✨ Integrate Nebius AI Studio into HuggingFace inference

.github/workflows/test.yml

@@ -42,6 +42,7 @@ jobs:
         env:
           HF_TOKEN: ${{ secrets.HF_TOKEN }}
           HF_FAL_KEY: dummy
+          HF_NEBIUS_KEY: dummy
           HF_REPLICATE_KEY: dummy
           HF_SAMBANOVA_KEY: dummy
           HF_TOGETHER_KEY: dummy
@@ -83,6 +84,7 @@ jobs:
         env:
           HF_TOKEN: ${{ secrets.HF_TOKEN }}
           HF_FAL_KEY: dummy
+          HF_NEBIUS_KEY: dummy
           HF_REPLICATE_KEY: dummy
           HF_SAMBANOVA_KEY: dummy
           HF_TOGETHER_KEY: dummy
@@ -151,6 +153,7 @@ jobs:
           NPM_CONFIG_REGISTRY: http://localhost:4874/
           HF_TOKEN: ${{ secrets.HF_TOKEN }}
           HF_FAL_KEY: dummy
+          HF_NEBIUS_KEY: dummy
           HF_REPLICATE_KEY: dummy
           HF_SAMBANOVA_KEY: dummy
           HF_TOGETHER_KEY: dummy

packages/inference/README.md

@@ -49,6 +49,7 @@ You can send inference requests to third-party providers with the inference clie
 Currently, we support the following providers:
 - [Fal.ai](https://fal.ai)
 - [Fireworks AI](https://fireworks.ai)
+- [Nebius](https://studio.nebius.ai)
 - [Replicate](https://replicate.com)
 - [Sambanova](https://sambanova.ai)
 - [Together](https://together.xyz)
@@ -71,12 +72,13 @@ When authenticated with a third-party provider key, the request is made directly
 Only a subset of models are supported when requesting third-party providers. You can check the list of supported models per pipeline tasks here:
 - [Fal.ai supported models](https://huggingface.co/api/partners/fal-ai/models)
 - [Fireworks AI supported models](https://huggingface.co/api/partners/fireworks-ai/models)
+- [Nebius supported models](https://huggingface.co/api/partners/nebius/models)
 - [Replicate supported models](https://huggingface.co/api/partners/replicate/models)
 - [Sambanova supported models](https://huggingface.co/api/partners/sambanova/models)
 - [Together supported models](https://huggingface.co/api/partners/together/models)
 - [HF Inference API (serverless)](https://huggingface.co/models?inference=warm&sort=trending)
 
-❗**Important note:** To be compatible, the third-party API must adhere to the "standard" shape API we expect on HF model pages for each pipeline task type. 
+❗**Important note:** To be compatible, the third-party API must adhere to the "standard" shape API we expect on HF model pages for each pipeline task type.
 This is not an issue for LLMs as everyone converged on the OpenAI API anyways, but can be more tricky for other tasks like "text-to-image" or "automatic-speech-recognition" where there exists no standard API. Let us know if any help is needed or if we can make things easier for you!
 
 👋**Want to add another provider?** Get in touch if you'd like to add support for another Inference provider, and/or request it on https://huggingface.co/spaces/huggingface/HuggingDiscussions/discussions/49
@@ -463,7 +465,7 @@ await hf.zeroShotImageClassification({
   model: 'openai/clip-vit-large-patch14-336',
   inputs: {
     image: await (await fetch('https://placekitten.com/300/300')).blob()
-  },  
+  },
   parameters: {
     candidate_labels: ['cat', 'dog']
   }

packages/inference/src/lib/makeRequestOptions.ts

@@ -1,5 +1,6 @@
 import { HF_HUB_URL, HF_ROUTER_URL } from "../config";
 import { FAL_AI_API_BASE_URL } from "../providers/fal-ai";
+import { NEBIUS_API_BASE_URL } from "../providers/nebius";
 import { REPLICATE_API_BASE_URL } from "../providers/replicate";
 import { SAMBANOVA_API_BASE_URL } from "../providers/sambanova";
 import { TOGETHER_API_BASE_URL } from "../providers/together";
@@ -144,6 +145,7 @@ export async function makeRequestOptions(
 			: JSON.stringify({
 					...otherArgs,
 					...(chatCompletion || provider === "together" ? { model } : undefined),
+					...(provider === "nebius" ? { model } : {}),
 			  }),
 		...(credentials ? { credentials } : undefined),
 		signal: options?.signal,
@@ -172,6 +174,22 @@ function makeUrl(params: {
 				: FAL_AI_API_BASE_URL;
 			return `${baseUrl}/${params.model}`;
 		}
+		case "nebius": {
+			const baseUrl = shouldProxy
+				? HF_HUB_INFERENCE_PROXY_TEMPLATE.replace("{{PROVIDER}}", params.provider)
+				: NEBIUS_API_BASE_URL;
+
+			if (params.taskHint === "text-to-image") {
+				return `${baseUrl}/v1/images/generations`;
+			}
+			if (params.taskHint === "text-generation") {
+				if (params.chatCompletion) {
+					return `${baseUrl}/v1/chat/completions`;
+				}
+				return `${baseUrl}/v1/completions`;
+			}
+			return baseUrl;
+		}
 		case "replicate": {
 			const baseUrl = shouldProxy
 				? HF_HUB_INFERENCE_PROXY_TEMPLATE.replace("{{PROVIDER}}", params.provider)

