@huggingface/inference - npm Package Compare versions

Comparing version 2.6.6 to 2.6.7

dist/index.d.ts

@@ -34,3 +34,3 @@
 	/**
-	 * (Default: "same-origin"). String | Boolean. Credentials to use for the request. If this is a string, it will be passed straight on. If it's a boolean, true will be "include" and false will not send credentials at all.
+	 * Credentials to use for the request. If this is a string, it will be passed straight on. If it's a boolean, true will be "include" and false will not send credentials at all (which defaults to "same-origin" inside browsers).
 	 */
@@ -707,2 +707,197 @@ includeCredentials?: string | boolean;
 /**
+ * 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";
+/**
+ * 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;
+}
+/**
  * 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).
@@ -709,0 +904,0 @@ */

dist/index.js

@@ -137,6 +137,4 @@ /// <reference path="./index.d.ts" />
     credentials = includeCredentials;
-  } else if (typeof includeCredentials === "boolean") {
-    credentials = includeCredentials ? "include" : void 0;
-  } else if (includeCredentials === void 0) {
-    credentials = "same-origin";
+  } else if (includeCredentials === true) {
+    credentials = "include";
   }
@@ -150,3 +148,3 @@ const info = {
     }),
-    credentials,
+    ...credentials && { credentials },
     signal: options?.signal
@@ -153,0 +151,0 @@ };

package.json

 {
   "name": "@huggingface/inference",
-  "version": "2.6.6",
+  "version": "2.6.7",
   "packageManager": "pnpm@8.10.5",
@@ -43,3 +43,3 @@ "license": "MIT",
     "@types/node": "18.13.0",
-    "@huggingface/tasks": "^0.6.0"
+    "@huggingface/tasks": "^0.8.0"
   },
@@ -46,0 +46,0 @@ "resolutions": {},

src/HfInference.ts

@@ -5,2 +5,5 @@ import * as tasks from "./tasks";
 
+/* eslint-disable @typescript-eslint/no-empty-interface */
+/* eslint-disable @typescript-eslint/no-unsafe-declaration-merging */
+
 type Task = typeof tasks;
@@ -64,6 +67,4 @@
 
-// eslint-disable-next-line @typescript-eslint/no-empty-interface
 export interface HfInference extends TaskWithNoAccessToken {}
 
-// eslint-disable-next-line @typescript-eslint/no-empty-interface
 export interface HfInferenceEndpoint extends TaskWithNoAccessTokenNoModel {}

src/lib/makeRequestOptions.ts

@@ -92,15 +92,10 @@ import type { InferenceTask, Options, RequestArgs } from "../types";
 
-	// Let users configure credentials, or disable them all together (or keep default behavior).
-	// ---
-	// This used to be an internal property only and never exposed to users. This means that most usages will never define this value
-	// So in order to make this backwards compatible, if it's undefined we go to "same-origin" (default behaviour before).
-	// If it's a boolean and set to true then set to "include". If false, don't define credentials at all (useful for edge runtimes)
-	// Then finally, if it's a string, use it as-is.
+	/**
+	 * For edge runtimes, leave 'credentials' undefined, otherwise cloudflare workers will error
+	 */
 	let credentials: RequestCredentials | undefined;
 	if (typeof includeCredentials === "string") {
 		credentials = includeCredentials as RequestCredentials;
-	} else if (typeof includeCredentials === "boolean") {
-		credentials = includeCredentials ? "include" : undefined;
-	} else if (includeCredentials === undefined) {
-		credentials = "same-origin";
+	} else if (includeCredentials === true) {
+		credentials = "include";
 	}
@@ -117,3 +112,3 @@
 			  }),
-		credentials,
+		...(credentials && { credentials }),
 		signal: options?.signal,
@@ -120,0 +115,0 @@ };

src/tasks/nlp/textGeneration.ts

@@ -1,2 +0,1 @@
-import type { TextGenerationInput, TextGenerationOutput } from "@huggingface/tasks/src/tasks/text-generation/inference";
 import { InferenceOutputError } from "../../lib/InferenceOutputError";
@@ -7,2 +6,205 @@ import type { BaseArgs, Options } from "../../types";
 /**
+ * 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";
+
+/**
  * 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).
@@ -9,0 +211,0 @@ */

src/tasks/nlp/textGenerationStream.ts

 import type { BaseArgs, Options } from "../../types";
 import { streamingRequest } from "../custom/streamingRequest";
+import type { TextGenerationInput } from "./textGeneration";
 
-import type { TextGenerationInput } from "@huggingface/tasks/src/tasks/text-generation/inference";
-
 export interface TextGenerationStreamToken {
@@ -7,0 +6,0 @@ /** Token ID from the model tokenizer */

src/types.ts

@@ -35,3 +35,3 @@ import type { PipelineType } from "@huggingface/tasks";
 	/**
-	 * (Default: "same-origin"). String | Boolean. Credentials to use for the request. If this is a string, it will be passed straight on. If it's a boolean, true will be "include" and false will not send credentials at all.
+	 * Credentials to use for the request. If this is a string, it will be passed straight on. If it's a boolean, true will be "include" and false will not send credentials at all (which defaults to "same-origin" inside browsers).
 	 */
@@ -38,0 +38,0 @@ includeCredentials?: string | boolean;

No alert changes

Improved metrics

Total package byte prevSize

208957

increased by4.12%

Lines of code

5729

increased by7.1%

No dependency changes