Use makeRequestOptions to generate inference snippets (#1273)

The broader goal of this PR is to use `makeRequestOptions` from JS
InferenceClient in order to get all the implementation details (correct
URL, correct authorization header, correct payload, etc.). JS
InferenceClient is supposed to be the ground truth in this case.

**In practice:**
- fixed `makeUrl` when chatCompletion + image-text-to-text (review
[here](https://github.com/huggingface/huggingface.js/pull/1273/files#diff-a6509c908fd0fb05fdbd3803492d6e9e2570d6dff2a21db575a76b26bff4d565)
+ other providers)
- fixed wrong URL in `openai` python snippet (e.g.
[here](https://github.com/huggingface/huggingface.js/pull/1273/files#diff-338b930b960057f85d0d5dd27032b73cd4834a6a7c7ce4db60af19395b8e56f9),
[here](https://github.com/huggingface/huggingface.js/pull/1273/files#diff-a253bcdfdf33df1ac53a4051a8ce7bb047a99f32618c048754d617cb55815c14))
- fixed DQA `requests` snippet
([here](https://github.com/huggingface/huggingface.js/pull/1273/files#diff-3a47136351b4572144f2fd42a2518da9be108b66fd5dca392d9d899a125b02d9))

**Technically, this PR:**
- splits `makeRequestOptions` in two parts: the async part that does the
model ID resolution (depending on task+provider) and the sync part which
generates the url, headers, body, etc. For snippets we only need the
second part which is a sync call. => new (internal) method
`makeRequestOptionsFromResolvedModel`
- moves most of the logic inside `snippetGenerator`
- logic is: _get inputs_ => _make request options_ => _prepare template
data_ => _iterate over clients_ => _generate snippets_
- **Next:** now that the logic is unified, adapting cURL and JS to use
the same logic should be fairly easy (e.g. "just" need to create the
jinja templates)
- => final goal is to handle all languages/clients/providers with the
same code and swap the templates
- update most providers to use `/chat/completions` endpoint when
`chatCompletion` is enabled
- Previously we were also checking that task is `text-generation` => now
we will also use /chat/completion on "image-text-to-text" models
- that was mostly a bug in existing codebase => detected it thanks to
the snippets
- updated `./packages/inference/package.json` to allow dev mode. Now
running `pnpm run dev` in `@inference` makes it much easier to work with
`@tasks-gen` (no need to rebuild each time I make a change)

---


**EDIT:** ~there is definitely a breaking change in how I handle the
`makeRequestOptions` split (hence the broken CI). Will fix this.~ =>
fixed.

---------

Co-authored-by: Simon Brandeis <[email protected]>

Author: Wauplin

packages/inference/package.json

@@ -56,7 +56,8 @@
 		"prepublishOnly": "pnpm run build",
 		"test": "vitest run --config vitest.config.mts",
 		"test:browser": "vitest run --browser.name=chrome --browser.headless --config vitest.config.mts",
-		"check": "tsc"
+		"check": "tsc",
+		"dev": "tsup src/index.ts --format cjs,esm --watch"
 	},
 	"dependencies": {
 		"@huggingface/tasks": "workspace:^",

packages/inference/src/lib/makeRequestOptions.ts

@@ -45,7 +45,8 @@ const providerConfigs: Record<InferenceProvider, ProviderConfig> = {
 };
 
 /**
- * Helper that prepares request arguments
+ * Helper that prepares request arguments.
+ * This async version handle the model ID resolution step.
  */
 export async function makeRequestOptions(
 	args: RequestArgs & {
@@ -56,17 +57,15 @@ export async function makeRequestOptions(
 		/** In most cases (unless we pass a endpointUrl) we know the task */
 		task?: InferenceTask;
 		chatCompletion?: boolean;
-		/* Used internally to generate inference snippets (in which case model mapping is done separately) */
-		skipModelIdResolution?: boolean;
 	}
 ): Promise<{ url: string; info: RequestInit }> {
-	const { accessToken, endpointUrl, provider: maybeProvider, model: maybeModel, ...remainingArgs } = args;
+	const { provider: maybeProvider, model: maybeModel } = args;
 	const provider = maybeProvider ?? "hf-inference";
 	const providerConfig = providerConfigs[provider];
+	const { task, chatCompletion } = options ?? {};
 
