@huggingface/transformers 3.2.3 → 3.2.4

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@huggingface/transformers",
-  "version": "3.2.3",
+  "version": "3.2.4",
   "description": "State-of-the-art Machine Learning for the web. Run 🤗 Transformers directly in your browser, with no need for a server!",
   "main": "./src/transformers.js",
   "types": "./types/transformers.d.ts",
@@ -24,7 +24,7 @@
   "scripts": {
     "format": "prettier --write .",
     "format:check": "prettier --check .",
-    "typegen": "tsc ./src/transformers.js --allowJs --declaration --emitDeclarationOnly --declarationMap --outDir types",
+    "typegen": "tsc --build",
     "dev": "webpack serve --no-client-overlay",
     "build": "webpack && npm run typegen",
     "test": "node --experimental-vm-modules node_modules/jest/bin/jest.js --verbose",

package/src/base/feature_extraction_utils.js

@@ -17,23 +17,23 @@ export class FeatureExtractor extends Callable {
     }
 
     /**
-     * Instantiate one of the processor classes of the library from a pretrained model.
+     * Instantiate one of the feature extractor classes of the library from a pretrained model.
      * 
-     * The processor class to instantiate is selected based on the `image_processor_type` (or `feature_extractor_type`; legacy)
-     * property of the config object (either passed as an argument or loaded from `pretrained_model_name_or_path` if possible)
+     * The feature extractor class to instantiate is selected based on the `feature_extractor_type` property of
+     * the config object (either passed as an argument or loaded from `pretrained_model_name_or_path` if possible)
      * 
      * @param {string} pretrained_model_name_or_path The name or path of the pretrained model. Can be either:
-     * - A string, the *model id* of a pretrained processor hosted inside a model repo on huggingface.co.
+     * - A string, the *model id* of a pretrained feature_extractor hosted inside a model repo on huggingface.co.
      *   Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced under a
      *   user or organization name, like `dbmdz/bert-base-german-cased`.
-     * - A path to a *directory* containing processor files, e.g., `./my_model_directory/`.
-     * @param {import('../utils/hub.js').PretrainedOptions} options Additional options for loading the processor.
+     * - A path to a *directory* containing feature_extractor files, e.g., `./my_model_directory/`.
+     * @param {import('../utils/hub.js').PretrainedOptions} options Additional options for loading the feature_extractor.
      * 
-     * @returns {Promise<FeatureExtractor>} A new instance of the Processor class.
+     * @returns {Promise<FeatureExtractor>} A new instance of the Feature Extractor class.
      */
     static async from_pretrained(pretrained_model_name_or_path, options) {
-        const preprocessorConfig = await getModelJSON(pretrained_model_name_or_path, FEATURE_EXTRACTOR_NAME, true, options);
-        return new this(preprocessorConfig);
+        const config = await getModelJSON(pretrained_model_name_or_path, FEATURE_EXTRACTOR_NAME, true, options);
+        return new this(config);
     }
 }
 

package/src/base/image_processors_utils.js

@@ -604,14 +604,20 @@ export class ImageProcessor extends Callable {
         this.do_thumbnail = config.do_thumbnail;
         this.size = config.size ?? config.image_size;
         this.do_resize = config.do_resize ?? (this.size !== undefined);
+        // @ts-expect-error TS2339
         this.size_divisibility = config.size_divisibility ?? config.size_divisor;
 
         this.do_center_crop = config.do_center_crop;
+        // @ts-expect-error TS2339
         this.crop_size = config.crop_size;
+        // @ts-expect-error TS2339
         this.do_convert_rgb = config.do_convert_rgb ?? true;
+        // @ts-expect-error TS2339
         this.do_crop_margin = config.do_crop_margin;
 
