@huggingface/transformers 4.0.0-next.1 → 4.0.0-next.10

package/src/configs.js

@@ -68,12 +68,20 @@ function getNormalizedConfig(config) {
         case 'florence2':
         case 'llava_onevision':
         case 'idefics3':
+        case 'granite_speech':
         case 'ultravox':
         case 'voxtral':
+        case 'voxtral_realtime':
         case 'smolvlm':
         case 'gemma3n':
+        case 'lfm2_vl':
         case 'chatterbox':
+        case 'lighton_ocr':
+        case 'glm_ocr':
         case 'mistral3':
+        case 'qwen2_5_vl':
+        case 'qwen3_vl':
+        case 'qwen3_vl_moe':
             // @ts-expect-error TS2339
             init_normalized_config = getNormalizedConfig(config.text_config);
             break;
@@ -115,6 +123,7 @@ function getNormalizedConfig(config) {
         case 'nanochat':
         case 'apertus':
         case 'arcee':
+        case 'afmoe':
         case 'lfm2':
         case 'lfm2_moe':
         case 'smollm3':
@@ -125,10 +134,19 @@ function getNormalizedConfig(config) {
         case 'granite':
         case 'granitemoehybrid':
         case 'cohere':
+        case 'cohere2':
         case 'mistral':
+        case 'voxtral_realtime_text':
+        case 'voxtral_realtime_encoder':
         case 'starcoder2':
         case 'qwen2':
+        case 'qwen2_moe':
         case 'qwen2_vl':
+        case 'qwen2_vl_text':
+        case 'qwen2_5_vl_text':
+        case 'qwen3_moe':
+        case 'qwen3_vl_text':
+        case 'qwen3_vl_moe_text':
         case 'phi':
         case 'phi3':
         case 'phi3_v':
@@ -140,6 +158,8 @@ function getNormalizedConfig(config) {
             mapping['dim_kv'] = 'head_dim';
             break;
         case 'qwen3':
+        case 'solar_open':
+        case 'glm_ocr_text':
         case 'gemma':
         case 'gemma2':
         case 'vaultgemma':
@@ -150,6 +170,7 @@ function getNormalizedConfig(config) {
         case 'ernie4_5':
         case 'hunyuan_v1_dense':
         case 'falcon_h1':
+        case 'nemotron_h':
         case 'ministral':
         case 'ministral3':
             mapping['num_heads'] = 'num_key_value_heads';
@@ -184,6 +205,9 @@ function getNormalizedConfig(config) {
             mapping['num_attention_heads'] = 'num_attention_heads';
             break;
         case 'youtu':
+        case 'deepseek_v3':
+        case 'glm_moe_dsa':
+        case 'mistral4':
             mapping['num_heads'] = 'num_key_value_heads';
             mapping['num_layers'] = 'num_hidden_layers';
             mapping['dim_kv'] = 'qk_head_dim';
@@ -242,6 +266,24 @@ function getNormalizedConfig(config) {
             mapping['num_encoder_heads'] = 'encoder_num_key_value_heads';
             mapping['encoder_hidden_size'] = mapping['decoder_hidden_size'] = 'hidden_size';
             break;
+        case 'cohere_asr':
+            mapping['num_decoder_layers'] = 'num_hidden_layers';
+            mapping['num_decoder_heads'] = 'num_key_value_heads';
+            mapping['decoder_hidden_size'] = 'hidden_size';
+            mapping['decoder_dim_kv'] = 'head_dim';
+            const {
+                num_hidden_layers: num_encoder_layers,
+                num_attention_heads: num_encoder_heads,
+                hidden_size: encoder_hidden_size,
+            } = /** @type {any} */ (config).encoder_config;
+            init_normalized_config = {
+                num_encoder_layers,
+                num_encoder_heads,
+                encoder_hidden_size,
+                // @ts-expect-error TS2339
+                encoder_dim_kv: config.head_dim,
+            };
+            break;
         case 'vision-encoder-decoder':
             // @ts-expect-error TS2339
             const decoderConfig = getNormalizedConfig(config.decoder);
@@ -283,17 +325,20 @@ function getNormalizedConfig(config) {
  * @returns {Record<string, number[]>}
  */
 export function getCacheShapes(config, options) {
+    if (!(config instanceof PretrainedConfig)) {
+        config = new PretrainedConfig(config);
+    }
+
+    const batch_size = options?.batch_size ?? 1;
     if (['lfm2', 'lfm2_moe'].includes(config.model_type)) {
         const pkv_prefix = options?.prefix ?? 'past_key_values';
         const conv_prefix = pkv_prefix === 'present' ? 'present' : 'past';
 
