@huggingface/tasks 0.9.1 → 0.10.1

package/src/tasks/index.ts

@@ -43,9 +43,8 @@ export type {
 	ChatCompletionInput,
 	ChatCompletionInputMessage,
 	ChatCompletionOutput,
-	ChatCompletionOutputChoice,
-	ChatCompletionFinishReason,
-	ChatCompletionOutputChoiceMessage,
+	ChatCompletionOutputComplete,
+	ChatCompletionOutputMessage,
 	ChatCompletionStreamOutput,
 	ChatCompletionStreamOutputChoice,
 	ChatCompletionStreamOutputDelta,
@@ -85,15 +84,15 @@ export type {
 	TextClassificationParameters,
 } from "./text-classification/inference";
 export type {
-	TextGenerationFinishReason,
-	TextGenerationPrefillToken,
+	TextGenerationOutputFinishReason,
+	TextGenerationOutputPrefillToken,
 	TextGenerationInput,
 	TextGenerationOutput,
 	TextGenerationOutputDetails,
-	TextGenerationParameters,
-	TextGenerationOutputSequenceDetails,
+	TextGenerationInputGenerateParameters,
+	TextGenerationOutputBestOfSequence,
 	TextGenerationOutputToken,
-	TextGenerationStreamDetails,
+	TextGenerationStreamOutputStreamDetails,
 	TextGenerationStreamOutput,
 } from "./text-generation/inference";
 export type * from "./video-classification/inference";

package/src/tasks/text-generation/inference.ts

@@ -5,246 +5,134 @@
  */
 
 /**
- * Inputs for Text Generation inference
+ * Text Generation Input.
+ *
+ * Auto-generated from TGI specs.
+ * For more details, check out
+ * https://github.com/huggingface/huggingface.js/blob/main/packages/tasks/scripts/inference-tgi-import.ts.
  */
 export interface TextGenerationInput {
-	/**
-	 * The text to initialize generation with
-	 */
 	inputs: string;
-	/**
-	 * Additional inference parameters
-	 */
-	parameters?: TextGenerationParameters;
-	/**
-	 * Whether to stream output tokens
-	 */
+	parameters?: TextGenerationInputGenerateParameters;
 	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.
-	 */
+export interface TextGenerationInputGenerateParameters {
 	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.
-	 */
+	frequency_penalty?: number;
+	grammar?: TextGenerationInputGrammarType;
 	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.
-	 */
+	stop?: string[];
 	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_n_tokens?: number;
 	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;
+	watermark?: boolean;
+	[property: string]: unknown;
+}
+
+export interface TextGenerationInputGrammarType {
+	type: Type;
 	/**
-	 * Watermarking with [A Watermark for Large Language Models](https://hf.co/papers/2301.10226)
+	 * A string that represents a [JSON Schema](https://json-schema.org/).
+	 *
+	 * JSON Schema is a declarative language that allows to annotate JSON documents
+	 * with types and descriptions.
 	 */
-	watermark?: boolean;
+	value: unknown;
 	[property: string]: unknown;
 }
 
+export type Type = "json" | "regex";
+
 /**
- * Outputs for Text Generation inference
+ * Text Generation Output.
+ *
+ * Auto-generated from TGI specs.
+ * For more details, check out
+ * https://github.com/huggingface/huggingface.js/blob/main/packages/tasks/scripts/inference-tgi-import.ts.
  */
 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
-	 */
+	best_of_sequences?: TextGenerationOutputBestOfSequence[];
+	finish_reason: TextGenerationOutputFinishReason;
 	generated_tokens: number;
-	prefill: TextGenerationPrefillToken[];
-	/**
-	 * The random seed used for generation
-	 */
+	prefill: TextGenerationOutputPrefillToken[];
 	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
-	 */
+export interface TextGenerationOutputBestOfSequence {
+	finish_reason: TextGenerationOutputFinishReason;
 	generated_text: string;
-	/**
-	 * The number of generated tokens
-	 */
 	generated_tokens: number;
-	prefill: TextGenerationPrefillToken[];
-	/**
-	 * The random seed used for generation
-	 */
+	prefill: TextGenerationOutputPrefillToken[];
 	seed?: number;
-	/**
-	 * The generated tokens and associated details
-	 */
 	tokens: TextGenerationOutputToken[];
-	/**
-	 * Most likely tokens
-	 */
 	top_tokens?: Array<TextGenerationOutputToken[]>;
 	[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 TextGenerationOutputFinishReason = "length" | "eos_token" | "stop_sequence";
 