+        // @ts-expect-error TS2339
         this.pad_size = config.pad_size;
+        // @ts-expect-error TS2339
         this.do_pad = config.do_pad;
 
         if (this.do_pad && !this.pad_size && this.size && this.size.width !== undefined && this.size.height !== undefined) {
@@ -820,6 +826,7 @@ export class ImageProcessor extends Callable {
         // Support both formats for backwards compatibility
         else if (Number.isInteger(size)) {
             shortest_edge = size;
+            // @ts-expect-error TS2339
             longest_edge = this.config.max_size ?? shortest_edge;
 
         } else if (size !== undefined) {
@@ -888,6 +895,7 @@ export class ImageProcessor extends Callable {
         } else if (size.min_pixels !== undefined && size.max_pixels !== undefined) {
             // Custom resize logic for Qwen2-VL models
             const { min_pixels, max_pixels } = size;
+            // @ts-expect-error TS2339
             const factor = this.config.patch_size * this.config.merge_size;
             return smart_resize(srcHeight, srcWidth, factor, min_pixels, max_pixels);
         } else {
@@ -903,6 +911,7 @@ export class ImageProcessor extends Callable {
     async resize(image) {
         const [newWidth, newHeight] = this.get_resize_output_image_size(image, this.size);
         return await image.resize(newWidth, newHeight, {
+            // @ts-expect-error TS2322
             resample: this.resample,
         });
     }
@@ -953,6 +962,7 @@ export class ImageProcessor extends Callable {
 
         // Resize the image using thumbnail method.
         if (this.do_thumbnail) {
+            // @ts-expect-error TS2345
             image = await this.thumbnail(image, this.size, this.resample);
         }
 
@@ -977,6 +987,7 @@ export class ImageProcessor extends Callable {
         // NOTE: All pixel-level manipulation (i.e., modifying `pixelData`)
         // occurs with data in the hwc format (height, width, channels), 
         // to emulate the behavior of the original Python code (w/ numpy).
+        /** @type {Float32Array} */
         let pixelData = Float32Array.from(image.data);
         let imgDims = [image.height, image.width, image.channels];
 

package/src/base/processing_utils.js

@@ -28,6 +28,7 @@ import { getModelJSON } from '../utils/hub.js';
 /**
  * @typedef {Object} ProcessorProperties Additional processor-specific properties.
  * @typedef {import('../utils/hub.js').PretrainedOptions & ProcessorProperties} PretrainedProcessorOptions
+ * @typedef {import('../tokenizers.js').PreTrainedTokenizer} PreTrainedTokenizer
  */
 
 
@@ -61,7 +62,7 @@ export class Processor extends Callable {
     }
 
     /**
-     * @returns {import('../tokenizers.js').PreTrainedTokenizer|undefined} The tokenizer of the processor, if it exists.
+     * @returns {PreTrainedTokenizer|undefined} The tokenizer of the processor, if it exists.
      */
     get tokenizer() {
         return this.components.tokenizer;
@@ -74,6 +75,11 @@ export class Processor extends Callable {
         return this.components.feature_extractor;
     }
 
+    /**
+     * @param {Parameters<PreTrainedTokenizer['apply_chat_template']>[0]} messages
+     * @param {Parameters<PreTrainedTokenizer['apply_chat_template']>[1]} options
+     * @returns {ReturnType<PreTrainedTokenizer['apply_chat_template']>}
+     */
     apply_chat_template(messages, options = {}) {
         if (!this.tokenizer) {
             throw new Error('Unable to apply chat template without a tokenizer.');
@@ -84,6 +90,10 @@ export class Processor extends Callable {
         });
     }
 