-        // Custom caching mechanism for LFM2
         /** @type {Record<string, number[]>} */
         const cache_values = {};
-        // @ts-expect-error TS2339
-        const { layer_types, num_attention_heads, num_key_value_heads, hidden_size, conv_L_cache } = config;
+        const { layer_types, num_attention_heads, num_key_value_heads, hidden_size, conv_L_cache } =
+            /** @type {any} */ (config);
         const head_dim = hidden_size / num_attention_heads;
-        const batch_size = options?.batch_size ?? 1;
         for (let i = 0; i < layer_types.length; ++i) {
             if (layer_types[i] === 'full_attention') {
                 for (const kv of ['key', 'value']) {
@@ -306,44 +351,99 @@ export function getCacheShapes(config, options) {
             }
         }
         return cache_values;
-    } else if (['granitemoehybrid', 'falcon_h1'].includes(config.model_type)) {
+    } else if (['granitemoehybrid', 'falcon_h1', 'nemotron_h'].includes(config.model_type)) {
         const pkv_prefix = options?.prefix ?? 'past_key_values';
         const conv_prefix = pkv_prefix === 'present' ? 'present' : 'past';
 
+        const c = /** @type {any} */ (config);
+
+        // Normalize config field names across model types
+        const layer_types = c.layer_types ?? c.layers_block_type;
+        const num_layers = c.num_hidden_layers ?? layer_types?.length;
+        const num_key_value_heads = c.num_key_value_heads;
+        const head_dim = c.head_dim ?? c.hidden_size / c.num_attention_heads;
+        const mamba_n_heads = c.mamba_n_heads ?? c.mamba_num_heads;
+        const mamba_d_head = c.mamba_d_head ?? c.mamba_head_dim;
+        const mamba_d_state = c.mamba_d_state ?? c.ssm_state_size;
+        const mamba_n_groups = c.mamba_n_groups ?? c.n_groups;
+        const mamba_d_conv = c.mamba_d_conv ?? c.conv_kernel;
+        const mamba_d_ssm =
+            c.mamba_d_ssm ?? (c.mamba_expand ? c.mamba_expand * c.hidden_size : mamba_n_heads * mamba_d_head);
+        const conv_d_inner = mamba_d_ssm + 2 * mamba_n_groups * mamba_d_state;
+
         /** @type {Record<string, number[]>} */
         const cache_values = {};
 
+        for (let i = 0; i < num_layers; ++i) {
+            if (!layer_types || layer_types[i] === 'mamba') {
+                cache_values[`${conv_prefix}_conv.${i}`] = [batch_size, conv_d_inner, mamba_d_conv];
+                cache_values[`${conv_prefix}_ssm.${i}`] = [batch_size, mamba_n_heads, mamba_d_head, mamba_d_state];
+            }
+            if (!layer_types || layer_types[i] === 'attention') {
+                for (const kv of ['key', 'value']) {
+                    cache_values[`${pkv_prefix}.${i}.${kv}`] = [batch_size, num_key_value_heads, 0, head_dim];
+                }
+            }
+        }
+        return cache_values;
+    } else if (['qwen3_next', 'qwen3_5_text', 'qwen3_5_moe_text', 'olmo_hybrid'].includes(config.model_type)) {
+        const pkv_prefix = options?.prefix ?? 'past_key_values';
+        const conv_prefix = pkv_prefix === 'present' ? 'present' : 'past';
+
+        /** @type {Record<string, number[]>} */
+        const cache_values = {};
         const {
+            head_dim,
             layer_types,
-            num_hidden_layers,
             num_attention_heads,
             num_key_value_heads,
             hidden_size,
-            mamba_d_conv,
-            mamba_n_heads,
-            mamba_d_head,
-            mamba_d_state,
-            mamba_n_groups,
-            mamba_expand,
-            mamba_d_ssm,
+            linear_num_value_heads,
+            linear_num_key_heads,
+            linear_key_head_dim,
+            linear_value_head_dim,
+            linear_conv_kernel_dim,
         } = /** @type {any} */ (config);
-        const head_dim = hidden_size / num_attention_heads;
-        const batch_size = options?.batch_size ?? 1;
 
