Add Cohere provider

.github/workflows/test.yml

@@ -50,6 +50,7 @@ jobs:
           HF_REPLICATE_KEY: dummy
           HF_SAMBANOVA_KEY: dummy
           HF_TOGETHER_KEY: dummy
+          HF_COHERE_KEY: dummy
 
   browser:
     runs-on: ubuntu-latest

README.md

@@ -23,7 +23,7 @@ await uploadFile({
   // Can work with native File in browsers
   file: {
     path: "pytorch_model.bin",
-    content: new Blob(...) 
+    content: new Blob(...)
   }
 });
 
@@ -39,7 +39,7 @@ await inference.chatCompletion({
   ],
   max_tokens: 512,
   temperature: 0.5,
-  provider: "sambanova", // or together, fal-ai, replicate, …
+  provider: "sambanova", // or together, fal-ai, replicate, cohere …
 });
 
 await inference.textToImage({
@@ -146,12 +146,12 @@ for await (const chunk of inference.chatCompletionStream({
   console.log(chunk.choices[0].delta.content);
 }
 
-/// Using a third-party provider: 
+/// Using a third-party provider:
 await inference.chatCompletion({
   model: "meta-llama/Llama-3.1-8B-Instruct",
   messages: [{ role: "user", content: "Hello, nice to meet you!" }],
   max_tokens: 512,
-  provider: "sambanova", // or together, fal-ai, replicate, …
+  provider: "sambanova", // or together, fal-ai, replicate, cohere …
 })
 
 await inference.textToImage({
@@ -211,7 +211,7 @@ await uploadFile({
   // Can work with native File in browsers
   file: {
     path: "pytorch_model.bin",
-    content: new Blob(...) 
+    content: new Blob(...)
   }
 });
 
