[Inference snippets] Templated snippets for inference snippet generation (#1255)

### What does this PR do?
- add `@huggingface/jinja` as a dependency for `@huggingface/inference`
- heavily refactor
[`./packages/inference/src/snippets/python.ts`](https://github.com/huggingface/huggingface.js/blob/ddd62b60d0a858dc808b6dd7b3589977681f5a8b/packages/inference/src/snippets/python.ts)
to use templates instead of inline strings
- add some snippet templates under
`packages/inference/src/snippets/templates/{language}/{client}/{templateName}.jinja`
    - `language` is Python
    - `client` is e.g. huggingface_hub, requests, etc.
- `templateName` is related to task name e.g. textToImage,
conversationalStream, etc.

=> overall goal is to make it rather easy to update the snippets for a
given task + client without entering the JS code
=> if conclusive, we can extend to cURL / JS snippets in a follow-up PR 

This PR also includes a lot of small fixes (usually whitespace<>tabs
inconsistencies) + added some tests. I did not change anything major in
the generated snippets.

### What to review?

IMO, no need to review all templates. Better to check the generated
snippets + the main JS file directly
(https://github.com/huggingface/huggingface.js/blob/ddd62b60d0a858dc808b6dd7b3589977681f5a8b/packages/inference/src/snippets/python.ts).
Sorry for the huge PR but I do believe it is a necessary move moving
forward 🤗

---------

Co-authored-by: Simon Brandeis <[email protected]>
Co-authored-by: coyotte508 <[email protected]>

Author: 3 people

.github/workflows/test.yml

@@ -105,7 +105,7 @@ jobs:
         run: |
           sleep 3
           pnpm i --filter root --filter inference... --filter hub... --filter tasks-gen --frozen-lockfile
-          pnpm --filter inference --filter hub --filter tasks publish --force --no-git-checks --registry http://localhost:4874/
+          pnpm --filter inference --filter hub --filter tasks --filter jinja publish --force --no-git-checks --registry http://localhost:4874/
 
       - name: E2E test - test yarn install
         working-directory: e2e/ts
@@ -136,7 +136,7 @@ jobs:
           deno-version: vx.x.x
       - name: E2E test - deno import from npm
         working-directory: e2e/deno
-        run: deno run --allow-net --allow-env=HF_TOKEN index.ts
+        run: deno run --allow-read --allow-net --allow-env=HF_TOKEN index.ts
         env:
           NPM_CONFIG_REGISTRY: http://localhost:4874/
           HF_TOKEN: ${{ secrets.HF_TOKEN }}

.vscode/settings.json

@@ -17,5 +17,5 @@
 	"search.exclude": {
 		"**/dist": true
 	},
-	"typescript.tsdk": "node_modules/typescript/lib"
+	"typescript.tsdk": "node_modules/typescript/lib",
 }

packages/inference/package.json

@@ -33,9 +33,16 @@
 	"main": "./dist/index.cjs",
 	"module": "./dist/index.js",
 	"exports": {
-		"types": "./dist/src/index.d.ts",
-		"require": "./dist/index.cjs",
-		"import": "./dist/index.js"
+		".": {
+			"types": "./dist/src/index.d.ts",
+			"require": "./dist/index.cjs",
+			"import": "./dist/index.js"
+		}
+	},
+	"browser": {
+		"./src/snippets/index.js": false,
+		"./dist/index.js": "./dist/browser/index.js",
+		"./dist/index.mjs": "./dist/browser/index.mjs"
 	},
 	"type": "module",
 	"scripts": {
@@ -52,7 +59,8 @@
 		"check": "tsc"
 	},
 	"dependencies": {
-		"@huggingface/tasks": "workspace:^"
+		"@huggingface/tasks": "workspace:^",
+		"@huggingface/jinja": "workspace:^"
 	},
 	"devDependencies": {
 		"@types/node": "18.13.0"

packages/inference/pnpm-lock.yaml

@@ -2,6 +2,6 @@ export { InferenceClient, InferenceClientEndpoint, HfInference } from "./Inferen
 export { InferenceOutputError } from "./lib/InferenceOutputError";
 export * from "./types";
 export * from "./tasks";
-
 import * as snippets from "./snippets/index.js";
+
 export { snippets };

packages/inference/src/index.ts

@@ -56,13 +56,15 @@ 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 ?? "hf-inference";
 	const providerConfig = providerConfigs[provider];
 
-	const { includeCredentials, task, chatCompletion, signal } = options ?? {};
+	const { includeCredentials, task, chatCompletion, signal, skipModelIdResolution } = options ?? {};
 
 	if (endpointUrl && provider !== "hf-inference") {
 		throw new Error(`Cannot use endpointUrl with a third-party provider.`);
@@ -81,15 +83,17 @@ export async function makeRequestOptions(
 	}
 	// eslint-disable-next-line @typescript-eslint/no-non-null-assertion
 	const hfModel = maybeModel ?? (await loadDefaultModel(task!));
-	const model = 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 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 authMethod = (() => {
 		if (providerConfig.clientSideRoutingOnly) {