+    /**
+     * @param {Parameters<PreTrainedTokenizer['batch_decode']>} args
+     * @returns {ReturnType<PreTrainedTokenizer['batch_decode']>}
+     */
     batch_decode(...args) {
         if (!this.tokenizer) {
             throw new Error('Unable to decode without a tokenizer.');
@@ -111,8 +121,8 @@ export class Processor extends Callable {
     /**
      * Instantiate one of the processor classes of the library from a pretrained model.
      * 
-     * The processor class to instantiate is selected based on the `feature_extractor_type` property of the config object
-     * (either passed as an argument or loaded from `pretrained_model_name_or_path` if possible)
+     * The processor class to instantiate is selected based on the `image_processor_type` (or `feature_extractor_type`; legacy)
+     * property of the config object (either passed as an argument or loaded from `pretrained_model_name_or_path` if possible)
      * 
      * @param {string} pretrained_model_name_or_path The name or path of the pretrained model. Can be either:
      * - A string, the *model id* of a pretrained processor hosted inside a model repo on huggingface.co.

package/src/configs.js

@@ -70,15 +70,19 @@ function getNormalizedConfig(config) {
         case 'florence2':
         case 'llava_onevision':
         case 'idefics3':
+            // @ts-expect-error TS2339
             init_normalized_config = getNormalizedConfig(config.text_config);
             break;
         case 'moondream1':
+            // @ts-expect-error TS2339
             init_normalized_config = getNormalizedConfig(config.phi_config);
             break;
         case 'musicgen':
+            // @ts-expect-error TS2339
             init_normalized_config = getNormalizedConfig(config.decoder);
             break;
         case 'multi_modality':
+            // @ts-expect-error TS2339
             init_normalized_config = getNormalizedConfig(config.language_config);
             break;
 
@@ -199,6 +203,7 @@ function getNormalizedConfig(config) {
             break;
 
         case 'vision-encoder-decoder':
+            // @ts-expect-error TS2339
             const decoderConfig = getNormalizedConfig(config.decoder);
 
             const add_encoder_pkv = 'num_decoder_layers' in decoderConfig;

package/src/env.js

@@ -26,7 +26,7 @@ import fs from 'fs';
 import path from 'path';
 import url from 'url';
 
-const VERSION = '3.2.3';
+const VERSION = '3.2.4';
 
 // Check if various APIs are available (depends on environment)
 const IS_BROWSER_ENV = typeof window !== "undefined" && typeof window.document !== "undefined";

package/src/models/auto/feature_extraction_auto.js

@@ -6,22 +6,6 @@ import * as AllFeatureExtractors from '../feature_extractors.js';
 
 export class AutoFeatureExtractor {
 
-    /**
-     * Instantiate one of the feature extractor classes of the library from a pretrained model.
-     * 
-     * The processor class to instantiate is selected based on the `feature_extractor_type` property of
-     * the config object (either passed as an argument or loaded from `pretrained_model_name_or_path` if possible)
-     * 
-     * @param {string} pretrained_model_name_or_path The name or path of the pretrained model. Can be either:
-     * - A string, the *model id* of a pretrained processor hosted inside a model repo on huggingface.co.
-     *   Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced under a
-     *   user or organization name, like `dbmdz/bert-base-german-cased`.
-     * - A path to a *directory* containing processor files, e.g., `./my_model_directory/`.
-     * @param {import('../../utils/hub.js').PretrainedOptions} options Additional options for loading the processor.
-     * 
-     * @returns {Promise<AllFeatureExtractors.ImageProcessor>} A new instance of the Processor class.
-     */
-
     /** @type {typeof FeatureExtractor.from_pretrained} */
     static async from_pretrained(pretrained_model_name_or_path, options={}) {
 

package/src/models/auto/processing_auto.js

@@ -40,22 +40,6 @@ import * as AllFeatureExtractors from '../feature_extractors.js';
  */
 export class AutoProcessor {
 