-        const conv_d_inner = (mamba_d_ssm ?? mamba_expand * hidden_size) + 2 * mamba_n_groups * mamba_d_state;
-        for (let i = 0; i < num_hidden_layers; ++i) {
-            if (!layer_types || layer_types[i] === 'mamba') {
-                cache_values[`${conv_prefix}_conv.${i}`] = [batch_size, conv_d_inner, mamba_d_conv];
-                cache_values[`${conv_prefix}_ssm.${i}`] = [batch_size, mamba_n_heads, mamba_d_head, mamba_d_state];
-            }
-            if (!layer_types || layer_types[i] === 'attention') {
+        const key_dim = linear_key_head_dim * linear_num_key_heads;
+        const value_dim = linear_value_head_dim * linear_num_value_heads;
+
+        const final_head_dim = head_dim ?? hidden_size / num_attention_heads;
+        for (let i = 0; i < layer_types.length; ++i) {
+            if (layer_types[i] === 'full_attention') {
                 for (const kv of ['key', 'value']) {
-                    cache_values[`${pkv_prefix}.${i}.${kv}`] = [batch_size, num_key_value_heads, 0, head_dim];
+                    cache_values[`${pkv_prefix}.${i}.${kv}`] = [batch_size, num_key_value_heads, 0, final_head_dim];
                 }
+            } else if (layer_types[i] === 'linear_attention') {
+                if (config.model_type === 'olmo_hybrid') {
+                    cache_values[`${conv_prefix}_conv.${i}.key`] = [batch_size, key_dim, linear_conv_kernel_dim];
+                    cache_values[`${conv_prefix}_conv.${i}.value`] = [batch_size, value_dim, linear_conv_kernel_dim];
+                    cache_values[`${conv_prefix}_conv.${i}.query`] = [batch_size, key_dim, linear_conv_kernel_dim];
+                } else {
+                    const conv_dim = key_dim * 2 + value_dim;
+                    cache_values[`${conv_prefix}_conv.${i}`] = [batch_size, conv_dim, linear_conv_kernel_dim];
+                }
+                cache_values[`${conv_prefix}_recurrent.${i}`] = [
+                    batch_size,
+                    linear_num_value_heads,
+                    linear_key_head_dim,
+                    linear_value_head_dim,
+                ];
+            } else {
+                throw new Error(`Unsupported layer type: ${layer_types[i]}`);
             }
         }
         return cache_values;
+    } else if (['lfm2_vl', 'qwen3_5', 'qwen3_5_moe', 'voxtral_realtime'].includes(config.model_type)) {
+        let subConfig;
+        if (config.model_type === 'voxtral_realtime' && options?.session_name === 'audio_encoder') {
+            subConfig = /** @type {any} */ (config).audio_config;
+        } else {
+            subConfig = /** @type {any} */ (config).text_config;
+        }
+        return getCacheShapes(subConfig, options);
     }
+
     return getKeyValueShapes(config, options);
 }
 

package/src/env.js

@@ -26,27 +26,43 @@ import fs from 'node:fs';
 import path from 'node:path';
 import url from 'node:url';
 
-const VERSION = '4.0.0-next.1';
+const VERSION = '4.0.0-next.10';
+
+const HAS_SELF = typeof self !== 'undefined';
 
-const IS_PROCESS_AVAILABLE = typeof process !== 'undefined';
-const IS_NODE_ENV = IS_PROCESS_AVAILABLE && process?.release?.name === 'node';
 const IS_FS_AVAILABLE = !isEmpty(fs);
 const IS_PATH_AVAILABLE = !isEmpty(path);
+const IS_WEB_CACHE_AVAILABLE = HAS_SELF && 'caches' in self;
 
 // Runtime detection
 const IS_DENO_RUNTIME = typeof globalThis.Deno !== 'undefined';
 const IS_BUN_RUNTIME = typeof globalThis.Bun !== 'undefined';
 
