[Inference Providers] Fix structured output (#1579)

equivalent PR to
huggingface/huggingface_hub#3082.
`sambanova` and `nebius` don’t fully follow OpenAI’s spec for the
`response_format` field. This PR adds internal mappings for each
provider. Seems like `fireworks-ai` is now following OpenAI specs for
this field (double checked their docs as well).

---------

Co-authored-by: SBrandeis <[email protected]>

Author: hanouticelina

packages/inference/src/providers/nebius.ts

@@ -25,6 +25,7 @@ import {
 	type TextToImageTaskHelper,
 } from "./providerHelper.js";
 import { InferenceClientProviderOutputError } from "../errors.js";
+import type { ChatCompletionInput } from "../../../tasks/dist/commonjs/index.js";
 
 const NEBIUS_API_BASE_URL = "https://api.studio.nebius.ai";
 
@@ -50,6 +51,17 @@ export class NebiusConversationalTask extends BaseConversationalTask {
 	constructor() {
 		super("nebius", NEBIUS_API_BASE_URL);
 	}
+
+	override preparePayload(params: BodyParams<ChatCompletionInput>): Record<string, unknown> {
+		const payload = super.preparePayload(params) as Record<string, unknown>;
+
+		const responseFormat = params.args.response_format;
+		if (responseFormat?.type === "json_schema" && responseFormat.json_schema?.schema) {
+			payload["guided_json"] = responseFormat.json_schema.schema;
+		}
+
+		return payload;
+	}
 }
 
 export class NebiusTextGenerationTask extends BaseTextGenerationTask {

packages/inference/src/providers/sambanova.ts

@@ -19,11 +19,25 @@ import type { BodyParams } from "../types.js";
 import type { FeatureExtractionTaskHelper } from "./providerHelper.js";
 import { BaseConversationalTask, TaskProviderHelper } from "./providerHelper.js";
 import { InferenceClientProviderOutputError } from "../errors.js";
+import type { ChatCompletionInput } from "../../../tasks/dist/commonjs/index.js";
 
 export class SambanovaConversationalTask extends BaseConversationalTask {
 	constructor() {
 		super("sambanova", "https://api.sambanova.ai");
 	}
+
+	override preparePayload(params: BodyParams<ChatCompletionInput>): Record<string, unknown> {
+		const responseFormat = params.args.response_format;
+
+		if (responseFormat?.type === "json_schema" && responseFormat.json_schema) {
+			if (responseFormat.json_schema.strict ?? true) {
+				responseFormat.json_schema.strict = false;
+			}
+		}
+		const payload = super.preparePayload(params) as Record<string, unknown>;
+
+		return payload;
+	}
 }
 
 export class SambanovaFeatureExtractionTask extends TaskProviderHelper implements FeatureExtractionTaskHelper {

packages/inference/src/providers/together.ts

@@ -24,6 +24,7 @@ import {
 	type TextToImageTaskHelper,
 } from "./providerHelper.js";
 import { InferenceClientProviderOutputError } from "../errors.js";
+import type { ChatCompletionInput } from "../../../tasks/dist/commonjs/index.js";
 
 const TOGETHER_API_BASE_URL = "https://api.together.xyz";
 
@@ -47,6 +48,22 @@ export class TogetherConversationalTask extends BaseConversationalTask {
 	constructor() {
 		super("together", TOGETHER_API_BASE_URL);
 	}
+
+	override preparePayload(params: BodyParams<ChatCompletionInput>): Record<string, unknown> {
+		const payload = super.preparePayload(params);
+		const response_format = payload.response_format as
+			| { type: "json_schema"; json_schema: { schema: unknown } }
+			| undefined;
+
+		if (response_format?.type === "json_schema" && response_format?.json_schema?.schema) {
+			payload.response_format = {
+				type: "json_schema",
+				schema: response_format.json_schema.schema,
+			};
+		}
+
+		return payload;
+	}
 }
 
 export class TogetherTextGenerationTask extends BaseTextGenerationTask {