Fireworks AI Conversational Models (#1177)

Added support for inference for Fireworks AI chat/instruct models

---------

Co-authored-by: Julien Chaumond <[email protected]>
Co-authored-by: Lucain <[email protected]>

Author: 3 people

.github/workflows/test.yml

@@ -45,6 +45,7 @@ jobs:
           HF_REPLICATE_KEY: dummy
           HF_SAMBANOVA_KEY: dummy
           HF_TOGETHER_KEY: dummy
+          HF_FIREWORKS_KEY: dummy
 
   browser:
     runs-on: ubuntu-latest
@@ -85,6 +86,7 @@ jobs:
           HF_REPLICATE_KEY: dummy
           HF_SAMBANOVA_KEY: dummy
           HF_TOGETHER_KEY: dummy
+          HF_FIREWORKS_KEY: dummy
 
   e2e:
     runs-on: ubuntu-latest
@@ -152,3 +154,4 @@ jobs:
           HF_REPLICATE_KEY: dummy
           HF_SAMBANOVA_KEY: dummy
           HF_TOGETHER_KEY: dummy
+          HF_FIREWORKS_KEY: dummy

packages/inference/README.md

@@ -46,7 +46,12 @@ Your access token should be kept private. If you need to protect it in front-end
 
 You can send inference requests to third-party providers with the inference client.
 
-Currently, we support the following providers: [Fal.ai](https://fal.ai), [Replicate](https://replicate.com), [Together](https://together.xyz) and [Sambanova](https://sambanova.ai).
+Currently, we support the following providers:
+- [Fal.ai](https://fal.ai)
+- [Fireworks AI](https://fireworks.ai)
+- [Replicate](https://replicate.com)
+- [Sambanova](https://sambanova.ai)
+- [Together](https://together.xyz)
 
 To send requests to a third-party provider, you have to pass the `provider` parameter to the inference function. Make sure your request is authenticated with an access token.
 ```ts
@@ -64,10 +69,11 @@ When authenticated with a Hugging Face access token, the request is routed throu
 When authenticated with a third-party provider key, the request is made directly against that provider's inference API.
 
 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](./src/providers/fal-ai.ts)
-- [Replicate supported models](./src/providers/replicate.ts)
-- [Sambanova supported models](./src/providers/sambanova.ts)
-- [Together supported models](./src/providers/together.ts)
+- [Fal.ai supported models](https://huggingface.co/api/partners/fal-ai/models)
+- [Fireworks AI supported models](https://huggingface.co/api/partners/fireworks-ai/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. 

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/lib/makeRequestOptions.ts

@@ -3,6 +3,7 @@ import { FAL_AI_API_BASE_URL } from "../providers/fal-ai";
 import { REPLICATE_API_BASE_URL } from "../providers/replicate";
 import { SAMBANOVA_API_BASE_URL } from "../providers/sambanova";
 import { TOGETHER_API_BASE_URL } from "../providers/together";
+import { FIREWORKS_AI_API_BASE_URL } from "../providers/fireworks-ai";
 import type { InferenceProvider } from "../types";
 import type { InferenceTask, Options, RequestArgs } from "../types";
 import { isUrl } from "./isUrl";
@@ -208,6 +209,15 @@ function makeUrl(params: {
 			}
 			return baseUrl;
 		}
+		case "fireworks-ai": {
+			const baseUrl = shouldProxy
+				? HF_HUB_INFERENCE_PROXY_TEMPLATE.replace("{{PROVIDER}}", params.provider)
+				: FIREWORKS_AI_API_BASE_URL;
+			if (params.taskHint === "text-generation" && params.chatCompletion) {
+				return `${baseUrl}/v1/chat/completions`;
+			}
+			return baseUrl;
+		}
 		default: {
 			const baseUrl = HF_HUB_INFERENCE_PROXY_TEMPLATE.replaceAll("{{PROVIDER}}", "hf-inference");
 			const url = params.forceTask

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/src/providers/fireworks-ai.ts

@@ -0,0 +1,18 @@
+export const FIREWORKS_AI_API_BASE_URL = "https://api.fireworks.ai/inference";
+
+/**
+ * See the registered mapping of HF model ID => Fireworks model ID here:
+ *
+ * https://huggingface.co/api/partners/fireworks/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 Fireworks 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 Fireworks, please open an issue on the present repo
+ * and we will tag Fireworks team members.
+ *
+ * Thanks!
+ */

packages/inference/src/types.ts

@@ -44,7 +44,14 @@ export interface Options {
 
 export type InferenceTask = Exclude<PipelineType, "other">;
 
-export const INFERENCE_PROVIDERS = ["fal-ai", "replicate", "sambanova", "together", "hf-inference"] as const;
+export const INFERENCE_PROVIDERS = [
+	"fal-ai",
+	"fireworks-ai",
+	"hf-inference",
+	"replicate",
+	"sambanova",
+	"together",
+] as const;
 export type InferenceProvider = (typeof INFERENCE_PROVIDERS)[number];
 
 export interface BaseArgs {

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;
@@ -1077,4 +1078,51 @@ describe.concurrent("HfInference", () => {
 			);
 		});
 	});
+
+	describe.concurrent(
+		"Fireworks",
+		() => {
+			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",
+					provider: "fireworks-ai",
+					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).toContain("two");
+				}
+			});
+
+			it("chatCompletion stream", async () => {
+				const stream = client.chatCompletionStream({
+					model: "deepseek-ai/DeepSeek-R1",
+					provider: "fireworks-ai",
+					messages: [{ role: "user", content: "Say this is a test" }],
+					stream: true,
+				}) as AsyncGenerator<ChatCompletionStreamOutput>;
+
+				let fullResponse = "";
+				for await (const chunk of stream) {
+					if (chunk.choices && chunk.choices.length > 0) {
+						const content = chunk.choices[0].delta?.content;
+						if (content) {
+							fullResponse += content;
+						}
+					}
+				}
+
+				// Verify we got a meaningful response
+				expect(fullResponse).toBeTruthy();
+				expect(fullResponse.length).toBeGreaterThan(0);
+			});
+		},
+		TIMEOUT
+	);
 });