@huggingface/tasks 0.2.2 → 0.3.0

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

@@ -3,6 +3,7 @@
  *
  * Using src/scripts/inference-codegen
  */
+
 /**
  * Inputs for Text Generation inference
  */
@@ -10,13 +11,14 @@ export interface TextGenerationInput {
 	/**
 	 * The text to initialize generation with
 	 */
-	data: string;
+	inputs: string;
 	/**
 	 * Additional inference parameters
 	 */
 	parameters?: TextGenerationParameters;
 	[property: string]: unknown;
 }
+
 /**
  * Additional inference parameters
  *
@@ -24,26 +26,43 @@ export interface TextGenerationInput {
  */
 export interface TextGenerationParameters {
 	/**
-	 * Whether to use logit sampling (true) or greedy search (false).
+	 * The number of sampling queries to run. Only the best one (in terms of total logprob) will
+	 * be returned.
 	 */
-	doSample?: boolean;
+	best_of?: number;
 	/**
-	 * Maximum number of generated tokens.
+	 * Whether or not to output decoder input details
 	 */
-	maxNewTokens?: number;
+	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.
 	 */
-	repetitionPenalty?: number;
+	repetition_penalty?: number;
 	/**
 	 * Whether to prepend the prompt to the generated text.
 	 */
-	returnFullText?: boolean;
+	return_full_text?: boolean;
+	/**
+	 * The random sampling seed.
+	 */
+	seed?: number;
 	/**
 	 * Stop generating tokens if a member of `stop_sequences` is generated.
 	 */
-	stopSequences?: string[];
+	stop_sequences?: string[];
 	/**
 	 * The value used to modulate the logits distribution.
 	 */
@@ -51,12 +70,12 @@ export interface TextGenerationParameters {
 	/**
 	 * The number of highest probability vocabulary tokens to keep for top-k-filtering.
 	 */
-	topK?: number;
+	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.
 	 */
-	topP?: number;
+	top_p?: number;
 	/**
 	 * Truncate input tokens to the given size.
 	 */
@@ -65,21 +84,111 @@ export interface TextGenerationParameters {
 	 * Typical Decoding mass. See [Typical Decoding for Natural Language
 	 * Generation](https://hf.co/papers/2202.00666) for more information
 	 */
-	typicalP?: number;
+	typical_p?: number;
 	/**
 	 * Watermarking with [A Watermark for Large Language Models](https://hf.co/papers/2301.10226)
 	 */
 	watermark?: boolean;
 	[property: string]: unknown;
 }
-export type TextGenerationOutput = TextGenerationOutputElement[];
+
 /**
  * Outputs for Text Generation inference
  */
-export interface TextGenerationOutputElement {
+export interface TextGenerationOutput {
+	/**
+	 * When enabled, details about the generation
+	 */
+	details?: TextGenerationOutputDetails;
 	/**
 	 * The generated text
 	 */
-	generatedText: string;
+	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?: TextGenerationSequenceDetails[];
+	/**
+	 * The reason why the generation was stopped.
+	 */
+	finish_reason: FinishReason;
+	/**
+	 * The number of generated tokens
+	 */
+	generated_tokens: number;
+	prefill: PrefillToken[];
+	/**
+	 * The random seed used for generation
+	 */
+	seed?: number;
+	/**
+	 * The generated tokens and associated details
+	 */
+	tokens: Token[];
+	[property: string]: unknown;
+}
+
+export interface TextGenerationSequenceDetails {
+	/**
+	 * The reason why the generation was stopped.
+	 */
+	finish_reason: FinishReason;
+	/**
+	 * The generated text
+	 */
+	generated_text: number;
+	/**
+	 * The number of generated tokens
+	 */
+	generated_tokens: number;
+	prefill: PrefillToken[];
+	/**
+	 * The random seed used for generation
+	 */
+	seed?: number;
+	/**
+	 * The generated tokens and associated details
+	 */
+	tokens: Token[];
+	[property: string]: unknown;
+}
+
+/**
+ * The generated sequence reached the maximum allowed length
+ *
+ * The model generated an end-of-sentence (EOS) token
+ *
+ * One of the sequence in stop_sequences was generated
+ */
+export type FinishReason = "length" | "eos_token" | "stop_sequence";
+
+export interface PrefillToken {
+	id: number;
+	logprob: number;
+	/**
+	 * The text associated with that token
+	 */
+	text: string;
+	[property: string]: unknown;
+}
+
+export interface Token {
+	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;
 }