+const IS_DENO_WEB_RUNTIME = IS_DENO_RUNTIME && IS_WEB_CACHE_AVAILABLE && !IS_FS_AVAILABLE;
+
+const IS_PROCESS_AVAILABLE = typeof process !== 'undefined';
+const IS_NODE_ENV = IS_PROCESS_AVAILABLE && process?.release?.name === 'node' && !IS_DENO_WEB_RUNTIME;
+
 // Check if various APIs are available (depends on environment)
 const IS_BROWSER_ENV = typeof window !== 'undefined' && typeof window.document !== 'undefined';
 const IS_WEBWORKER_ENV =
-    typeof self !== 'undefined' &&
+    HAS_SELF &&
     ['DedicatedWorkerGlobalScope', 'ServiceWorkerGlobalScope', 'SharedWorkerGlobalScope'].includes(
         self.constructor?.name,
     );
-const IS_WEB_CACHE_AVAILABLE = typeof self !== 'undefined' && 'caches' in self;
+const IS_WEB_ENV = IS_BROWSER_ENV || IS_WEBWORKER_ENV || IS_DENO_WEB_RUNTIME;
+
 const IS_WEBGPU_AVAILABLE = IS_NODE_ENV || (typeof navigator !== 'undefined' && 'gpu' in navigator);
 const IS_WEBNN_AVAILABLE = typeof navigator !== 'undefined' && 'ml' in navigator;
+const IS_CRYPTO_AVAILABLE = typeof crypto !== 'undefined' && typeof crypto.getRandomValues === 'function';
+
+const IS_CHROME_AVAILABLE =
+    // @ts-ignore - chrome may not exist in all environments
+    typeof chrome !== 'undefined' && typeof chrome.runtime !== 'undefined' && typeof chrome.runtime.id === 'string';
+
+const IS_SERVICE_WORKER_ENV =
+    // @ts-ignore - ServiceWorkerGlobalScope may not exist in all environments
+    typeof ServiceWorkerGlobalScope !== 'undefined' && HAS_SELF && self instanceof ServiceWorkerGlobalScope;
 
 /**
  * Check if the current environment is Safari browser.
@@ -86,6 +102,15 @@ export const apis = Object.freeze({
     /** Whether we are running in a web worker environment */
     IS_WEBWORKER_ENV,
 
+    /** Whether we are running in a web-like environment (browser, web worker, or Deno web runtime) */
+    IS_WEB_ENV,
+
+    /** Whether we are running in a service worker environment */
+    IS_SERVICE_WORKER_ENV,
+
+    /** Whether we are running in Deno's web runtime (CDN imports, Cache API available, no filesystem) */
+    IS_DENO_WEB_RUNTIME,
+
     /** Whether the Cache API is available */
     IS_WEB_CACHE_AVAILABLE,
 
@@ -109,6 +134,12 @@ export const apis = Object.freeze({
 
     /** Whether the path API is available */
     IS_PATH_AVAILABLE,
+
+    /** Whether the crypto API is available */
+    IS_CRYPTO_AVAILABLE,
+
+    /** Whether the Chrome runtime API is available */
+    IS_CHROME_AVAILABLE,
 });
 
 const RUNNING_LOCALLY = IS_FS_AVAILABLE && IS_PATH_AVAILABLE;
@@ -134,12 +165,48 @@ const DEFAULT_CACHE_DIR = RUNNING_LOCALLY ? path.join(dirname__, '/.cache/') : n
 const DEFAULT_LOCAL_MODEL_PATH = '/models/';
 const localModelPath = RUNNING_LOCALLY ? path.join(dirname__, DEFAULT_LOCAL_MODEL_PATH) : DEFAULT_LOCAL_MODEL_PATH;
 
