@huggingface/transformers 4.0.0-next.8 → 4.0.0-next.9

package/dist/transformers.web.js

@@ -14,7 +14,7 @@ var node_path_default = {};
 var node_url_default = {};
 
 // src/env.js
-var VERSION = "4.0.0-next.8";
+var VERSION = "4.0.0-next.9";
 var HAS_SELF = typeof self !== "undefined";
 var IS_FS_AVAILABLE = !isEmpty(node_fs_default);
 var IS_PATH_AVAILABLE = !isEmpty(node_path_default);
@@ -11830,6 +11830,7 @@ var processors_exports = {};
 __export(processors_exports, {
   ChatterboxProcessor: () => ChatterboxProcessor,
   Florence2Processor: () => Florence2Processor,
+  Gemma3Processor: () => Gemma3Processor,
   Gemma3nProcessor: () => Gemma3nProcessor,
   Glm46VProcessor: () => Glm46VProcessor,
   GraniteSpeechProcessor: () => GraniteSpeechProcessor,
@@ -14829,6 +14830,7 @@ __export(image_processors_exports, {
   DonutImageProcessor: () => DonutImageProcessor,
   EfficientNetImageProcessor: () => EfficientNetImageProcessor,
   GLPNFeatureExtractor: () => GLPNFeatureExtractor,
+  Gemma3ImageProcessor: () => Gemma3ImageProcessor,
   Glm46VImageProcessor: () => Glm46VImageProcessor,
   GroundingDinoImageProcessor: () => GroundingDinoImageProcessor,
   Idefics3ImageProcessor: () => Idefics3ImageProcessor,
@@ -15013,6 +15015,10 @@ var EfficientNetImageProcessor = class extends ImageProcessor {
   }
 };
 
+// src/models/gemma3/image_processing_gemma3.js
+var Gemma3ImageProcessor = class extends ImageProcessor {
+};
+
 // src/models/qwen2_vl/image_processing_qwen2_vl.js
 var Qwen2VLImageProcessor = class extends ImageProcessor {
   constructor(config) {
@@ -16236,6 +16242,48 @@ var Florence2Processor = class extends Processor {
   }
 };
 
