New way of injecting mapping into tests + record VCR

Author: julien-c

packages/inference/src/lib/getProviderModelId.ts

@@ -30,8 +30,8 @@ export async function getProviderModelId(
 		options.taskHint === "text-generation" && options.chatCompletion ? "conversational" : options.taskHint;
 
 	// A dict called HARDCODED_MODEL_ID_MAPPING takes precedence in all cases (useful for dev purposes)
-	if (HARDCODED_MODEL_ID_MAPPING[params.model]) {
-		return HARDCODED_MODEL_ID_MAPPING[params.model];
+	if (HARDCODED_MODEL_ID_MAPPING[params.provider]?.[params.model]) {
+		return HARDCODED_MODEL_ID_MAPPING[params.provider][params.model];
 	}
 
 	let inferenceProviderMapping: InferenceProviderMapping | null;

packages/inference/src/providers/consts.ts

@@ -1,15 +1,25 @@
-import type { ModelId } from "../types";
+import type { InferenceProvider } from "../types";
+import { type ModelId } from "../types";
 
 type ProviderId = string;
-
 /**
  * If you want to try to run inference for a new model locally before it's registered on huggingface.co
  * for a given Inference Provider,
  * you can add it to the following dictionary, for dev purposes.
+ *
+ * We also inject into this dictionary from tests.
  */
-export const HARDCODED_MODEL_ID_MAPPING: Record<ModelId, ProviderId> = {
+export const HARDCODED_MODEL_ID_MAPPING: Record<InferenceProvider, Record<ModelId, ProviderId>> = {
 	/**
 	 * "HF model ID" => "Model ID on Inference Provider's side"
+	 *
+	 * Example:
+	 * "Qwen/Qwen2.5-Coder-32B-Instruct": "Qwen2.5-Coder-32B-Instruct",
 	 */
-	// "Qwen/Qwen2.5-Coder-32B-Instruct": "Qwen2.5-Coder-32B-Instruct",
+	"fal-ai": {},
+	"fireworks-ai": {},
+	"hf-inference": {},
+	replicate: {},
+	sambanova: {},
+	together: {},
 };

packages/inference/test/HfInference.spec.ts

@@ -6,6 +6,7 @@ import { chatCompletion, HfInference } from "../src";
 import { textToVideo } from "../src/tasks/cv/textToVideo";
 import { readTestFile } from "./test-files";
 import "./vcr";
+import { HARDCODED_MODEL_ID_MAPPING } from "../src/providers/consts";
 
 const TIMEOUT = 60000 * 3;
 const env = import.meta.env;
@@ -1083,6 +1084,10 @@ describe.concurrent("HfInference", () => {
 		() => {
 			const client = new HfInference(env.HF_FIREWORKS_KEY);
 
+			HARDCODED_MODEL_ID_MAPPING["fireworks-ai"] = {
+				"deepseek-ai/DeepSeek-R1": "accounts/fireworks/models/deepseek-r1",
+			};
+
 			it("chatCompletion", async () => {
 				const res = await client.chatCompletion({
 					model: "deepseek-ai/DeepSeek-R1",