-export interface TextGenerationPrefillToken {
+export interface TextGenerationOutputPrefillToken {
 	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
-	 */
+	logprob: number;
 	special: boolean;
-	/**
-	 * The text associated with that token
-	 */
 	text: string;
 	[property: string]: unknown;
 }
 
 /**
- * Text Generation Stream Output
+ * Text Generation Stream Output.
+ *
+ * Auto-generated from TGI specs.
+ * For more details, check out
+ * https://github.com/huggingface/huggingface.js/blob/main/packages/tasks/scripts/inference-tgi-import.ts.
  */
 export interface TextGenerationStreamOutput {
-	/**
-	 * Generation details. Only available when the generation is finished.
-	 */
-	details?: TextGenerationStreamDetails;
-	/**
-	 * The complete generated text. Only available when the generation is finished.
-	 */
+	details?: TextGenerationStreamOutputStreamDetails;
 	generated_text?: string;
-	/**
-	 * The token index within the stream. Optional to support older clients that omit it.
-	 */
-	index?: number;
-	/**
-	 * Generated token.
-	 */
-	token: TextGenerationOutputToken;
+	index: number;
+	token: TextGenerationStreamOutputToken;
+	top_tokens?: TextGenerationStreamOutputToken[];
 	[property: string]: unknown;
 }
 
-/**
- * Generation details. Only available when the generation is finished.
- */
-export interface TextGenerationStreamDetails {
-	/**
-	 * The reason why the generation was stopped.
-	 */
-	finish_reason: TextGenerationFinishReason;
-	/**
-	 * The number of generated tokens
-	 */
+export interface TextGenerationStreamOutputStreamDetails {
+	finish_reason: TextGenerationOutputFinishReason;
 	generated_tokens: number;
-	/**
-	 * The random seed used for generation
-	 */
-	seed: number;
+	seed?: number;
+	[property: string]: unknown;
+}
+
+export interface TextGenerationStreamOutputToken {
+	id: number;
+	logprob: number;
+	special: boolean;
+	text: string;
 	[property: string]: unknown;
 }

