Featherless provider: max_tokens (#1474)

pohnean · SBrandeis · web-flow · commit 2413423d3bf2 · 2025-05-23T13:31:47.000+02:00
This PR fixes the parameters that are passed into the featherless api,
in particular the max_token.

---------

Co-authored-by: Simon Brandeis &lt;33657802+SBrandeis@users.noreply.github.com&gt;
Co-authored-by: SBrandeis &lt;simon@huggingface.co&gt;
diff --git a/packages/inference/src/providers/featherless-ai.ts b/packages/inference/src/providers/featherless-ai.ts
@@ -7,6 +7,7 @@ import type {
 import { InferenceOutputError } from "../lib/InferenceOutputError.js";
 import type { BodyParams } from "../types.js";
 import { BaseConversationalTask, BaseTextGenerationTask } from "./providerHelper.js";
+import { omit } from "../utils/omit.js";
 
 interface FeatherlessAITextCompletionOutput extends Omit<ChatCompletionOutput, "choices"> {
 	choices: Array<{
@@ -33,9 +34,14 @@ export class FeatherlessAITextGenerationTask extends BaseTextGenerationTask {
 
 	override preparePayload(params: BodyParams<TextGenerationInput>): Record<string, unknown> {
 		return {
-			...params.args,
-			...params.args.parameters,
 			model: params.model,
+			...omit(params.args, ["inputs", "parameters"]),
+			...(params.args.parameters
+				? {
+						max_tokens: params.args.parameters.max_new_tokens,
+						...omit(params.args.parameters, "max_new_tokens"),
+				  }
+				: undefined),
 			prompt: params.args.inputs,
 		};
 	}