+// src/models/gemma3/processing_gemma3.js
+var Gemma3Processor = class extends Processor {
+  static tokenizer_class = AutoTokenizer;
+  static image_processor_class = AutoImageProcessor;
+  static uses_processor_config = true;
+  static uses_chat_template_file = true;
+  constructor(config, components, chat_template) {
+    super(config, components, chat_template);
+    this.image_seq_length = this.config.image_seq_length;
+    const { boi_token, image_token, eoi_token } = this.tokenizer.config;
+    this.boi_token = boi_token;
+    this.image_token = image_token;
+    this.eoi_token = eoi_token;
+    const image_tokens_expanded = image_token.repeat(this.image_seq_length);
+    this.full_image_sequence = `
+
+${boi_token}${image_tokens_expanded}${eoi_token}
+
+`;
+  }
+  /**
+   * @param {string|string[]} text
+   * @param {import('../../utils/image.js').RawImage|import('../../utils/image.js').RawImage[]} [images]
+   * @param {Object} [options]
+   */
+  async _call(text, images = null, options = {}) {
+    if (typeof text === "string") {
+      text = [text];
+    }
+    let image_inputs;
+    if (images) {
+      image_inputs = await this.image_processor(images, options);
+      text = text.map((prompt) => prompt.replaceAll(this.boi_token, this.full_image_sequence));
+    }
+    const text_inputs = this.tokenizer(text, options);
+    return {
+      ...text_inputs,
+      ...image_inputs
+    };
+  }
+};
+
 // src/models/gemma3n/processing_gemma3n.js
 var Gemma3nProcessor = class extends Processor {
   static image_processor_class = AutoImageProcessor;
@@ -19416,8 +19464,7 @@ var MODEL_TYPES = {
   ImageAudioTextToText: 13,
   Supertonic: 14,
   Chatterbox: 15,
-  MultimodalLanguageModelOnly: 16,
-  VoxtralRealtime: 17
+  VoxtralRealtime: 16
 };
 var MODEL_TYPE_CONFIG = {
   [MODEL_TYPES.DecoderOnly]: {
@@ -19474,12 +19521,12 @@ var MODEL_TYPE_CONFIG = {
     can_generate: true,
     forward: image_text_to_text_forward,
     prepare_inputs: multimodal_text_to_text_prepare_inputs_for_generation,
-    sessions: (config) => {
+    sessions: (config, options, textOnly) => {
       const s = {
         embed_tokens: "embed_tokens",
-        vision_encoder: "vision_encoder",
         decoder_model_merged: "decoder_model_merged"
       };
+      if (!textOnly) s["vision_encoder"] = "vision_encoder";
       if (config.is_encoder_decoder) s["model"] = "encoder_model";
       return s;
     },
@@ -19501,12 +19548,17 @@ var MODEL_TYPE_CONFIG = {
   [MODEL_TYPES.ImageAudioTextToText]: {
     can_generate: true,
     prepare_inputs: multimodal_text_to_text_prepare_inputs_for_generation,
-    sessions: () => ({
-      embed_tokens: "embed_tokens",
-      audio_encoder: "audio_encoder",
-      vision_encoder: "vision_encoder",
-      decoder_model_merged: "decoder_model_merged"
-    }),
+    sessions: (config, options, textOnly) => {
+      const s = {
+        embed_tokens: "embed_tokens",
+        decoder_model_merged: "decoder_model_merged"
+      };
+      if (!textOnly) {
+        s["audio_encoder"] = "audio_encoder";
+        s["vision_encoder"] = "vision_encoder";
+      }
+      return s;
+    },
     optional_configs: { generation_config: "generation_config.json" }
   },
   [MODEL_TYPES.Phi3V]: {
@@ -19557,14 +19609,6 @@ var MODEL_TYPE_CONFIG = {
     cache_sessions: { model: true },
     optional_configs: { generation_config: "generation_config.json" }
   },
-  [MODEL_TYPES.MultimodalLanguageModelOnly]: {
-    can_generate: true,
-    forward: image_text_to_text_forward,
-    prepare_inputs: multimodal_text_to_text_prepare_inputs_for_generation,
-    sessions: () => ({ embed_tokens: "embed_tokens", decoder_model_merged: "decoder_model_merged" }),
-    cache_sessions: { decoder_model_merged: true },
-    optional_configs: { generation_config: "generation_config.json" }
-  },
   [MODEL_TYPES.VoxtralRealtime]: {
     can_generate: true,
     prepare_inputs: decoder_prepare_inputs_for_generation,
@@ -19590,6 +19634,19 @@ function getSessionsConfig(modelType, config, options = {}) {
     optional_configs: typeConfig.optional_configs
   };
 }
+function resolveTypeConfig(modelName, config) {
+  let modelType = MODEL_TYPE_MAPPING.get(modelName);
+  let textOnly = false;
+  const nativeArch = config?.architectures?.[0];
+  if (nativeArch && nativeArch !== modelName && modelName?.endsWith("ForCausalLM") && nativeArch.endsWith("ForConditionalGeneration")) {
+    const nativeType = MODEL_TYPE_MAPPING.get(nativeArch);
+    if (nativeType !== void 0) {
+      modelType = nativeType;
+      textOnly = true;
+    }
+  }
+  return { typeConfig: MODEL_TYPE_CONFIG[modelType] ?? MODEL_TYPE_CONFIG.default, textOnly, modelType };
+}
 var MODEL_TYPE_MAPPING = /* @__PURE__ */ new Map();
 var MODEL_NAME_TO_CLASS_MAPPING = /* @__PURE__ */ new Map();
 var MODEL_CLASS_TO_NAME_MAPPING = /* @__PURE__ */ new Map();
@@ -19609,8 +19666,7 @@ var PreTrainedModel = class extends Callable2 {
     this.sessions = sessions;
     this.configs = configs;
     const modelName = MODEL_CLASS_TO_NAME_MAPPING.get(this.constructor);
-    const modelType = MODEL_TYPE_MAPPING.get(modelName);
-    const typeConfig = MODEL_TYPE_CONFIG[modelType] ?? MODEL_TYPE_CONFIG.default;
+    const { typeConfig } = resolveTypeConfig(modelName, config);
     this.can_generate = typeConfig.can_generate;
     this._forward = typeConfig.forward;
     this._prepare_inputs_for_generation = typeConfig.prepare_inputs;
@@ -19673,9 +19729,8 @@ var PreTrainedModel = class extends Callable2 {
       session_options
     };
     const modelName = MODEL_CLASS_TO_NAME_MAPPING.get(this);
-    const modelType = MODEL_TYPE_MAPPING.get(modelName);
     config = options.config = await AutoConfig.from_pretrained(pretrained_model_name_or_path, options);
-    const typeConfig = MODEL_TYPE_CONFIG[modelType] ?? MODEL_TYPE_CONFIG.default;
+    const { typeConfig, textOnly, modelType } = resolveTypeConfig(modelName, config);
     if (modelType === void 0) {
       const type = modelName ?? config?.model_type;
       if (type !== "custom") {
@@ -19684,7 +19739,7 @@ var PreTrainedModel = class extends Callable2 {
         );
       }
     }
-    const sessions = typeConfig.sessions(config, options);
+    const sessions = typeConfig.sessions(config, options, textOnly);
     const promises = [
       constructSessions(pretrained_model_name_or_path, sessions, options, typeConfig.cache_sessions)
     ];
@@ -20739,6 +20794,7 @@ __export(models_exports, {
   Gemma2Model: () => Gemma2Model,
   Gemma2PreTrainedModel: () => Gemma2PreTrainedModel,
   Gemma3ForCausalLM: () => Gemma3ForCausalLM,
+  Gemma3ForConditionalGeneration: () => Gemma3ForConditionalGeneration,
   Gemma3Model: () => Gemma3Model,
   Gemma3PreTrainedModel: () => Gemma3PreTrainedModel,
   Gemma3nForCausalLM: () => Gemma3nForCausalLM,
@@ -22457,12 +22513,35 @@ var Gemma2Model = class extends Gemma2PreTrainedModel {
 var Gemma2ForCausalLM = class extends Gemma2PreTrainedModel {
 };
 
