➕ Add @huggingface/tasks as a dep of @huggingface/inference

.github/workflows/inference-publish.yml

@@ -52,6 +52,44 @@ jobs:
           git add ../..
           git commit -m "🔖 @hugginface/inference $BUMPED_VERSION"
           git tag "inference-v$BUMPED_VERSION"
+
+      - name: Make sure that the latest version of @huggingface/tasks is consistent with the local version
+        run: |
+          LOCAL_TASKS_VERSION=$(node -p "require('./package.json').version")
+          REMOTE_TASKS_VERSION=$(npm view @huggingface/tasks version)
+
+          # If the versions are different, error 
+          if [ "$LOCAL_TASKS_VERSION" != "$REMOTE_TASKS_VERSION" ]; then
+            echo "Error: The local @huggingface/tasks package version ($LOCAL_TASKS_VERSION) differs from the remote version ($REMOTE_TASKS_VERSION). Release halted."
+            exit 1
+          fi
+
+          npm pack @huggingface/tasks
+          mv huggingface-tasks-$LOCAL_TASKS_VERSION.tgz tasks-local.tgz
+
+          npm pack @huggingface/tasks@$REMOTE_TASKS_VERSION
+          mv huggingface-tasks-$REMOTE_TASKS_VERSION.tgz tasks-remote.tgz
+
+          # Compute checksum of local tar. We need to extract both tar since the remote compression might be different
+          tar -xf tasks-local.tgz
+          LOCAL_CHECKSUM=$(cd package && tar --mtime='1970-01-01' --mode=755 -cf - . | sha256sum | cut -d' ' -f1)
+          echo "Local package checksum: $LOCAL_CHECKSUM"
+
+          rm -Rf package
+
+          tar -xf tasks-remote.tgz
+          REMOTE_CHECKSUM=$(cd package && tar --mtime='1970-01-01' --mode=755 -cf - . | sha256sum | cut -d' ' -f1)
+          echo "Remote package checksum: $REMOTE_CHECKSUM"
+
+          rm -Rf package
+
+          if [ "$LOCAL_CHECKSUM" != "$REMOTE_CHECKSUM" ]; then
+            echo "Checksum Verification Failed: The local @huggingface/tasks package differs from the remote version. Release halted. Local Checksum: $LOCAL_CHECKSUM, Remote Checksum: $REMOTE_CHECKSUM"
+            exit 1
+          fi
+          echo "Checksum Verification Successful: The local and remote @huggingface/tasks packages are consistent. Proceeding with the @huggingface/widgets package release. Local Checksum: $LOCAL_CHECKSUM, Remote Checksum: $REMOTE_CHECKSUM."
+        working-directory: packages/tasks
+
       - run: pnpm publish --no-git-checks .
         env:
           NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}

.github/workflows/test.yml

@@ -59,7 +59,7 @@ jobs:
         run: |
           sleep 3
           pnpm i --filter inference... --filter hub... --frozen-lockfile
-          pnpm --filter inference --filter hub publish --force --no-git-checks --registry http://localhost:4874/
+          pnpm --filter inference --filter hub --filter tasks publish --force --no-git-checks --registry http://localhost:4874/
 
       - name: E2E test - test yarn install
         working-directory: e2e/ts

packages/inference/package.json

@@ -40,7 +40,7 @@
 	"type": "module",
 	"scripts": {
 		"build": "tsup src/index.ts --format cjs,esm --clean && pnpm run dts",
-		"dts": "tsx scripts/generate-dts.ts",
+		"dts": "tsx scripts/generate-dts.ts && tsc --noEmit dist/index.d.ts",
 		"lint": "eslint --quiet --fix --ext .cjs,.ts .",
 		"lint:check": "eslint --ext .cjs,.ts .",
 		"format": "prettier --write .",
@@ -51,8 +51,10 @@
 		"test:browser": "vitest run --browser.name=chrome --browser.headless --config vitest.config.mts",
 		"check": "tsc"
 	},
+	"dependencies": {
+		"@huggingface/tasks": "workspace:^"
+	},
 	"devDependencies": {
-		"@huggingface/tasks": "workspace:^",
 		"@types/node": "18.13.0"
 	},
 	"resolutions": {}

packages/inference/scripts/generate-dts.ts

@@ -3,6 +3,8 @@
 import { readFileSync, writeFileSync, appendFileSync, readdirSync } from "node:fs";
 import { TASKS_DATA } from "@huggingface/tasks";
 
