Merge branch 'main' of https://github.com/huggingface/huggingface.js into kai/hyperbolic-integration

Author: connorch

README.md

@@ -96,7 +96,7 @@ You can run our packages with vanilla JS, without any bundler, by using a CDN or
 
 ```html
 <script type="module">
-    import { HfInference } from 'https://cdn.jsdelivr.net/npm/@huggingface/[email protected].0/+esm';
+    import { HfInference } from 'https://cdn.jsdelivr.net/npm/@huggingface/[email protected].3/+esm';
     import { createRepo, commit, deleteRepo, listFiles } from "https://cdn.jsdelivr.net/npm/@huggingface/[email protected]/+esm";
 </script>
 ```

packages/agents/src/tools/imageToText.ts

@@ -15,12 +15,9 @@ export const imageToTextTool: Tool = {
 		if (typeof data === "string") throw "Input must be a blob.";
 
 		return (
-			await inference.imageToText(
-				{
-					data,
-				},
-				{ wait_for_model: true }
-			)
+			await inference.imageToText({
+				data,
+			})
 		).generated_text;
 	},
 };

packages/agents/src/tools/speechToText.ts

@@ -15,12 +15,9 @@ export const speechToTextTool: Tool = {
 		if (typeof data === "string") throw "Input must be a blob.";
 
 		return (
-			await inference.automaticSpeechRecognition(
-				{
-					data,
-				},
-				{ wait_for_model: true }
-			)
+			await inference.automaticSpeechRecognition({
+				data,
+			})
 		).text;
 	},
 };

packages/agents/src/tools/textToImage.ts

@@ -19,11 +19,8 @@ export const textToImageTool: Tool = {
 		const data = await input;
 		if (typeof data !== "string") throw "Input must be a string.";
 
-		return await inference.textToImage(
-			{
-				inputs: data,
-			},
-			{ wait_for_model: true }
-		);
+		return await inference.textToImage({
+			inputs: data,
+		});
 	},
 };

packages/agents/src/tools/textToSpeech.ts

@@ -19,12 +19,9 @@ export const textToSpeechTool: Tool = {
 		const data = await input;
 		if (typeof data !== "string") throw "Input must be a string.";
 
-		return inference.textToSpeech(
-			{
-				inputs: data,
-				model: "espnet/kan-bayashi_ljspeech_vits",
-			},
-			{ wait_for_model: true }
-		);
+		return inference.textToSpeech({
+			inputs: data,
+			model: "espnet/kan-bayashi_ljspeech_vits",
+		});
 	},
 };

packages/inference/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@huggingface/inference",
-	"version": "3.3.0",
+	"version": "3.3.3",
 	"packageManager": "[email protected]",
 	"license": "MIT",
 	"author": "Tim Mikeladze <[email protected]>",

packages/inference/src/lib/makeRequestOptions.ts

@@ -40,8 +40,7 @@ export async function makeRequestOptions(
 	let otherArgs = remainingArgs;
 	const provider = maybeProvider ?? "hf-inference";
 
-	const { forceTask, includeCredentials, taskHint, wait_for_model, use_cache, dont_load_model, chatCompletion } =
-		options ?? {};
+	const { forceTask, includeCredentials, taskHint, chatCompletion } = options ?? {};
 
 	if (endpointUrl && provider !== "hf-inference") {
 		throw new Error(`Cannot use endpointUrl with a third-party provider.`);
@@ -103,18 +102,6 @@ export async function makeRequestOptions(
 		headers["Content-Type"] = "application/json";
 	}
 
-	if (provider === "hf-inference") {
-		if (wait_for_model) {
-			headers["X-Wait-For-Model"] = "true";
-		}
-		if (use_cache === false) {
-			headers["X-Use-Cache"] = "false";
-		}
-		if (dont_load_model) {
-			headers["X-Load-Model"] = "0";
-		}
-	}
-
 	if (provider === "replicate") {
 		headers["Prefer"] = "wait";
 	}

packages/inference/src/tasks/custom/request.ts

@@ -18,11 +18,8 @@ export async function request<T>(
 	const { url, info } = await makeRequestOptions(args, options);
 	const response = await (options?.fetch ?? fetch)(url, info);
 
-	if (options?.retry_on_error !== false && response.status === 503 && !options?.wait_for_model) {
-		return request(args, {
-			...options,
-			wait_for_model: true,
-		});
+	if (options?.retry_on_error !== false && response.status === 503) {
+		return request(args, options);
 	}
 
 	if (!response.ok) {

packages/inference/src/tasks/custom/streamingRequest.ts

@@ -20,11 +20,8 @@ export async function* streamingRequest<T>(
 	const { url, info } = await makeRequestOptions({ ...args, stream: true }, options);
 	const response = await (options?.fetch ?? fetch)(url, info);
 
-	if (options?.retry_on_error !== false && response.status === 503 && !options?.wait_for_model) {
-		return yield* streamingRequest(args, {
-			...options,
-			wait_for_model: true,
-		});
+	if (options?.retry_on_error !== false && response.status === 503) {
+		return yield* streamingRequest(args, options);
 	}
 	if (!response.ok) {
 		if (response.headers.get("Content-Type")?.startsWith("application/json")) {

packages/inference/src/tasks/cv/textToImage.ts

@@ -1,6 +1,6 @@
 import type { TextToImageInput, TextToImageOutput } from "@huggingface/tasks";
 import { InferenceOutputError } from "../../lib/InferenceOutputError";
-import type { BaseArgs, Options } from "../../types";
+import type { BaseArgs, InferenceProvider, Options } from "../../types";
 import { omit } from "../../utils/omit";
 import { request } from "../custom/request";
 
@@ -15,29 +15,40 @@ interface OutputUrlImageGeneration {
 	output: string[];
 }
 
+function getResponseFormatArg(provider: InferenceProvider) {
+	switch (provider) {
+		case "fal-ai":
+			return { sync_mode: true };
+		case "nebius":
+			return { response_format: "b64_json" };
+		case "replicate":
+			return undefined;
+		case "together":
+			return { response_format: "base64" };
+		default:
+			return undefined;
+	}
+}
+
 /**
  * This task reads some text input and outputs an image.
  * Recommended model: stabilityai/stable-diffusion-2
  */
 export async function textToImage(args: TextToImageArgs, options?: Options): Promise<Blob> {
 	const payload =
-		args.provider === "together" ||
-		args.provider === "fal-ai" ||
-		args.provider === "replicate" ||
-		args.provider === "nebius" ||
-		args.provider === "hyperbolic"
-			? {
+		!args.provider || args.provider === "hf-inference" || args.provider === "sambanova"
+			? args
+			: {
 					...omit(args, ["inputs", "parameters"]),
 					...args.parameters,
-					...(args.provider !== "replicate" ? { response_format: "base64" } : undefined),
-					...(args.provider === "nebius" ? { response_format: "b64_json" } : undefined),
+					...getResponseFormatArg(args.provider),
 					prompt: args.inputs,
-			  }
-			: args;
+			  };
 	const res = await request<TextToImageOutput | Base64ImageGeneration | OutputUrlImageGeneration>(payload, {
 		...options,
 		taskHint: "text-to-image",
 	});
+
 	if (res && typeof res === "object") {
 		if (args.provider === "fal-ai" && "images" in res && Array.isArray(res.images) && res.images[0].url) {
 			const image = await fetch(res.images[0].url);