+// Ensure default fetch is called with the correct receiver in browser environments.
+const DEFAULT_FETCH = typeof globalThis.fetch === 'function' ? globalThis.fetch.bind(globalThis) : undefined;
+
+/**
+ * Log levels for controlling output verbosity.
+ *
+ * Each level is represented by a number, where higher numbers include all lower level messages.
+ * Use these values to set `env.logLevel`.
+ *
+ * @example
+ * import { env, LogLevel } from '@huggingface/transformers';
+ *
+ * // Set log level to show only errors
+ * env.logLevel = LogLevel.ERROR;
+ *
+ * // Set log level to show errors, warnings, and info
+ * env.logLevel = LogLevel.INFO;
+ *
+ * // Disable all logging
+ * env.logLevel = LogLevel.NONE;
+ *
+ */
+export const LogLevel = Object.freeze({
+    /** All messages including debug output (value: 10) */
+    DEBUG: 10,
+    /** Errors, warnings, and info messages (value: 20) */
+    INFO: 20,
+    /** Errors and warnings (value: 30) */
+    WARNING: 30,
+    /** Only error messages (value: 40) */
+    ERROR: 40,
+    /** No logging output (value: 50) */
+    NONE: 50,
+});
+
 /**
  * Global variable given visible to users to control execution. This provides users a simple way to configure Transformers.js.
  * @typedef {Object} TransformersEnvironment
  * @property {string} version This version of Transformers.js.
- * @property {{onnx: Partial<import('onnxruntime-common').Env>}} backends Expose environment variables of different backends,
+ * @property {{onnx: Partial<import('onnxruntime-common').Env> & { setLogLevel?: (logLevel: number) => void }}} backends Expose environment variables of different backends,
  * allowing users to set these variables if they want to.
+ * @property {number} logLevel The logging level. Use LogLevel enum values. Defaults to LogLevel.ERROR.
  * @property {boolean} allowRemoteModels Whether to allow loading of remote files, defaults to `true`.
  * If set to `false`, it will have the same effect as setting `local_files_only=true` when loading pipelines, models, tokenizers, processors, etc.
  * @property {string} remoteHost Host URL to load models from. Defaults to the Hugging Face Hub.
@@ -154,12 +221,18 @@ const localModelPath = RUNNING_LOCALLY ? path.join(dirname__, DEFAULT_LOCAL_MODE
  * @property {boolean} useCustomCache Whether to use a custom cache system (defined by `customCache`), defaults to `false`.
  * @property {import('./utils/cache.js').CacheInterface|null} customCache The custom cache to use. Defaults to `null`. Note: this must be an object which
  * implements the `match` and `put` functions of the Web Cache API. For more information, see https://developer.mozilla.org/en-US/docs/Web/API/Cache.
- * @property {boolean} useWasmCache Whether to pre-load and cache WASM binaries for ONNX Runtime. Defaults to `true` when cache is available.
- * This can improve performance by avoiding repeated downloads of WASM files. Note: Only the WASM binary is cached.
- * The MJS loader file still requires network access unless you use a Service Worker.
+ * @property {boolean} useWasmCache Whether to pre-load and cache WASM binaries and the WASM factory (.mjs) for ONNX Runtime.
+ * Defaults to `true` when cache is available. This can improve performance and enables offline usage by avoiding repeated downloads.
  * @property {string} cacheKey The cache key to use for storing models and WASM binaries. Defaults to 'transformers-cache'.
+ * @property {boolean} experimental_useCrossOriginStorage Whether to use the Cross-Origin Storage API to cache model files
+ * across origins, allowing different sites to share the same cached model weights. Defaults to `false`.
+ * Requires the Cross-Origin Storage Chrome extension: {@link https://chromewebstore.google.com/detail/cross-origin-storage/denpnpcgjgikjpoglpjefakmdcbmlgih}.
+ * The `experimental_` prefix indicates that the underlying browser API is not yet standardised and may change or be
+ * removed without a major version bump. For more information, see {@link https://github.com/WICG/cross-origin-storage}.
+ * @property {(input: string | URL, init?: any) => Promise<any>} fetch The fetch function to use. Defaults to `fetch`.
  */
 
+let logLevel = LogLevel.WARNING; // Default log level
 /** @type {TransformersEnvironment} */
 export const env = {
     version: VERSION,
@@ -171,17 +244,27 @@ export const env = {
         onnx: {},
     },
 
+    /////////////////// Logging settings ///////////////////
+    get logLevel() {
+        return logLevel;
+    },
+    set logLevel(level) {
+        logLevel = level;
+
+        // invoke hook to set ONNX Runtime log level when Transformers.js log level changes
+        env.backends.onnx?.setLogLevel?.(level);
+    },
     /////////////////// Model settings ///////////////////
     allowRemoteModels: true,
     remoteHost: 'https://huggingface.co/',
     remotePathTemplate: '{model}/resolve/{revision}/',
 