package/src/tasks/text-generation/spec/input.json

@@ -1,94 +1,195 @@
 {
 	"$id": "/inference/schemas/text-generation/input.json",
 	"$schema": "http://json-schema.org/draft-06/schema#",
-	"description": "Inputs for Text Generation inference",
+	"description": "Text Generation Input.\n\nAuto-generated from TGI specs.\nFor more details, check out https://github.com/huggingface/huggingface.js/blob/main/packages/tasks/scripts/inference-tgi-import.ts.",
 	"title": "TextGenerationInput",
 	"type": "object",
+	"required": ["inputs"],
 	"properties": {
 		"inputs": {
-			"description": "The text to initialize generation with",
-			"type": "string"
+			"type": "string",
+			"example": "My name is Olivier and I"
 		},
 		"parameters": {
-			"description": "Additional inference parameters",
-			"$ref": "#/$defs/TextGenerationParameters"
+			"$ref": "#/$defs/TextGenerationInputGenerateParameters"
 		},
 		"stream": {
-			"description": "Whether to stream output tokens",
-			"type": "boolean"
+			"type": "boolean",
+			"default": "false"
 		}
 	},
 	"$defs": {
-		"TextGenerationParameters": {
-			"title": "TextGenerationParameters",
-			"description": "Additional inference parameters for Text Generation",
+		"TextGenerationInputGenerateParameters": {
 			"type": "object",
 			"properties": {
 				"best_of": {
 					"type": "integer",
-					"description": "The number of sampling queries to run. Only the best one (in terms of total logprob) will be returned."
+					"default": "null",
+					"example": 1,
+					"nullable": true,
+					"minimum": 0,
+					"exclusiveMinimum": 0
 				},
 				"decoder_input_details": {
 					"type": "boolean",
-					"description": "Whether or not to output decoder input details"
+					"default": "false"
 				},
 				"details": {
 					"type": "boolean",
-					"description": "Whether or not to output details"
+					"default": "true"
 				},
 				"do_sample": {
 					"type": "boolean",
-					"description": "Whether to use logits sampling instead of greedy decoding when generating new tokens."
+					"default": "false",
+					"example": true
+				},
+				"frequency_penalty": {
+					"type": "number",
+					"format": "float",
+					"default": "null",
+					"example": 0.1,
+					"nullable": true,
+					"exclusiveMinimum": -2
+				},
+				"grammar": {
+					"allOf": [
+						{
+							"$ref": "#/$defs/TextGenerationInputGrammarType"
+						}
+					],
+					"default": "null",
+					"nullable": true
 				},
 				"max_new_tokens": {
 					"type": "integer",
-					"description": "The maximum number of tokens to generate."
+					"format": "int32",
+					"default": "100",
+					"example": "20",
+					"nullable": true,
+					"minimum": 0
 				},
 				"repetition_penalty": {
 					"type": "number",
-					"description": "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."
+					"format": "float",
+					"default": "null",
+					"example": 1.03,
+					"nullable": true,
+					"exclusiveMinimum": 0
 				},
 				"return_full_text": {
 					"type": "boolean",
-					"description": "Whether to prepend the prompt to the generated text."
+					"default": "null",
+					"example": false,
+					"nullable": true
 				},
 				"seed": {
 					"type": "integer",
-					"description": "The random sampling seed."
+					"format": "int64",
+					"default": "null",
+					"example": "null",
+					"nullable": true,
+					"minimum": 0,
+					"exclusiveMinimum": 0
 				},
-				"stop_sequences": {
+				"stop": {
 					"type": "array",
 					"items": {
 						"type": "string"
 					},
-					"description": "Stop generating tokens if a member of `stop_sequences` is generated."
+					"example": ["photographer"],
+					"maxItems": 4
 				},
 				"temperature": {
 					"type": "number",
-					"description": "The value used to modulate the logits distribution."
+					"format": "float",
+					"default": "null",
+					"example": 0.5,
+					"nullable": true,
+					"exclusiveMinimum": 0
 				},
 				"top_k": {
 					"type": "integer",
-					"description": "The number of highest probability vocabulary tokens to keep for top-k-filtering."
+					"format": "int32",
+					"default": "null",
+					"example": 10,
+					"nullable": true,
+					"exclusiveMinimum": 0
+				},
+				"top_n_tokens": {
+					"type": "integer",
+					"format": "int32",
+					"default": "null",
+					"example": 5,
+					"nullable": true,
+					"minimum": 0,
+					"exclusiveMinimum": 0
 				},
 				"top_p": {
 					"type": "number",
-					"description": "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."
+					"format": "float",
+					"default": "null",
+					"example": 0.95,
+					"nullable": true,
+					"maximum": 1,
+					"exclusiveMinimum": 0
 				},
 				"truncate": {
 					"type": "integer",
-					"description": "Truncate input tokens to the given size."
+					"default": "null",
+					"example": "null",
+					"nullable": true,
+					"minimum": 0
 				},
 				"typical_p": {
 					"type": "number",
-					"description": "Typical Decoding mass. See [Typical Decoding for Natural Language Generation](https://hf.co/papers/2202.00666) for more information"
+					"format": "float",
+					"default": "null",
+					"example": 0.95,
+					"nullable": true,
+					"maximum": 1,
+					"exclusiveMinimum": 0
 				},
 				"watermark": {
 					"type": "boolean",
-					"description": "Watermarking with [A Watermark for Large Language Models](https://hf.co/papers/2301.10226)"
+					"default": "false",
+					"example": true
 				}
-			}
+			},
+			"title": "TextGenerationInputGenerateParameters"
+		},
+		"TextGenerationInputGrammarType": {
+			"oneOf": [
+				{
+					"type": "object",
+					"required": ["type", "value"],
+					"properties": {
+						"type": {
+							"type": "string",
+							"enum": ["json"]
+						},
+						"value": {
+							"description": "A string that represents a [JSON Schema](https://json-schema.org/).\n\nJSON Schema is a declarative language that allows to annotate JSON documents\nwith types and descriptions."
+						}
+					}
+				},
+				{
+					"type": "object",
+					"required": ["type", "value"],
+					"properties": {
+						"type": {
+							"type": "string",
+							"enum": ["regex"]
+						},
+						"value": {
+							"type": "string"
+						}
+					}
+				}
+			],
+			"discriminator": {
+				"propertyName": "type"
+			},
+			"title": "TextGenerationInputGrammarType"
 		}
-	},
-	"required": ["inputs"]
+	}
 }