-    /**
-     * Instantiate one of the processor classes of the library from a pretrained model.
-     * 
-     * The processor class to instantiate is selected based on the `image_processor_type` (or `feature_extractor_type`; legacy)
-     * property of the config object (either passed as an argument or loaded from `pretrained_model_name_or_path` if possible)
-     * 
-     * @param {string} pretrained_model_name_or_path The name or path of the pretrained model. Can be either:
-     * - A string, the *model id* of a pretrained processor hosted inside a model repo on huggingface.co.
-     *   Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced under a
-     *   user or organization name, like `dbmdz/bert-base-german-cased`.
-     * - A path to a *directory* containing processor files, e.g., `./my_model_directory/`.
-     * @param {import('../../utils/hub.js').PretrainedOptions} options Additional options for loading the processor.
-     * 
-     * @returns {Promise<Processor>} A new instance of the Processor class.
-     */
-
     /** @type {typeof Processor.from_pretrained} */
     static async from_pretrained(pretrained_model_name_or_path, options={}) {
 

package/src/models/convnext/image_processing_convnext.js

@@ -9,6 +9,7 @@ export class ConvNextImageProcessor extends ImageProcessor {
         /**
          * Percentage of the image to crop. Only has an effect if this.size < 384.
          */
+        // @ts-expect-error TS2339
         this.crop_pct = this.config.crop_pct ?? (224 / 256);
     }
 

package/src/models/efficientnet/image_processing_efficientnet.js

@@ -5,6 +5,7 @@ import {
 export class EfficientNetImageProcessor extends ImageProcessor {
     constructor(config) {
         super(config);
+        // @ts-expect-error TS2339
         this.include_top = this.config.include_top ?? true;
         if (this.include_top) {
             this.image_std = this.image_std.map(x => x * x);

package/src/models/florence2/processing_florence2.js

@@ -10,8 +10,11 @@ export class Florence2Processor extends Processor {
         super(config, components);
 
         const {
+            // @ts-expect-error TS2339
             tasks_answer_post_processing_type,
+            // @ts-expect-error TS2339
             task_prompts_without_inputs,
+            // @ts-expect-error TS2339
             task_prompts_with_input,
         } = this.image_processor.config;
 

package/src/models/idefics3/image_processing_idefics3.js

@@ -146,6 +146,8 @@ export class Idefics3ImageProcessor extends ImageProcessor {
 
                     const start_offset = i * pixel_attention_mask_stride + num_patches * h * w;
                     const end_offset = (i + 1) * pixel_attention_mask_stride;
+
+                    // @ts-expect-error
                     pixel_attention_mask_data.fill(false, start_offset, end_offset);
                 }
             }

package/src/models/janus/image_processing_janus.js

@@ -13,6 +13,7 @@ export class VLMImageProcessor extends ImageProcessor {
             },
             ...config,
         });
+        // @ts-expect-error TS2339
         this.constant_values = this.config.background_color.map(x => x * this.rescale_factor)
     }
 

package/src/models/mgp_str/processing_mgp_str.js

@@ -119,6 +119,8 @@ export class MgpstrProcessor extends Processor {
      * - bpe_preds: The list of BPE decoded sentences.
      * - wp_preds: The list of wp decoded sentences.
      */
+    // @ts-expect-error The type of this method is not compatible with the one
+    // in the base class. It might be a good idea to fix this.
     batch_decode([char_logits, bpe_logits, wp_logits]) {
         const [char_preds, char_scores] = this._decode_helper(char_logits, 'char');
         const [bpe_preds, bpe_scores] = this._decode_helper(bpe_logits, 'bpe');

package/src/models/paligemma/processing_paligemma.js

@@ -41,6 +41,7 @@ export class PaliGemmaProcessor extends Processor {
         }
 
         const bos_token = this.tokenizer.bos_token;
+        // @ts-expect-error TS2339
         const image_seq_length = this.image_processor.config.image_seq_length;
         let input_strings;
         if (text.some((t) => t.includes(IMAGE_TOKEN))) {

package/src/models/phi3_v/processing_phi3_v.js

@@ -14,7 +14,7 @@ export class Phi3VProcessor extends Processor {
      * 
      * @param {string|string[]} text 
      * @param {RawImage|RawImage[]} images 
-     * @param  {...any} args 
+     * @param  { { padding?: boolean, truncation?: boolean, num_crops?: number } | undefined } options
      * @returns {Promise<any>}
      */
     async _call(text, images = null, {