-    allowLocalModels: !(IS_BROWSER_ENV || IS_WEBWORKER_ENV),
+    allowLocalModels: !(IS_BROWSER_ENV || IS_WEBWORKER_ENV || IS_DENO_WEB_RUNTIME), // Default to true for non-web environments, false for web environments
     localModelPath: localModelPath,
     useFS: IS_FS_AVAILABLE,
 
     /////////////////// Cache settings ///////////////////
-    useBrowserCache: IS_WEB_CACHE_AVAILABLE && !IS_DENO_RUNTIME,
+    useBrowserCache: IS_WEB_CACHE_AVAILABLE,
 
     useFSCache: IS_FS_AVAILABLE,
     cacheDir: DEFAULT_CACHE_DIR,
@@ -191,6 +274,12 @@ export const env = {
 
     useWasmCache: IS_WEB_CACHE_AVAILABLE || IS_FS_AVAILABLE,
     cacheKey: 'transformers-cache',
+
+    experimental_useCrossOriginStorage: false,
+
+    /////////////////// Custom fetch /////////////////////
+    fetch: DEFAULT_FETCH,
+
     //////////////////////////////////////////////////////
 };
 

package/src/generation/logits_sampler.js

@@ -6,6 +6,7 @@ import { Callable } from '../utils/generic.js';
 import { Tensor, topk } from '../utils/tensor.js';
 
 import { max, softmax } from '../utils/maths.js';
+import { _weightedIndex } from '../utils/random.js';
 import { GenerationConfig } from '../generation/configuration_utils.js';
 
 /**
@@ -64,24 +65,11 @@ export class LogitsSampler extends Callable {
 
     /**
      * Selects an item randomly based on the specified probabilities.
-     * @param {import("../transformers.js").DataArray} probabilities An array of probabilities to use for selection.
+     * @param {Float32Array} probabilities An array of probabilities to use for selection.
      * @returns {number} The index of the selected item.
      */
     randomSelect(probabilities) {
-        // Return index of chosen item
-        let sumProbabilities = 0;
-        for (let i = 0; i < probabilities.length; ++i) {
-            sumProbabilities += probabilities[i];
-        }
-
-        let r = Math.random() * sumProbabilities;
-        for (let i = 0; i < probabilities.length; ++i) {
-            r -= probabilities[i];
-            if (r <= 0) {
-                return i;
-            }
-        }
-        return 0; // return first (most probable) as a fallback
+        return _weightedIndex(probabilities);
     }
 
     /**

package/src/generation/parameters.js

@@ -21,7 +21,7 @@
  * Custom logits processors that complement the default logits processors built from arguments and
  * generation config. If a logit processor is passed that is already created with the arguments or a
  * generation config an error is thrown. This feature is intended for advanced users.
- * @property {import('./stopping_criteria.js').StoppingCriteriaList} [stopping_criteria=null] (`StoppingCriteriaList`, *optional*):
+ * @property {import('./stopping_criteria.js').StoppingCriteria|import('./stopping_criteria.js').StoppingCriteria[]|import('./stopping_criteria.js').StoppingCriteriaList} [stopping_criteria=null] (`StoppingCriteriaList`, *optional*):
  * Custom stopping criteria that complements the default stopping criteria built from arguments and a
  * generation config. If a stopping criteria is passed that is already created with the arguments or a
  * generation config an error is thrown. This feature is intended for advanced users.

package/src/generation/streamers.js

@@ -70,6 +70,9 @@ export class TextStreamer extends BaseStreamer {
         this.token_cache = [];
         this.print_len = 0;
         this.next_tokens_are_prompt = true;
+
+        // Track special token IDs for special handling during streaming.
+        this.special_ids = new Set(this.tokenizer.all_special_ids.map(BigInt));
     }
 
     /**
@@ -90,6 +93,24 @@ export class TextStreamer extends BaseStreamer {
         const tokens = value[0];
         this.token_callback_function?.(tokens);
 