+const taskImports = new Set<string>();
+
 const tasks = Object.keys(TASKS_DATA)
 	.sort()
 	.filter((task) => task !== "other");
@@ -36,6 +38,16 @@ for (const dir of dirs) {
 
 		const fileContent = readFileSync(`./src/tasks/${dir}/${file}`, "utf-8");
 
+		// detect imports from @huggingface/tasks
+		for (const imports of fileContent.matchAll(/import type {(.*)} from "@huggingface\/tasks";/g)) {
+			// Convert A, B, C to ["A", "B", "C"]
+			const imported = imports[1].split(",").map((x) => x.trim());
+
+			for (const imp of imported) {
+				taskImports.add(imp);
+			}
+		}
+
 		for (const type of extractTypesAndInterfaces(fileContent)) {
 			appendFileSync("./dist/index.d.ts", type + "\n");
 		}
@@ -87,6 +99,13 @@ appendFileSync(
 		"\n}\n"
 );
 
+// Prepend import from @huggingface/tasks
+writeFileSync(
+	"./dist/index.d.ts",
+	`import type { ${[...taskImports].join(", ")} } from "@huggingface/tasks";\n` +
+		readFileSync("./dist/index.d.ts", "utf-8")
+);
+
 function* extractTypesAndInterfaces(fileContent: string): Iterable<string> {
 	let index = 0;
 

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

@@ -1,209 +1,9 @@
+import type { TextGenerationInput, TextGenerationOutput } from "@huggingface/tasks";
 import { InferenceOutputError } from "../../lib/InferenceOutputError";
 import type { BaseArgs, Options } from "../../types";
 import { request } from "../custom/request";
 