packages/inference/src/providers/nebius.ts

@@ -0,0 +1,18 @@
+export const NEBIUS_API_BASE_URL = "https://api.studio.nebius.ai";
+
+/**
+ * See the registered mapping of HF model ID => Nebius model ID here:
+ *
+ * https://huggingface.co/api/partners/nebius/models
+ *
+ * This is a publicly available mapping.
+ *
+ * If you want to try to run inference for a new model locally before it's registered on huggingface.co,
+ * you can add it to the dictionary "HARDCODED_MODEL_ID_MAPPING" in consts.ts, for dev purposes.
+ *
+ * - If you work at Nebius and want to update this mapping, please use the model mapping API we provide on huggingface.co
+ * - If you're a community member and want to add a new supported HF model to Nebius, please open an issue on the present repo
+ * and we will tag Nebius team members.
+ *
+ * Thanks!
+ */

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

@@ -21,11 +21,15 @@ interface OutputUrlImageGeneration {
  */
 export async function textToImage(args: TextToImageArgs, options?: Options): Promise<Blob> {
 	const payload =
-		args.provider === "together" || args.provider === "fal-ai" || args.provider === "replicate"
+		args.provider === "together" ||
+		args.provider === "fal-ai" ||
+		args.provider === "replicate" ||
+		args.provider === "nebius"
 			? {
 					...omit(args, ["inputs", "parameters"]),
 					...args.parameters,
 					...(args.provider !== "replicate" ? { response_format: "base64" } : undefined),
+					...(args.provider === "nebius" ? { response_format: "b64_json" } : {}),
 					prompt: args.inputs,
 			  }
 			: args;

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

@@ -15,14 +15,17 @@ export async function chatCompletion(
 		taskHint: "text-generation",
 		chatCompletion: true,
 	});
+
 	const isValidOutput =
 		typeof res === "object" &&
 		Array.isArray(res?.choices) &&
 		typeof res?.created === "number" &&
 		typeof res?.id === "string" &&
 		typeof res?.model === "string" &&
-		/// Together.ai does not output a system_fingerprint
-		(res.system_fingerprint === undefined || typeof res.system_fingerprint === "string") &&
+		/// Together.ai and Nebius do not output a system_fingerprint
+		(res.system_fingerprint === undefined ||
+			res.system_fingerprint === null ||
+			typeof res.system_fingerprint === "string") &&
 		typeof res?.usage === "object";
 
 	if (!isValidOutput) {

packages/inference/src/types.ts

@@ -47,6 +47,7 @@ export type InferenceTask = Exclude<PipelineType, "other">;
 export const INFERENCE_PROVIDERS = [
 	"fal-ai",
 	"fireworks-ai",
+  "nebius", 
 	"hf-inference",
 	"replicate",
 	"sambanova",

packages/inference/test/HfInference.spec.ts

@@ -1064,6 +1064,50 @@ describe.concurrent("HfInference", () => {
 		TIMEOUT
 	);
 
+	describe.concurrent(
+		"Nebius",
+		() => {
+			const client = new HfInference(env.NEBIUS_API_KEY);
+
+			it("chatCompletion", async () => {
+				const res = await client.chatCompletion({
+					model: "meta-llama/Meta-Llama-3.1-8B-Instruct",
+					provider: "nebius",
+					messages: [{ role: "user", content: "Complete this sentence with words, one plus one is equal " }],
+				});
+				if (res.choices && res.choices.length > 0) {
+					const completion = res.choices[0].message?.content;
+					expect(completion).toMatch(/(two|2)/i);
+				}
+			});
+
+			it("chatCompletion stream", async () => {
+				const stream = client.chatCompletionStream({
+					model: "meta-llama/Meta-Llama-3.1-70B-Instruct",
+					provider: "nebius",
+					messages: [{ role: "user", content: "Complete the equation 1 + 1 = , just the answer" }],
+				}) as AsyncGenerator<ChatCompletionStreamOutput>;
+				let out = "";
+				for await (const chunk of stream) {
+					if (chunk.choices && chunk.choices.length > 0) {
+						out += chunk.choices[0].delta.content;
+					}
+				}
+				expect(out).toMatch(/(two|2)/i);
+			});
+
+			it("textToImage", async () => {
+				const res = await client.textToImage({
+					model: "black-forest-labs/flux-schnell",
+					provider: "nebius",
+					inputs: "award winning high resolution photo of a giant tortoise",
+				});
+				expect(res).toBeInstanceOf(Blob);
+			});
+		},
+		TIMEOUT
+	);
+
 	describe.concurrent("3rd party providers", () => {
 		it("chatCompletion - fails with unsupported model", async () => {
 			expect(