+        // Handle special tokens: flush any existing text, then print or skip them
+        if (tokens.length === 1 && this.special_ids.has(tokens[0])) {
+            if (this.decode_kwargs.skip_special_tokens) return;
+
+            // Flush any existing cached text first
+            if (this.token_cache.length > 0) {
+                const text = this.tokenizer.decode(this.token_cache, this.decode_kwargs);
+                const printable_text = text.slice(this.print_len);
+                this.on_finalized_text(printable_text, false);
+                this.token_cache = [];
+                this.print_len = 0;
+            }
+            // Print the special token immediately
+            const special_text = this.tokenizer.decode(tokens, this.decode_kwargs);
+            this.on_finalized_text(special_text, false);
+            return;
+        }
+
         // Add the new token to the cache and decodes the entire thing.
         this.token_cache = mergeArrays(this.token_cache, tokens);
         const text = this.tokenizer.decode(this.token_cache, this.decode_kwargs);

package/src/image_processors_utils.js

@@ -5,6 +5,7 @@ import { RawImage } from './utils/image.js';
 import { calculateReflectOffset } from './utils/core.js';
 import { getModelJSON } from './utils/hub.js';
 import { IMAGE_PROCESSOR_NAME } from './utils/constants.js';
+import { logger } from './utils/logger.js';
 
 /**
  * Named tuple to indicate the order we are using is (height x width),
@@ -13,7 +14,7 @@ import { IMAGE_PROCESSOR_NAME } from './utils/constants.js';
  */
 
 /**
- * @typedef {object} ImageProcessorResult
+ * @typedef {Object} ImageProcessorResult
  * @property {Tensor} pixel_values The pixel values of the batched preprocessed images.
  * @property {HeightWidth[]} original_sizes Array of two-dimensional tuples like [[480, 640]].
  * @property {HeightWidth[]} reshaped_input_sizes Array of two-dimensional tuples like [[1000, 1330]].
@@ -403,13 +404,24 @@ function compute_segments(
  * @param {number} [factor=28] The factor to use for resizing.
  * @param {number} [min_pixels=56*56] The minimum number of pixels.
  * @param {number} [max_pixels=14*14*4*1280] The maximum number of pixels.
- * @returns {[number, number]} The new height and width of the image.
+ * @param {number} [temporal_factor=1] The temporal factor to include in the pixel budget (e.g. temporal_patch_size for video/3D models).
+ * @returns {[number, number]} The new width and height of the image.
  * @throws {Error} If the height or width is smaller than the factor.
  */
-function smart_resize(height, width, factor = 28, min_pixels = 56 * 56, max_pixels = 14 * 14 * 4 * 1280) {
+export function smart_resize(
+    height,
+    width,
+    factor = 28,
+    min_pixels = 56 * 56,
+    max_pixels = 14 * 14 * 4 * 1280,
+    temporal_factor = 1,
+) {
     if (height < factor || width < factor) {
-        throw new Error(`height:${height} or width:${width} must be larger than factor:${factor}`);
-    } else if (Math.max(height, width) / Math.min(height, width) > 200) {
+        const scale = Math.max(factor / height, factor / width);
+        height = Math.round(height * scale);
+        width = Math.round(width * scale);
+    }
+    if (Math.max(height, width) / Math.min(height, width) > 200) {
         throw new Error(
             `absolute aspect ratio must be smaller than 200, got ${Math.max(height, width) / Math.min(height, width)}`,
         );
@@ -418,17 +430,17 @@ function smart_resize(height, width, factor = 28, min_pixels = 56 * 56, max_pixe
     let h_bar = Math.round(height / factor) * factor;
     let w_bar = Math.round(width / factor) * factor;
 
-    if (h_bar * w_bar > max_pixels) {
-        const beta = Math.sqrt((height * width) / max_pixels);
-        h_bar = Math.floor(height / beta / factor) * factor;
-        w_bar = Math.floor(width / beta / factor) * factor;
-    } else if (h_bar * w_bar < min_pixels) {
-        const beta = Math.sqrt(min_pixels / (height * width));
+    if (temporal_factor * h_bar * w_bar > max_pixels) {
+        const beta = Math.sqrt((temporal_factor * height * width) / max_pixels);
+        h_bar = Math.max(factor, Math.floor(height / beta / factor) * factor);
+        w_bar = Math.max(factor, Math.floor(width / beta / factor) * factor);
+    } else if (temporal_factor * h_bar * w_bar < min_pixels) {
+        const beta = Math.sqrt(min_pixels / (temporal_factor * height * width));
         h_bar = Math.ceil((height * beta) / factor) * factor;
         w_bar = Math.ceil((width * beta) / factor) * factor;
     }
 