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

@@ -5,7 +5,7 @@
 	"title": "TextGenerationInput",
 	"type": "object",
 	"properties": {
-		"data": {
+		"inputs": {
 			"description": "The text to initialize generation with",
 			"type": "string"
 		},
@@ -20,23 +20,39 @@
 			"description": "Additional inference parameters for Text Generation",
 			"type": "object",
 			"properties": {
-				"doSample": {
+				"best_of": {
+					"type": "integer",
+					"description": "The number of sampling queries to run. Only the best one (in terms of total logprob) will be returned."
+				},
+				"decoder_input_details": {
+					"type": "boolean",
+					"description": "Whether or not to output decoder input details"
+				},
+				"details": {
 					"type": "boolean",
-					"description": "Whether to use logit sampling (true) or greedy search (false)."
+					"description": "Whether or not to output details"
 				},
-				"maxNewTokens": {
+				"do_sample": {
+					"type": "boolean",
+					"description": "Whether to use logits sampling instead of greedy decoding when generating new tokens."
+				},
+				"max_new_tokens": {
 					"type": "integer",
-					"description": "Maximum number of generated tokens."
+					"description": "The maximum number of tokens to generate."
 				},
-				"repetitionPenalty": {
+				"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."
 				},
-				"returnFullText": {
+				"return_full_text": {
 					"type": "boolean",
 					"description": "Whether to prepend the prompt to the generated text."
 				},
-				"stopSequences": {
+				"seed": {
+					"type": "integer",
+					"description": "The random sampling seed."
+				},
+				"stop_sequences": {
 					"type": "array",
 					"items": {
 						"type": "string"
@@ -47,11 +63,11 @@
 					"type": "number",
 					"description": "The value used to modulate the logits distribution."
 				},
-				"topK": {
+				"top_k": {
 					"type": "integer",
 					"description": "The number of highest probability vocabulary tokens to keep for top-k-filtering."
 				},
-				"topP": {
+				"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."
 				},
@@ -59,7 +75,7 @@
 					"type": "integer",
 					"description": "Truncate input tokens to the given size."
 				},
-				"typicalP": {
+				"typical_p": {
 					"type": "number",
 					"description": "Typical Decoding mass. See [Typical Decoding for Natural Language Generation](https://hf.co/papers/2202.00666) for more information"
 				},
@@ -70,5 +86,5 @@
 			}
 		}
 	},
-	"required": ["data"]
+	"required": ["inputs"]
 }

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

@@ -3,15 +3,118 @@
 	"$schema": "http://json-schema.org/draft-06/schema#",
 	"description": "Outputs for Text Generation inference",
 	"title": "TextGenerationOutput",
-	"type": "array",
-	"items": {
-		"type": "object",
-		"properties": {
-			"generatedText": {
-				"type": "string",
-				"description": "The generated text"
-			}
+	"type": "object",
+	"properties": {
+		"generated_text": {
+			"type": "string",
+			"description": "The generated text"
 		},
-		"required": ["generatedText"]
+		"details": {
+			"description": "When enabled, details about the generation",
+			"title": "TextGenerationOutputDetails",
+			"allOf": [
+				{ "$ref": "#/$defs/SequenceDetails" },
+				{
+					"type": "object",
+					"properties": {
+						"best_of_sequences": {
+							"type": "array",
+							"description": "Details about additional sequences when best_of is provided",
+							"items": {
+								"allOf": [
+									{ "$ref": "#/$defs/SequenceDetails" },
+									{
+										"type": "object",
+										"properties": {
+											"generated_text": {
+												"type": "integer",
+												"description": "The generated text"
+											}
+										},
+										"required": ["generated_text"]
+									}
+								]
+							}
+						}
+					}
+				}
+			]
+		}
+	},
+	"required": ["generated_text"],
+
+	"$defs": {
+		"Token": {
+			"type": "object",
+			"title": "Token",
+			"properties": {
+				"id": {
+					"type": "integer"
+				},
+				"logprob": {
+					"type": "number"
+				},
+				"special": {
+					"type": "boolean",
+					"description": "Whether or not that token is a special one"
+				},
+				"text": {
+					"type": "string",
+					"description": "The text associated with that token"
+				}
+			},
+			"required": ["id", "logprob", "special", "text"]
+		},
+		"SequenceDetails": {
+			"type": "object",
+			"title": "TextGenerationSequenceDetails",
+			"properties": {
+				"finish_reason": {
+					"type": "string",
+					"description": "The reason why the generation was stopped.",
+					"oneOf": [
+						{ "const": "length", "description": "The generated sequence reached the maximum allowed length" },
+						{ "const": "eos_token", "description": "The model generated an end-of-sentence (EOS) token" },
+						{ "const": "stop_sequence", "description": "One of the sequence in stop_sequences was generated" }
+					]
+				},
+				"generated_tokens": {
+					"type": "integer",
+					"description": "The number of generated tokens"
+				},
+				"prefill": {
+					"type": "array",
+					"items": {
+						"title": "PrefillToken",
+						"type": "object",
+						"properties": {
+							"id": {
+								"type": "integer"
+							},
+							"logprob": {
+								"type": "number"
+							},
+							"text": {
+								"type": "string",
+								"description": "The text associated with that token"
+							}
+						},
+						"required": ["id", "logprob", "text"]
+					}
+				},
+				"seed": {
+					"type": "integer",
+					"description": "The random seed used for generation"
+				},
+				"tokens": {
+					"type": "array",
+					"description": "The generated tokens and associated details",
+					"items": {
+						"$ref": "#/$defs/Token"
+					}
+				}
+			},
+			"required": ["finish_reason", "generated_tokens", "prefill", "tokens"]
+		}
 	}
 }

