
+  chatModels: [],
+  description: 'Xorbits Inference (Xinference) 是一个开源平台，用于简化各种 AI 模型的运行和集成。借助 Xinference，您可以使用任何开源 LLM、嵌入模型和多模态模型在云端或本地环境中运行推理，并创建强大的 AI 应用。',
+  id: 'xinference',
+  modelsUrl: 'https://inference.readthedocs.io/zh-cn/latest/models/builtin/index.html',
+  name: 'Xinference',
+  settings: {
+    proxyUrl: {
+      placeholder: 'http://localhost:9997/v1',
+    },
+    sdkType: 'openai',
+  },
+  url: 'https://inference.readthedocs.io/zh-cn/v0.12.3/index.html',
+};
+
+export default Xinference;

src/libs/agent-runtime/runtimeMap.ts

@@ -45,6 +45,7 @@ import { LobeVLLMAI } from './vllm';
 import { LobeVolcengineAI } from './volcengine';
 import { LobeWenxinAI } from './wenxin';
 import { LobeXAI } from './xai';
+import { LobeXinferenceAI } from './xinference';
 import { LobeZeroOneAI } from './zeroone';
 import { LobeZhipuAI } from './zhipu';
 
@@ -97,6 +98,7 @@ export const providerRuntimeMap = {
   volcengine: LobeVolcengineAI,
   wenxin: LobeWenxinAI,
   xai: LobeXAI,
+  xinference: LobeXinferenceAI,
   zeroone: LobeZeroOneAI,
   zhipu: LobeZhipuAI,
 };

src/libs/agent-runtime/types/type.ts

@@ -74,6 +74,7 @@ export enum ModelProvider {
   Volcengine = 'volcengine',
   Wenxin = 'wenxin',
   XAI = 'xai',
+  Xinference = 'xinference',
   ZeroOne = 'zeroone',
   ZhiPu = 'zhipu',
 }

src/libs/agent-runtime/xinference/index.ts

@@ -0,0 +1,53 @@
+import { ModelProvider } from '../types';
+import { LobeOpenAICompatibleFactory } from '../utils/openaiCompatibleFactory';
+
+import type { ChatModelCard } from '@/types/llm';
+
+export interface XinferenceModelCard {
+  context_length: number;
+  id: string;
+  model_ability: string[];
+  model_description: string;
+  model_type: string;
+  name: string;
+}
+
+export const LobeXinferenceAI = LobeOpenAICompatibleFactory({
+  baseURL: 'http://localhost:9997/v1',
+  debug: {
+    chatCompletion: () => process.env.DEBUG_XINFERENCE_CHAT_COMPLETION === '1',
+  },
+  models: async ({ client }) => {
+    const { LOBE_DEFAULT_MODEL_LIST } = await import('@/config/aiModels');
+
+    const modelsPage = await client.models.list() as any;
+    const modelList: XinferenceModelCard[] = modelsPage.data;
+
+    return modelList
+      .map((model) => {
+        const knownModel = LOBE_DEFAULT_MODEL_LIST.find((m) => model.id.toLowerCase() === m.id.toLowerCase());
+
+        return {
+          contextWindowTokens: model.context_length,
+          description: model.model_description,
+          displayName: model.name,
+          enabled: knownModel?.enabled || false,
+          functionCall:
+            (model.model_ability && model.model_ability.includes("tools"))
+            || knownModel?.abilities?.functionCall
+            || false,
+          id: model.id,
+          reasoning:
+            (model.model_ability && model.model_ability.includes("reasoning"))
+            || knownModel?.abilities?.reasoning
+            || false,
+          vision:
+            (model.model_ability && model.model_ability.includes("vision"))
+            || knownModel?.abilities?.vision
+            || false,
+        };
+      })
+      .filter(Boolean) as ChatModelCard[];
+  },
+  provider: ModelProvider.Xinference,
+});

src/types/user/settings/keyVaults.ts

@@ -84,6 +84,7 @@ export interface UserKeyVaults extends SearchEngineKeyVaults {
   volcengine?: OpenAICompatibleKeyVault;
   wenxin?: OpenAICompatibleKeyVault;
   xai?: OpenAICompatibleKeyVault;
+  xinference?: OpenAICompatibleKeyVault;
   zeroone?: OpenAICompatibleKeyVault;
   zhipu?: OpenAICompatibleKeyVault;
 }