Merge branch 'main' into main

Author: mingcheng

README.md

@@ -98,7 +98,7 @@ You can run our packages with vanilla JS, without any bundler, by using a CDN or
 
 ```html
 <script type="module">
-    import { InferenceClient } from 'https://cdn.jsdelivr.net/npm/@huggingface/inference@4.11.3/+esm';
+    import { InferenceClient } from 'https://cdn.jsdelivr.net/npm/@huggingface/inference@4.12.0/+esm';
     import { createRepo, commit, deleteRepo, listFiles } from "https://cdn.jsdelivr.net/npm/@huggingface/[email protected]/+esm";
 </script>
 ```

packages/inference/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@huggingface/inference",
-	"version": "4.11.3",
+	"version": "4.12.0",
 	"packageManager": "[email protected]",
 	"license": "MIT",
 	"author": "Hugging Face and Tim Mikeladze <[email protected]>",

packages/inference/src/errors.ts

@@ -17,6 +17,13 @@ export class InferenceClientInputError extends InferenceClientError {
 	}
 }
 
+export class InferenceClientRoutingError extends InferenceClientError {
+	constructor(message: string) {
+		super(message);
+		this.name = "RoutingError";
+	}
+}
+
 interface HttpRequest {
 	url: string;
 	method: string;

packages/inference/src/lib/getInferenceProviderMapping.ts

@@ -124,6 +124,17 @@ export async function getInferenceProviderMapping(
 	}
 ): Promise<InferenceProviderMappingEntry | null> {
 	const logger = getLogger();
+	if (params.provider === ("auto" as InferenceProvider) && params.task === "conversational") {
+		// Special case for auto + conversational to avoid extra API calls
+		// Call directly the server-side auto router
+		return {
+			hfModelId: params.modelId,
+			provider: "auto",
+			providerId: params.modelId,
+			status: "live",
+			task: "conversational",
+		};
+	}
 	if (HARDCODED_MODEL_INFERENCE_MAPPING[params.provider][params.modelId]) {
 		return HARDCODED_MODEL_INFERENCE_MAPPING[params.provider][params.modelId];
 	}

packages/inference/src/package.ts

@@ -1,3 +1,3 @@
 // Generated file from package.json. Issues importing JSON directly when publishing on commonjs/ESM - see https://github.com/microsoft/TypeScript/issues/51783
-export const PACKAGE_VERSION = "4.11.3";
+export const PACKAGE_VERSION = "4.12.0";
 export const PACKAGE_NAME = "@huggingface/inference";

packages/inference/src/providers/providerHelper.ts

@@ -47,7 +47,7 @@ import type {
 	ZeroShotImageClassificationOutput,
 } from "@huggingface/tasks";
 import { HF_ROUTER_URL } from "../config.js";
-import { InferenceClientProviderOutputError } from "../errors.js";
+import { InferenceClientProviderOutputError, InferenceClientRoutingError } from "../errors.js";
 import type { AudioToAudioOutput } from "../tasks/audio/audioToAudio.js";
 import type { BaseArgs, BodyParams, HeaderParams, InferenceProvider, RequestArgs, UrlParams } from "../types.js";
 import { toArray } from "../utils/toArray.js";
@@ -62,7 +62,7 @@ import type { ImageSegmentationArgs } from "../tasks/cv/imageSegmentation.js";
 export abstract class TaskProviderHelper {
 	constructor(
 		readonly provider: InferenceProvider,
-		private baseUrl: string,
+		protected baseUrl: string,
 		readonly clientSideRoutingOnly: boolean = false
 	) {}
 
@@ -369,3 +369,16 @@ export class BaseTextGenerationTask extends TaskProviderHelper implements TextGe
 		throw new InferenceClientProviderOutputError("Expected Array<{generated_text: string}>");
 	}
 }
+
+export class AutoRouterConversationalTask extends BaseConversationalTask {
+	constructor() {
+		super("auto" as InferenceProvider, "https://router.huggingface.co");
+	}
+
+	override makeBaseUrl(params: UrlParams): string {
+		if (params.authMethod !== "hf-token") {
+			throw new InferenceClientRoutingError("Cannot select auto-router when using non-Hugging Face API key.");
+		}
+		return this.baseUrl;
+	}
+}

packages/inference/src/tasks/nlp/chatCompletion.ts

@@ -3,6 +3,8 @@ import { resolveProvider } from "../../lib/getInferenceProviderMapping.js";
 import { getProviderHelper } from "../../lib/getProviderHelper.js";
 import type { BaseArgs, Options } from "../../types.js";
 import { innerRequest } from "../../utils/request.js";
+import type { ConversationalTaskHelper, TaskProviderHelper } from "../../providers/providerHelper.js";
+import { AutoRouterConversationalTask } from "../../providers/providerHelper.js";
 
 /**
  * Use the chat completion endpoint to generate a response to a prompt, using OpenAI message completion API no stream
@@ -11,8 +13,14 @@ export async function chatCompletion(
 	args: BaseArgs & ChatCompletionInput,
 	options?: Options
 ): Promise<ChatCompletionOutput> {
-	const provider = await resolveProvider(args.provider, args.model, args.endpointUrl);
-	const providerHelper = getProviderHelper(provider, "conversational");
+	let providerHelper: ConversationalTaskHelper & TaskProviderHelper;
+	if (!args.provider || args.provider === "auto") {
+		// Special case: we have a dedicated auto-router for conversational models. No need to fetch provider mapping.
+		providerHelper = new AutoRouterConversationalTask();
+	} else {
+		const provider = await resolveProvider(args.provider, args.model, args.endpointUrl);
+		providerHelper = getProviderHelper(provider, "conversational");
+	}
 	const { data: response } = await innerRequest<ChatCompletionOutput>(args, providerHelper, {
 		...options,
 		task: "conversational",

packages/tasks/package.json

@@ -1,7 +1,7 @@
 {
 	"name": "@huggingface/tasks",
 	"packageManager": "[email protected]",
-	"version": "0.19.56",
+	"version": "0.19.57",
 	"description": "List of ML tasks for huggingface.co/tasks",
 	"repository": "https://github.com/huggingface/huggingface.js.git",
 	"publishConfig": {

packages/tasks/src/model-libraries.ts

@@ -809,8 +809,10 @@ export const MODEL_LIBRARIES_UI_ELEMENTS = {
 		prettyLabel: "PaddleOCR",
 		repoName: "PaddleOCR",
 		repoUrl: "https://github.com/PaddlePaddle/PaddleOCR",
+		docsUrl: "https://www.paddleocr.ai/",
 		snippets: snippets.paddleocr,
 		filter: true,
+		countDownloads: `path_extension:"safetensors" OR path:"inference.pdiparams"`,
 	},
 	peft: {
 		prettyLabel: "PEFT",