package/src/models/pyannote/feature_extraction_pyannote.js

@@ -52,6 +52,7 @@ export class PyAnnoteFeatureExtractor extends FeatureExtractor {
 
             let current_speaker = -1;
             for (let i = 0; i < scores.length; ++i) {
+                /** @type {number[]} */
                 const probabilities = softmax(scores[i]);
                 const [score, id] = max(probabilities);
                 const [start, end] = [i, i + 1];

package/src/models/qwen2_vl/processing_qwen2_vl.js

@@ -28,6 +28,7 @@ export class Qwen2VLProcessor extends Processor {
         }
 
         if (image_grid_thw) {
+            // @ts-expect-error TS2551
             let merge_length = this.image_processor.config.merge_size ** 2;
             let index = 0;
 

package/src/models/seamless_m4t/feature_extraction_seamless_m4t.js

@@ -133,8 +133,8 @@ export class SeamlessM4TFeatureExtractor extends FeatureExtractor {
                         'int64',
                         new BigInt64Array(numPaddedFrames),
                         [1, numPaddedFrames],
-                    )
-                    padded_attention_mask.data.fill(1n, 0, num_frames);
+                    );
+                    /** @type {BigInt64Array} */ (padded_attention_mask.data).fill(1n, 0, num_frames);
                 }
             }
         }

package/src/models/whisper/feature_extraction_whisper.js

@@ -44,7 +44,7 @@ export class WhisperFeatureExtractor extends FeatureExtractor {
         )
 
         const data = features.data;
-        const maxValue = max(data)[0];
+        const maxValue = max(/** @type {Float32Array} */(data))[0];
 
         for (let i = 0; i < data.length; ++i) {
             data[i] = (Math.max(data[i], maxValue - 8.0) + 4.0) / 4.0;

package/src/models.js

@@ -270,8 +270,11 @@ async function getSession(pretrained_model_name_or_path, fileName, options) {
     } else if (session_options.externalData !== undefined) {
         externalDataPromises = session_options.externalData.map(async (ext) => {
             // if the external data is a string, fetch the file and replace the string with its content
+            // @ts-expect-error TS2339
             if (typeof ext.data === "string") {
+                // @ts-expect-error TS2339
                 const ext_buffer = await getModelFile(pretrained_model_name_or_path, ext.data, true, options);
+                // @ts-expect-error TS2698
                 return { ...ext, data: ext_buffer };
             }
             return ext;
@@ -1519,6 +1522,7 @@ export class PreTrainedModel extends Callable {
                 if (this.config.model_type === 'musicgen') {
                     // Custom logic (TODO: move to Musicgen class)
                     decoder_input_ids = Array.from({
+                        // @ts-expect-error TS2339
                         length: batch_size * this.config.decoder.num_codebooks
                     }, () => [decoder_start_token_id]);
 
@@ -1848,11 +1852,13 @@ export class PreTrainedModel extends Callable {
     async encode_image({ pixel_values }) {
         // image_inputs === { pixel_values }
         const features = (await sessionRun(this.sessions['vision_encoder'], { pixel_values })).image_features;
+        // @ts-expect-error TS2339
         if (!this.config.num_image_tokens) {
             console.warn(
                 'The number of image tokens was not set in the model configuration. ' +
                 `Setting it to the number of features detected by the vision encoder (${features.dims[1]}).`
             )
+            // @ts-expect-error TS2339
             this.config.num_image_tokens = features.dims[1];
         }
         return features;
@@ -3280,6 +3286,7 @@ export class WhisperForConditionalGeneration extends WhisperPreTrainedModel {
 
         if (generation_config.return_token_timestamps) {
             outputs["token_timestamps"] = this._extract_token_timestamps(
+                // @ts-expect-error TS2345
                 outputs,
                 generation_config.alignment_heads,
                 generation_config.num_frames,
@@ -3315,6 +3322,7 @@ export class WhisperForConditionalGeneration extends WhisperPreTrainedModel {
             );
         }
 