package/src/tasks/text-to-audio/inference.ts

@@ -3,6 +3,7 @@
  *
  * Using src/scripts/inference-codegen
  */
+
 /**
  * Inputs for Text To Audio inference
  */
@@ -10,13 +11,14 @@ export interface TextToAudioInput {
 	/**
 	 * The input text data
 	 */
-	data: string;
+	inputs: string;
 	/**
 	 * Additional inference parameters
 	 */
 	parameters?: TextToAudioParameters;
 	[property: string]: unknown;
 }
+
 /**
  * Additional inference parameters
  *
@@ -29,6 +31,7 @@ export interface TextToAudioParameters {
 	generate?: GenerationParameters;
 	[property: string]: unknown;
 }
+
 /**
  * Parametrization of the text generation process
  *
@@ -38,18 +41,18 @@ export interface GenerationParameters {
 	/**
 	 * Whether to use sampling instead of greedy decoding when generating new tokens.
 	 */
-	doSample?: boolean;
+	do_sample?: boolean;
 	/**
 	 * Controls the stopping condition for beam-based methods.
 	 */
-	earlyStopping?: EarlyStoppingUnion;
+	early_stopping?: EarlyStoppingUnion;
 	/**
 	 * If set to float strictly between 0 and 1, only tokens with a conditional probability
 	 * greater than epsilon_cutoff will be sampled. In the paper, suggested values range from
 	 * 3e-4 to 9e-4, depending on the size of the model. See [Truncation Sampling as Language
 	 * Model Desmoothing](https://hf.co/papers/2210.15191) for more details.
 	 */
-	epsilonCutoff?: number;
+	epsilon_cutoff?: number;
 	/**
 	 * Eta sampling is a hybrid of locally typical sampling and epsilon sampling. If set to
 	 * float strictly between 0 and 1, a token is only considered if it is greater than either
@@ -59,37 +62,37 @@ export interface GenerationParameters {
 	 * See [Truncation Sampling as Language Model Desmoothing](https://hf.co/papers/2210.15191)
 	 * for more details.
 	 */
-	etaCutoff?: number;
+	eta_cutoff?: number;
 	/**
 	 * The maximum length (in tokens) of the generated text, including the input.
 	 */
-	maxLength?: number;
+	max_length?: number;
 	/**
 	 * The maximum number of tokens to generate. Takes precedence over maxLength.
 	 */
-	maxNewTokens?: number;
+	max_new_tokens?: number;
 	/**
 	 * The minimum length (in tokens) of the generated text, including the input.
 	 */
-	minLength?: number;
+	min_length?: number;
 	/**
 	 * The minimum number of tokens to generate. Takes precedence over maxLength.
 	 */
-	minNewTokens?: number;
+	min_new_tokens?: number;
 	/**
 	 * Number of groups to divide num_beams into in order to ensure diversity among different
 	 * groups of beams. See [this paper](https://hf.co/papers/1610.02424) for more details.
 	 */
-	numBeamGroups?: number;
+	num_beam_groups?: number;
 	/**
 	 * Number of beams to use for beam search.
 	 */
-	numBeams?: number;
+	num_beams?: number;
 	/**
 	 * The value balances the model confidence and the degeneration penalty in contrastive
 	 * search decoding.
 	 */
-	penaltyAlpha?: number;
+	penalty_alpha?: number;
 	/**
 	 * The value used to modulate the next token probabilities.
 	 */
@@ -97,12 +100,12 @@ export interface GenerationParameters {
 	/**
 	 * The number of highest probability vocabulary tokens to keep for top-k-filtering.
 	 */
-	topK?: number;
+	top_k?: number;
 	/**
 	 * If set to float < 1, only the smallest set of most probable tokens with probabilities
 	 * that add up to top_p or higher are kept for generation.
 	 */
-	topP?: number;
+	top_p?: number;
 	/**
 	 * Local typicality measures how similar the conditional probability of predicting a target
 	 * token next is to the expected conditional probability of predicting a random token next,
@@ -110,29 +113,31 @@ export interface GenerationParameters {
 	 * most locally typical tokens with probabilities that add up to typical_p or higher are
 	 * kept for generation. See [this paper](https://hf.co/papers/2202.00666) for more details.
 	 */
-	typicalP?: number;
+	typical_p?: number;
 	/**
 	 * Whether the model should use the past last key/values attentions to speed up decoding
 	 */
-	useCache?: boolean;
+	use_cache?: boolean;
 	[property: string]: unknown;
 }
+
 /**
  * Controls the stopping condition for beam-based methods.
  */
 export type EarlyStoppingUnion = boolean | "never";
-export type TextToAudioOutput = TextToAudioOutputElement[];
+
 /**
  * Outputs of inference for the Text To Audio task
  */
-export interface TextToAudioOutputElement {
+export interface TextToAudioOutput {
 	/**
 	 * The generated audio waveform.
 	 */
 	audio: unknown;
+	samplingRate: unknown;
 	/**
 	 * The sampling rate of the generated audio waveform.
 	 */
-	samplingRate: number;
+	sampling_rate?: number;
 	[property: string]: unknown;
 }