@@ -244,7 +244,7 @@ console.log(messages); // contains the data
 
 // or you can run the code directly, however you can't check that the code is safe to execute this way, use at your own risk.
 const messages = await agent.run("Draw a picture of a cat wearing a top hat. Then caption the picture and read it out loud.")
-console.log(messages); 
+console.log(messages);
 ```
 
 There are more features of course, check each library's README!

packages/inference/README.md

@@ -56,6 +56,7 @@ Currently, we support the following providers:
 - [Sambanova](https://sambanova.ai)
 - [Together](https://together.xyz)
 - [Blackforestlabs](https://blackforestlabs.ai)
+- [Cohere](https://cohere.com)
 
 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
@@ -80,6 +81,7 @@ Only a subset of models are supported when requesting third-party providers. You
 - [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)
+- [Cohere supported models](https://huggingface.co/api/partners/cohere/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/makeRequestOptions.ts

@@ -8,6 +8,7 @@ import { NOVITA_API_BASE_URL } from "../providers/novita";
 import { FIREWORKS_AI_API_BASE_URL } from "../providers/fireworks-ai";
 import { HYPERBOLIC_API_BASE_URL } from "../providers/hyperbolic";
 import { BLACKFORESTLABS_AI_API_BASE_URL } from "../providers/black-forest-labs";
+import { COHERE_API_BASE_URL } from "../providers/cohere";
 import type { InferenceProvider } from "../types";
 import type { InferenceTask, Options, RequestArgs } from "../types";
 import { isUrl } from "./isUrl";
@@ -256,6 +257,15 @@ function makeUrl(params: {
 			}
 			return baseUrl;
 		}
+		case "cohere": {
+			const baseUrl = shouldProxy
+				? HF_HUB_INFERENCE_PROXY_TEMPLATE.replace("{{PROVIDER}}", params.provider)
+				: COHERE_API_BASE_URL;
+			if (params.taskHint === "text-generation") {
+				return `${baseUrl}/v2/chat`;
+			}
+			return baseUrl;
+		}
 		default: {
 			const baseUrl = HF_HUB_INFERENCE_PROXY_TEMPLATE.replaceAll("{{PROVIDER}}", "hf-inference");
 			if (params.taskHint && ["feature-extraction", "sentence-similarity"].includes(params.taskHint)) {

packages/inference/src/providers/cohere.ts

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

packages/inference/src/providers/consts.ts

@@ -26,4 +26,5 @@ export const HARDCODED_MODEL_ID_MAPPING: Record<InferenceProvider, Record<ModelI
 	sambanova: {},
 	together: {},
 	novita: {},
+	cohere: {},
 };

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

@@ -3,33 +3,158 @@ import type { BaseArgs, Options } from "../../types";
 import { request } from "../custom/request";
 import type { ChatCompletionInput, ChatCompletionOutput } from "@huggingface/tasks";
 
+export type CohereTextGenerationOutputFinishReason =
+	| "COMPLETE"
+	| "STOP_SEQUENCE"
+	| "MAX_TOKENS"
+	| "TOOL_CALL"
+	| "ERROR";
+
+interface CohereChatCompletionOutput {
+	id: string;
+	finish_reason: CohereTextGenerationOutputFinishReason;
+	message: CohereMessage;
+	usage: CohereChatCompletionOutputUsage;
+	logprobs?: CohereLogprob[]; // Optional field for log probabilities
+}
+
+interface CohereMessage {
+	role: string;
+	content: Array<{
+		type: string;
+		text: string;
+	}>;
+	tool_calls?: CohereToolCall[]; // Optional field for tool calls
+}
+
+interface CohereChatCompletionOutputUsage {
+	billed_units: CohereInputOutputTokens;
+	tokens: CohereInputOutputTokens;
+}
+
+interface CohereInputOutputTokens {
+	input_tokens: number;
+	output_tokens: number;
+}
+
+interface CohereLogprob {
+	logprob: number;
+	token: string;
+	top_logprobs: CohereTopLogprob[];
+}
+
+interface CohereTopLogprob {
+	logprob: number;
+	token: string;
+}
+
+interface CohereToolCall {
+	function: CohereFunctionDefinition;
+	id: string;
+	type: string;
+}
+
+interface CohereFunctionDefinition {
+	arguments: unknown;
+	description?: string;
+	name: string;
+}
+
+function convertCohereToChatCompletionOutput(res: CohereChatCompletionOutput): ChatCompletionOutput {
+	// Create a ChatCompletionOutput object from the CohereChatCompletionOutput
+	return {
+		id: res.id,
+		created: Date.now(),
+		model: "cohere-model",
+		system_fingerprint: "cohere-fingerprint",
+		usage: {
+			completion_tokens: res.usage.tokens.output_tokens,
+			prompt_tokens: res.usage.tokens.input_tokens,
+			total_tokens: res.usage.tokens.input_tokens + res.usage.tokens.output_tokens,
+		},
+		choices: [
+			{
+				finish_reason: res.finish_reason,
+				index: 0,
+				message: {
+					role: res.message.role,
+					content: res.message.content.map((c) => c.text).join(" "),
+					tool_calls: res.message.tool_calls?.map((toolCall) => ({
+						function: {
+							arguments: toolCall.function.arguments,
+							description: toolCall.function.description,
+							name: toolCall.function.name,
+						},
+						id: toolCall.id,
+						type: toolCall.type,
+					})),
+				},
+				logprobs: res.logprobs
+					? {
+							content: res.logprobs.map((logprob) => ({
+								logprob: logprob.logprob,
+								token: logprob.token,
+								top_logprobs: logprob.top_logprobs.map((topLogprob) => ({
+									logprob: topLogprob.logprob,
+									token: topLogprob.token,
+								})),
+							})),
+					  }
+					: undefined,
+			},
+		],
+	};
+}
+
 /**
  * Use the chat completion endpoint to generate a response to a prompt, using OpenAI message completion API no stream
  */
 export async function chatCompletion(
 	args: BaseArgs & ChatCompletionInput,
 	options?: Options
 ): Promise<ChatCompletionOutput> {
-	const res = await request<ChatCompletionOutput>(args, {
-		...options,
-		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 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) {
-		throw new InferenceOutputError("Expected ChatCompletionOutput");
+	if (args.provider === "cohere") {
+		const res = await request<CohereChatCompletionOutput>(args, {
+			...options,
+			taskHint: "text-generation",
+			chatCompletion: true,
+		});
+
+		const isValidOutput =
+			typeof res === "object" &&
+			typeof res?.id === "string" &&
+			typeof res?.finish_reason === "string" &&
+			typeof res?.message === "object" &&
+			Array.isArray(res?.message.content) &&
+			typeof res?.usage === "object";
+
+		if (!isValidOutput) {
+			throw new InferenceOutputError("Expected CohereChatCompletionOutput");
+		}
+
+		return convertCohereToChatCompletionOutput(res);
+	} else {
+		const res = await request<ChatCompletionOutput>(args, {
+			...options,
+			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 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) {
+			throw new InferenceOutputError("Expected ChatCompletionOutput");
+		}
+		return res;
 	}
-	return res;
 }