+        // @ts-expect-error TS2339
         let median_filter_width = this.config.median_filter_width;
         if (median_filter_width === undefined) {
             console.warn("Model config has no `median_filter_width`, using default value of 7.")
@@ -3325,6 +3333,7 @@ export class WhisperForConditionalGeneration extends WhisperPreTrainedModel {
         const batch = generate_outputs.cross_attentions;
         // Create a list with `decoder_layers` elements, each a tensor of shape
         // (batch size, attention_heads, output length, input length).
+        // @ts-expect-error TS2339
         const cross_attentions = Array.from({ length: this.config.decoder_layers },
             // Concatenate the cross attentions for each layer across sequence length dimension.
             (_, i) => cat(batch.map(x => x[i]), 2)
@@ -3468,6 +3477,7 @@ export class LlavaForConditionalGeneration extends LlavaPreTrainedModel {
         attention_mask,
     }) {
 
+        // @ts-expect-error TS2339
         const image_token_index = this.config.image_token_index;
 
         const idsList = input_ids.tolist();
@@ -4453,6 +4463,7 @@ export class Qwen2VLForConditionalGeneration extends Qwen2VLPreTrainedModel {
                 const image_nums = vision_tokens.filter(x => x == image_token_id).length;
                 const video_nums = vision_tokens.filter(x => x == video_token_id).length;
 
+                /** @type {number[][]} */
                 let llm_pos_ids_list = [];
                 let st = 0;
                 let remain_images = image_nums;
@@ -4522,6 +4533,7 @@ export class Qwen2VLForConditionalGeneration extends Qwen2VLPreTrainedModel {
                 // NOTE: Each item in llm_pos_ids_list is an array of shape (3, text_len),
                 // meaning to perform concatenation along dim=1, we can do the following:
                 const num_items = llm_pos_ids_list.reduce((acc, x) => acc + x.length, 0);
+                /** @type {number[]} */
                 const llm_positions = new Array(num_items);
                 let index = 0;
                 for (let x = 0; x < 3; ++x) {
@@ -4562,9 +4574,10 @@ export class Qwen2VLForConditionalGeneration extends Qwen2VLPreTrainedModel {
                     { length: 3 * data.length },
                     (_, i) => data[i % data.length]
                 );
+                /** @type {bigint[]} */
                 const mrope_position_deltas = Array.from(
                     { length: dims[0] },
-                    (_, i) => max(data.subarray(dims[1] * i, dims[1] * (i + 1)))[0] + 1 + dims[1]
+                    (_, i) => max(data.subarray(dims[1] * i, dims[1] * (i + 1)))[0] + 1n + BigInt(dims[1])
                 );
 