-    return [h_bar, w_bar];
+    return [w_bar, h_bar];
 }
 
 /**
@@ -450,7 +462,7 @@ export function post_process_panoptic_segmentation(
     target_sizes = null,
 ) {
     if (label_ids_to_fuse === null) {
-        console.warn('`label_ids_to_fuse` unset. No instance will be fused.');
+        logger.warn('`label_ids_to_fuse` unset. No instance will be fused.');
         label_ids_to_fuse = new Set();
     }
 
@@ -592,6 +604,7 @@ export class ImageProcessor extends Callable {
         if (
             this.do_pad &&
             !this.pad_size &&
+            !this.size_divisibility &&
             this.size &&
             this.size.width !== undefined &&
             this.size.height !== undefined
@@ -864,11 +877,6 @@ export class ImageProcessor extends Callable {
             return [newWidth, newHeight];
         } else if (this.size_divisibility !== undefined) {
             return enforce_size_divisibility([srcWidth, srcHeight], this.size_divisibility);
-        } else if (this.min_pixels !== undefined && this.max_pixels !== undefined) {
-            // Custom resize logic for Qwen2-VL models
-            // @ts-expect-error TS2339
-            const factor = this.config.patch_size * this.config.merge_size;
-            return smart_resize(srcHeight, srcWidth, factor, this.min_pixels, this.max_pixels);
         } else {
             throw new Error(
                 `Could not resize image due to unsupported \`this.size\` option in config: ${JSON.stringify(size)}`,
@@ -890,7 +898,7 @@ export class ImageProcessor extends Callable {
     }
 
     /**
-     * @typedef {object} PreprocessedImage
+     * @typedef {Object} PreprocessedImage
      * @property {HeightWidth} original_size The original size of the image.
      * @property {HeightWidth} reshaped_input_size The reshaped input size of the image.
      * @property {Tensor} pixel_values The pixel values of the preprocessed image.
@@ -1000,10 +1008,8 @@ export class ImageProcessor extends Callable {
                 const padded = this.pad_image(pixelData, [image.height, image.width, image.channels], this.pad_size);
                 [pixelData, imgDims] = padded; // Update pixel data and image dimensions
             } else if (this.size_divisibility) {
-                const [paddedWidth, paddedHeight] = enforce_size_divisibility(
-                    [imgDims[1], imgDims[0]],
-                    this.size_divisibility,
-                );
+                const paddedWidth = Math.ceil(imgDims[1] / this.size_divisibility) * this.size_divisibility;
+                const paddedHeight = Math.ceil(imgDims[0] / this.size_divisibility) * this.size_divisibility;
                 [pixelData, imgDims] = this.pad_image(pixelData, imgDims, { width: paddedWidth, height: paddedHeight });
             }
         }

package/src/models/afmoe/modeling_afmoe.js

@@ -0,0 +1,5 @@
+import { PreTrainedModel } from '../modeling_utils.js';
+
+export class AfmoePreTrainedModel extends PreTrainedModel {}
+export class AfmoeModel extends AfmoePreTrainedModel {}
+export class AfmoeForCausalLM extends AfmoePreTrainedModel {}

package/src/models/auto/image_processing_auto.js

@@ -2,6 +2,7 @@ import { getModelJSON } from '../../utils/hub.js';
 import { ImageProcessor } from '../../image_processors_utils.js';
 import * as AllImageProcessors from '../image_processors.js';
 import { GITHUB_ISSUE_URL, IMAGE_PROCESSOR_NAME } from '../../utils/constants.js';
+import { logger } from '../../utils/logger.js';
 
 export class AutoImageProcessor {
     /** @type {typeof ImageProcessor.from_pretrained} */
@@ -20,7 +21,7 @@ export class AutoImageProcessor {
         if (!image_processor_class) {
             if (key !== undefined) {
                 // Only log a warning if the class is not found and the key is set.
-                console.warn(
+                logger.warn(
                     `Image processor type '${key}' not found, assuming base ImageProcessor. Please report this at ${GITHUB_ISSUE_URL}.`,
                 );
             }