-/**
- * Inputs for Text Generation inference
- */
-export interface TextGenerationInput {
-	/**
-	 * The text to initialize generation with
-	 */
-	inputs: string;
-	/**
-	 * Additional inference parameters
-	 */
-	parameters?: TextGenerationParameters;
-	/**
-	 * Whether to stream output tokens
-	 */
-	stream?: boolean;
-	[property: string]: unknown;
-}
-
-/**
- * Additional inference parameters
- *
- * Additional inference parameters for Text Generation
- */
-export interface TextGenerationParameters {
-	/**
-	 * The number of sampling queries to run. Only the best one (in terms of total logprob) will
-	 * be returned.
-	 */
-	best_of?: number;
-	/**
-	 * Whether or not to output decoder input details
-	 */
-	decoder_input_details?: boolean;
-	/**
-	 * Whether or not to output details
-	 */
-	details?: boolean;
-	/**
-	 * Whether to use logits sampling instead of greedy decoding when generating new tokens.
-	 */
-	do_sample?: boolean;
-	/**
-	 * The maximum number of tokens to generate.
-	 */
-	max_new_tokens?: number;
-	/**
-	 * The parameter for repetition penalty. A value of 1.0 means no penalty. See [this
-	 * paper](https://hf.co/papers/1909.05858) for more details.
-	 */
-	repetition_penalty?: number;
-	/**
-	 * Whether to prepend the prompt to the generated text.
-	 */
-	return_full_text?: boolean;
-	/**
-	 * The random sampling seed.
-	 */
-	seed?: number;
-	/**
-	 * Stop generating tokens if a member of `stop_sequences` is generated.
-	 */
-	stop_sequences?: string[];
-	/**
-	 * The value used to modulate the logits distribution.
-	 */
-	temperature?: number;
-	/**
-	 * The number of highest probability vocabulary tokens to keep for top-k-filtering.
-	 */
-	top_k?: number;
-	/**
-	 * If set to < 1, only the smallest set of most probable tokens with probabilities that add
-	 * up to `top_p` or higher are kept for generation.
-	 */
-	top_p?: number;
-	/**
-	 * Truncate input tokens to the given size.
-	 */
-	truncate?: number;
-	/**
-	 * Typical Decoding mass. See [Typical Decoding for Natural Language
-	 * Generation](https://hf.co/papers/2202.00666) for more information
-	 */
-	typical_p?: number;
-	/**
-	 * Watermarking with [A Watermark for Large Language Models](https://hf.co/papers/2301.10226)
-	 */
-	watermark?: boolean;
-	[property: string]: unknown;
-}
-
-/**
- * Outputs for Text Generation inference
- */
-export interface TextGenerationOutput {
-	/**
-	 * When enabled, details about the generation
-	 */
-	details?: TextGenerationOutputDetails;
-	/**
-	 * The generated text
-	 */
-	generated_text: string;
-	[property: string]: unknown;
-}
-
-/**
- * When enabled, details about the generation
- */
-export interface TextGenerationOutputDetails {
-	/**
-	 * Details about additional sequences when best_of is provided
-	 */
-	best_of_sequences?: TextGenerationOutputSequenceDetails[];
-	/**
-	 * The reason why the generation was stopped.
-	 */
-	finish_reason: TextGenerationFinishReason;
-	/**
-	 * The number of generated tokens
-	 */
-	generated_tokens: number;
-	prefill: TextGenerationPrefillToken[];
-	/**
-	 * The random seed used for generation
-	 */
-	seed?: number;
-	/**
-	 * The generated tokens and associated details
-	 */
-	tokens: TextGenerationOutputToken[];
-	/**
-	 * Most likely tokens
-	 */
-	top_tokens?: Array<TextGenerationOutputToken[]>;
-	[property: string]: unknown;
-}
-
-export interface TextGenerationOutputSequenceDetails {
-	finish_reason: TextGenerationFinishReason;
-	/**
-	 * The generated text
-	 */
-	generated_text: string;
-	/**
-	 * The number of generated tokens
-	 */
-	generated_tokens: number;
-	prefill: TextGenerationPrefillToken[];
-	/**
-	 * The random seed used for generation
-	 */
-	seed?: number;
-	/**
-	 * The generated tokens and associated details
-	 */
-	tokens: TextGenerationOutputToken[];
-	/**
-	 * Most likely tokens
-	 */
-	top_tokens?: Array<TextGenerationOutputToken[]>;
-	[property: string]: unknown;
-}
-
-export interface TextGenerationPrefillToken {
-	id: number;
-	logprob: number;
-	/**
-	 * The text associated with that token
-	 */
-	text: string;
-	[property: string]: unknown;
-}
-
-/**
- * Generated token.
- */
-export interface TextGenerationOutputToken {
-	id: number;
-	logprob?: number;
-	/**
-	 * Whether or not that token is a special one
-	 */
-	special: boolean;
-	/**
-	 * The text associated with that token
-	 */
-	text: string;
-	[property: string]: unknown;
-}
-
-/**
- * The reason why the generation was stopped.
- *
- * length: The generated sequence reached the maximum allowed length
- *
- * eos_token: The model generated an end-of-sentence (EOS) token
- *
- * stop_sequence: One of the sequence in stop_sequences was generated
- */
-export type TextGenerationFinishReason = "length" | "eos_token" | "stop_sequence";
+export type { TextGenerationInput, TextGenerationOutput };
 
 /**
  * Use to continue text from a prompt. This is a very generic task. Recommended model: gpt2 (it’s a simple model, but fun to play with).

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

@@ -1,6 +1,6 @@
+import type { TextGenerationInput } from "@huggingface/tasks";
 import type { BaseArgs, Options } from "../../types";
 import { streamingRequest } from "../custom/streamingRequest";
-import type { TextGenerationInput } from "./textGeneration";
 
 export interface TextGenerationStreamToken {
 	/** Token ID from the model tokenizer */