                 return [
@@ -5135,7 +5148,7 @@ export class DPTModel extends DPTPreTrainedModel { }
  * 
  * **Example:** Depth estimation w/ `Xenova/dpt-hybrid-midas`.
  * ```javascript
- * import { DPTForDepthEstimation, AutoProcessor, RawImage, interpolate, max } from '@huggingface/transformers';
+ * import { DPTForDepthEstimation, AutoProcessor, RawImage, interpolate_4d } from '@huggingface/transformers';
  * 
  * // Load model and processor
  * const model_id = 'Xenova/dpt-hybrid-midas';
@@ -5144,7 +5157,7 @@ export class DPTModel extends DPTPreTrainedModel { }
  * 
  * // Load image from URL
  * const url = 'http://images.cocodataset.org/val2017/000000039769.jpg';
- * const image = await RawImage.fromURL(url);
+ * const image = await RawImage.read(url);
  * 
  * // Prepare image for the model
  * const inputs = await processor(image);
@@ -5153,10 +5166,15 @@ export class DPTModel extends DPTPreTrainedModel { }
  * const { predicted_depth } = await model(inputs);
  * 
  * // Interpolate to original size
- * const prediction = interpolate(predicted_depth, image.size.reverse(), 'bilinear', false);
+ * const prediction = (await interpolate_4d(predicted_depth.unsqueeze(1), {
+     * size: image.size.reverse(),
+     * mode: 'bilinear',
+ * })).squeeze(1);
  * 
  * // Visualize the prediction
- * const formatted = prediction.mul_(255 / max(prediction.data)[0]).to('uint8');
+ * const min = prediction.min().item();
+ * const max = prediction.max().item();
+ * const formatted = prediction.sub_(min).div_(max - min).mul_(255).to('uint8');
  * const depth = RawImage.fromTensor(formatted);
  * // RawImage {
  * //   data: Uint8Array(307200) [ 85, 85, 84, ... ],
@@ -5206,11 +5224,7 @@ export class GLPNPreTrainedModel extends PreTrainedModel { }
 export class GLPNModel extends GLPNPreTrainedModel { }
 
 /**
- * GLPN Model transformer with a lightweight depth estimation head on top e.g. for KITTI, NYUv2.
- * 
- * **Example:** Depth estimation w/ `Xenova/glpn-kitti`.
- * ```javascript
- * import { GLPNForDepthEstimation, AutoProcessor, RawImage, interpolate, max } from '@huggingface/transformers';
+ * import { GLPNForDepthEstimation, AutoProcessor, RawImage, interpolate_4d } from '@huggingface/transformers';
  * 
  * // Load model and processor
  * const model_id = 'Xenova/glpn-kitti';
@@ -5219,7 +5233,7 @@ export class GLPNModel extends GLPNPreTrainedModel { }
  * 
  * // Load image from URL
  * const url = 'http://images.cocodataset.org/val2017/000000039769.jpg';
- * const image = await RawImage.fromURL(url);
+ * const image = await RawImage.read(url);
  * 
  * // Prepare image for the model
  * const inputs = await processor(image);
@@ -5228,13 +5242,18 @@ export class GLPNModel extends GLPNPreTrainedModel { }
  * const { predicted_depth } = await model(inputs);
  * 
  * // Interpolate to original size
- * const prediction = interpolate(predicted_depth, image.size.reverse(), 'bilinear', false);
+ * const prediction = (await interpolate_4d(predicted_depth.unsqueeze(1), {
+     * size: image.size.reverse(),
+     * mode: 'bilinear',
+ * })).squeeze(1);
  * 
  * // Visualize the prediction
- * const formatted = prediction.mul_(255 / max(prediction.data)[0]).to('uint8');
+ * const min = prediction.min().item();
+ * const max = prediction.max().item();
+ * const formatted = prediction.sub_(min).div_(max - min).mul_(255).to('uint8');
  * const depth = RawImage.fromTensor(formatted);
  * // RawImage {
- * //   data: Uint8Array(307200) [ 207, 169, 154, ... ],
+ * //   data: Uint8Array(307200) [ 85, 85, 84, ... ],
  * //   width: 640,
  * //   height: 480,
  * //   channels: 1
@@ -6201,10 +6220,12 @@ export class SpeechT5ForTextToSpeech extends SpeechT5PreTrainedModel {
 
         const { encoder_outputs, encoder_attention_mask } = await encoderForward(this, model_inputs);
 