-	const { includeCredentials, task, chatCompletion, signal, skipModelIdResolution } = options ?? {};
-
-	if (endpointUrl && provider !== "hf-inference") {
+	// Validate inputs
+	if (args.endpointUrl && provider !== "hf-inference") {
 		throw new Error(`Cannot use endpointUrl with a third-party provider.`);
 	}
 	if (maybeModel && isUrl(maybeModel)) {
@@ -81,19 +80,43 @@ export async function makeRequestOptions(
 	if (providerConfig.clientSideRoutingOnly && !maybeModel) {
 		throw new Error(`Provider ${provider} requires a model ID to be passed directly.`);
 	}
+
 	// eslint-disable-next-line @typescript-eslint/no-non-null-assertion
 	const hfModel = maybeModel ?? (await loadDefaultModel(task!));
-	const model = skipModelIdResolution
-		? hfModel
-		: providerConfig.clientSideRoutingOnly
-		  ? // eslint-disable-next-line @typescript-eslint/no-non-null-assertion
-		    removeProviderPrefix(maybeModel!, provider)
-		  : // For closed-models API providers, one needs to pass the model ID directly (e.g. "gpt-3.5-turbo")
-		    await getProviderModelId({ model: hfModel, provider }, args, {
-					task,
-					chatCompletion,
-					fetch: options?.fetch,
-		    });
+	const resolvedModel = providerConfig.clientSideRoutingOnly
+		? // eslint-disable-next-line @typescript-eslint/no-non-null-assertion
+		  removeProviderPrefix(maybeModel!, provider)
+		: await getProviderModelId({ model: hfModel, provider }, args, {
+				task,
+				chatCompletion,
+				fetch: options?.fetch,
+		  });
+
+	// Use the sync version with the resolved model
+	return makeRequestOptionsFromResolvedModel(resolvedModel, args, options);
+}
+
+/**
+ * Helper that prepares request arguments. - for internal use only
+ * This sync version skips the model ID resolution step
+ */
+export function makeRequestOptionsFromResolvedModel(
+	resolvedModel: string,
+	args: RequestArgs & {
+		data?: Blob | ArrayBuffer;
+		stream?: boolean;
+	},
+	options?: Options & {
+		task?: InferenceTask;
+		chatCompletion?: boolean;
+	}
+): { url: string; info: RequestInit } {
+	const { accessToken, endpointUrl, provider: maybeProvider, model, ...remainingArgs } = args;
+
+	const provider = maybeProvider ?? "hf-inference";
+	const providerConfig = providerConfigs[provider];
+
+	const { includeCredentials, task, chatCompletion, signal } = options ?? {};
 
 	const authMethod = (() => {
 		if (providerConfig.clientSideRoutingOnly) {
@@ -123,7 +146,7 @@ export async function makeRequestOptions(
 					authMethod !== "provider-key"
 						? HF_HUB_INFERENCE_PROXY_TEMPLATE.replace("{{PROVIDER}}", provider)
 						: providerConfig.baseUrl,
-				model,
+				model: resolvedModel,
 				chatCompletion,
 				task,
 		  });
@@ -154,7 +177,7 @@ export async function makeRequestOptions(
 		: JSON.stringify(
 				providerConfig.makeBody({
 					args: remainingArgs as Record<string, unknown>,
-					model,
+					model: resolvedModel,
 					task,
 					chatCompletion,
 				})

packages/inference/src/providers/fireworks-ai.ts

@@ -30,7 +30,7 @@ const makeHeaders = (params: HeaderParams): Record<string, string> => {
 };
 
 const makeUrl = (params: UrlParams): string => {
-	if (params.task === "text-generation" && params.chatCompletion) {
+	if (params.chatCompletion) {
 		return `${params.baseUrl}/inference/v1/chat/completions`;
 	}
 	return `${params.baseUrl}/inference`;

packages/inference/src/providers/hf-inference.ts

@@ -29,7 +29,7 @@ const makeUrl = (params: UrlParams): string => {
 		/// when deployed on hf-inference, those two tasks are automatically compatible with one another.
 		return `${params.baseUrl}/pipeline/${params.task}/${params.model}`;
 	}
-	if (params.task === "text-generation" && params.chatCompletion) {
+	if (params.chatCompletion) {
 		return `${params.baseUrl}/models/${params.model}/v1/chat/completions`;
 	}
 	return `${params.baseUrl}/models/${params.model}`;

packages/inference/src/providers/nebius.ts

@@ -33,10 +33,10 @@ const makeUrl = (params: UrlParams): string => {
 	if (params.task === "text-to-image") {
 		return `${params.baseUrl}/v1/images/generations`;
 	}
+	if (params.chatCompletion) {
+		return `${params.baseUrl}/v1/chat/completions`;
+	}
 	if (params.task === "text-generation") {
-		if (params.chatCompletion) {
-			return `${params.baseUrl}/v1/chat/completions`;
-		}
 		return `${params.baseUrl}/v1/completions`;
 	}
 	return params.baseUrl;

packages/inference/src/providers/novita.ts

@@ -30,10 +30,9 @@ const makeHeaders = (params: HeaderParams): Record<string, string> => {
 };
 
 const makeUrl = (params: UrlParams): string => {
-	if (params.task === "text-generation") {
-		if (params.chatCompletion) {
-			return `${params.baseUrl}/v3/openai/chat/completions`;
-		}
+	if (params.chatCompletion) {
+		return `${params.baseUrl}/v3/openai/chat/completions`;
+	} else if (params.task === "text-generation") {
 		return `${params.baseUrl}/v3/openai/completions`;
 	} else if (params.task === "text-to-video") {
 		return `${params.baseUrl}/v3/hf/${params.model}`;

packages/inference/src/providers/sambanova.ts

@@ -30,7 +30,7 @@ const makeHeaders = (params: HeaderParams): Record<string, string> => {
 };
 
 const makeUrl = (params: UrlParams): string => {
-	if (params.task === "text-generation" && params.chatCompletion) {
+	if (params.chatCompletion) {
 		return `${params.baseUrl}/v1/chat/completions`;
 	}
 	return params.baseUrl;

packages/inference/src/providers/together.ts

@@ -33,10 +33,10 @@ const makeUrl = (params: UrlParams): string => {
 	if (params.task === "text-to-image") {
 		return `${params.baseUrl}/v1/images/generations`;
 	}
+	if (params.chatCompletion) {
+		return `${params.baseUrl}/v1/chat/completions`;
+	}
 	if (params.task === "text-generation") {
-		if (params.chatCompletion) {
-			return `${params.baseUrl}/v1/chat/completions`;
-		}
 		return `${params.baseUrl}/v1/completions`;
 	}
 	return params.baseUrl;