package/src/tasks/text-to-audio/spec/input.json

@@ -5,7 +5,7 @@
 	"title": "TextToAudioInput",
 	"type": "object",
 	"properties": {
-		"data": {
+		"inputs": {
 			"description": "The input text data",
 			"type": "string"
 		},
@@ -27,5 +27,5 @@
 			}
 		}
 	},
-	"required": ["data"]
+	"required": ["inputs"]
 }

package/src/tasks/text-to-audio/spec/output.json

@@ -3,18 +3,15 @@
 	"$schema": "http://json-schema.org/draft-06/schema#",
 	"description": "Outputs of inference for the Text To Audio task",
 	"title": "TextToAudioOutput",
-	"type": "array",
-	"items": {
-		"type": "object",
-		"properties": {
-			"audio": {
-				"description": "The generated audio waveform."
-			},
-			"samplingRate": {
-				"type": "number",
-				"description": "The sampling rate of the generated audio waveform."
-			}
+	"type": "object",
+	"properties": {
+		"audio": {
+			"description": "The generated audio waveform."
 		},
-		"required": ["audio", "samplingRate"]
-	}
+		"sampling_rate": {
+			"type": "number",
+			"description": "The sampling rate of the generated audio waveform."
+		}
+	},
+	"required": ["audio", "samplingRate"]
 }

package/src/tasks/text-to-image/inference.ts

@@ -11,7 +11,7 @@ export interface TextToImageInput {
 	/**
 	 * The input text data (sometimes called "prompt"
 	 */
-	data: string;
+	inputs: string;
 	/**
 	 * Additional inference parameters
 	 */
@@ -29,16 +29,16 @@ export interface TextToImageParameters {
 	 * For diffusion models. A higher guidance scale value encourages the model to generate
 	 * images closely linked to the text prompt at the expense of lower image quality.
 	 */
-	guidanceScale?: number;
+	guidance_scale?: number;
 	/**
 	 * One or several prompt to guide what NOT to include in image generation.
 	 */
-	negativePrompt?: string[];
+	negative_prompt?: string[];
 	/**
 	 * For diffusion models. The number of denoising steps. More denoising steps usually lead to
 	 * a higher quality image at the expense of slower inference.
 	 */
-	numInferenceSteps?: number;
+	num_inference_steps?: number;
 	/**
 	 * For diffusion models. Override the scheduler with a compatible one
 	 */
@@ -46,7 +46,7 @@ export interface TextToImageParameters {
 	/**
 	 * The size in pixel of the output image
 	 */
-	targetSize?: TargetSize;
+	target_size?: TargetSize;
 	[property: string]: unknown;
 }
 
@@ -62,9 +62,7 @@ export interface TargetSize {
 /**
  * Outputs of inference for the Text To Image task
  */
-export type TextToImageOutput = unknown[] | boolean | number | number | null | TextToImageOutputObject | string;
-
-export interface TextToImageOutputObject {
+export interface TextToImageOutput {
 	/**
 	 * The generated image
 	 */