+        // @ts-expect-error TS2339
         const r = encoder_outputs.dims[1] / this.config.reduction_factor;
         const maxlen = Math.floor(r * maxlenratio);
         const minlen = Math.floor(r * minlenratio);
 
+        // @ts-expect-error TS2339
         const num_mel_bins = this.config.num_mel_bins;
 
         let spectrogramParts = [];
@@ -6569,11 +6590,13 @@ export class MusicgenForConditionalGeneration extends PreTrainedModel { // NOTE:
      */
     _apply_and_filter_by_delay_pattern_mask(outputs) {
         const [bs_x_codebooks, seqLength] = outputs.dims;
+        // @ts-expect-error TS2339
         const num_codebooks = this.config.decoder.num_codebooks;
         const upperBound = (seqLength - num_codebooks);
 
         let newDataSize = 0;
         for (let i = 0; i < outputs.size; ++i) {
+            // @ts-expect-error TS2339
             if (outputs.data[i] === this.config.decoder.pad_token_id) {
                 continue;
             }
@@ -6603,7 +6626,9 @@ export class MusicgenForConditionalGeneration extends PreTrainedModel { // NOTE:
         let clonedInputIds = structuredClone(input_ids);
         for (let i = 0; i < clonedInputIds.length; ++i) {
             for (let j = 0; j < clonedInputIds[i].length; ++j) {
+                // @ts-expect-error TS2339
                 if ((i % this.config.decoder.num_codebooks) >= j) {
+                    // @ts-expect-error TS2339
                     clonedInputIds[i][j] = BigInt(this.config.decoder.pad_token_id);
                 }
             }
@@ -6760,6 +6785,9 @@ export class MultiModalityCausalLM extends MultiModalityPreTrainedModel {
         'past_key_values',
     ];
 
+    /**
+     * @param {ConstructorParameters<typeof MultiModalityPreTrainedModel>} args
+     */
     constructor(...args) {
         super(...args);
 
@@ -7728,10 +7756,17 @@ export class SequenceClassifierOutput extends ModelOutput {
     /**
      * @param {Object} output The output of the model.
      * @param {Tensor} output.logits classification (or regression if config.num_labels==1) scores (before SoftMax).
+     * @param {Record<string, Tensor>} [output.attentions] Object of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`.
+     * Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
      */
-    constructor({ logits }) {
+    constructor({ logits, ...attentions }) {
         super();
         this.logits = logits;
+        const attentions_list = Object.values(attentions);
+        if (attentions_list.length > 0) {
+            // Only set attentions if they are not empty
+            this.attentions = attentions_list;
+        }
     }
 }
 

package/src/ops/registry.js

@@ -36,6 +36,16 @@ export class TensorOpRegistry {
         // executionProviders: ['webgpu'],
     };
 
+    static get nearest_interpolate_4d() {
+        if (!this._nearest_interpolate_4d) {
+            this._nearest_interpolate_4d = wrap(
+                [8, 10, 18, 0, 58, 129, 1, 10, 41, 10, 1, 120, 10, 0, 10, 0, 10, 1, 115, 18, 1, 121, 34, 6, 82, 101, 115, 105, 122, 101, 42, 18, 10, 4, 109, 111, 100, 101, 34, 7, 110, 101, 97, 114, 101, 115, 116, 160, 1, 3, 18, 1, 114, 90, 31, 10, 1, 120, 18, 26, 10, 24, 8, 1, 18, 20, 10, 3, 18, 1, 98, 10, 3, 18, 1, 99, 10, 3, 18, 1, 104, 10, 3, 18, 1, 119, 90, 15, 10, 1, 115, 18, 10, 10, 8, 8, 7, 18, 4, 10, 2, 8, 4, 98, 31, 10, 1, 121, 18, 26, 10, 24, 8, 1, 18, 20, 10, 3, 18, 1, 98, 10, 3, 18, 1, 99, 10, 3, 18, 1, 104, 10, 3, 18, 1, 119, 66, 2, 16, 21],
+                this.session_options,
+                'y',
+            );
+        }
+        return this._nearest_interpolate_4d;
+    }
     static get bilinear_interpolate_4d() {
         if (!this._bilinear_interpolate_4d) {
             this._bilinear_interpolate_4d = wrap(