+// src/models/llava/modeling_llava.js
+var LlavaPreTrainedModel = class extends PreTrainedModel {
+  forward_params = ["input_ids", "attention_mask", "pixel_values", "position_ids", "past_key_values"];
+};
+var LlavaForConditionalGeneration = class extends LlavaPreTrainedModel {
+  _merge_input_ids_with_image_features(kwargs) {
+    const vision_hidden_size = kwargs.image_features.dims.at(-1);
+    const reshaped_image_hidden_states = kwargs.image_features.view(-1, vision_hidden_size);
+    return default_merge_input_ids_with_image_features({
+      // @ts-ignore
+      image_token_id: this.config.image_token_index ?? this.config.image_token_id,
+      ...kwargs,
+      image_features: reshaped_image_hidden_states
+    });
+  }
+};
+var Moondream1ForConditionalGeneration = class extends LlavaForConditionalGeneration {
+};
+var LlavaQwen2ForCausalLM = class extends LlavaForConditionalGeneration {
+};
+
 // src/models/gemma3/modeling_gemma3.js
 var Gemma3PreTrainedModel = class extends PreTrainedModel {
 };
 var Gemma3Model = class extends Gemma3PreTrainedModel {
 };
-var Gemma3ForCausalLM = class extends Gemma3PreTrainedModel {
+var Gemma3ForConditionalGeneration = class extends LlavaForConditionalGeneration {
+};
+var Gemma3ForCausalLM = class extends Gemma3ForConditionalGeneration {
 };
 
 // src/models/gemma3n/modeling_gemma3n.js
@@ -22828,40 +22907,45 @@ var Qwen2VLForConditionalGeneration = class extends Qwen2VLPreTrainedModel {
     });
   }
   prepare_inputs_for_generation(input_ids, model_inputs, generation_config) {
-    if (model_inputs.attention_mask && !model_inputs.position_ids) {
-      if (!model_inputs.past_key_values) {
-        [model_inputs.position_ids, model_inputs.rope_deltas] = this.get_rope_index(
+    if (!model_inputs.attention_mask || model_inputs.position_ids) {
+      return model_inputs;
+    }
+    const session = this.sessions["decoder_model_merged"] ?? this.sessions["model"];
+    if (!session.inputNames.includes("position_ids")) {
+      return model_inputs;
+    }
+    if (!model_inputs.past_key_values) {
+      [model_inputs.position_ids, model_inputs.rope_deltas] = this.get_rope_index(
+        model_inputs.input_ids,
+        model_inputs.image_grid_thw,
+        model_inputs.video_grid_thw,
+        model_inputs.attention_mask
+      );
+    } else {
+      model_inputs.pixel_values = null;
+      const past_length = model_inputs.past_key_values.get_seq_length();
+      if (past_length < model_inputs.input_ids.dims[1]) {
+        const [full_position_ids, rope_deltas] = this.get_rope_index(
           model_inputs.input_ids,
           model_inputs.image_grid_thw,
           model_inputs.video_grid_thw,
           model_inputs.attention_mask
         );
+        model_inputs.rope_deltas = rope_deltas;
+        model_inputs.position_ids = full_position_ids.slice(null, null, [past_length, null]);
+        model_inputs.input_ids = model_inputs.input_ids.slice(null, [past_length, null]);
       } else {
-        model_inputs.pixel_values = null;
-        const past_length = model_inputs.past_key_values.get_seq_length();
-        if (past_length < model_inputs.input_ids.dims[1]) {
-          const [full_position_ids, rope_deltas] = this.get_rope_index(
+        if (!model_inputs.rope_deltas) {
+          [, model_inputs.rope_deltas] = this.get_rope_index(
             model_inputs.input_ids,
             model_inputs.image_grid_thw,
             model_inputs.video_grid_thw,
             model_inputs.attention_mask
           );
-          model_inputs.rope_deltas = rope_deltas;
-          model_inputs.position_ids = full_position_ids.slice(null, null, [past_length, null]);
-          model_inputs.input_ids = model_inputs.input_ids.slice(null, [past_length, null]);
-        } else {
-          if (!model_inputs.rope_deltas) {
-            [, model_inputs.rope_deltas] = this.get_rope_index(
-              model_inputs.input_ids,
-              model_inputs.image_grid_thw,
-              model_inputs.video_grid_thw,
-              model_inputs.attention_mask
-            );
-          }
-          const delta = BigInt(past_length);
-          const rope_deltas_list = model_inputs.rope_deltas.map((x) => delta + x);
-          model_inputs.position_ids = stack([rope_deltas_list, rope_deltas_list, rope_deltas_list], 0);
         }
+        const delta = BigInt(past_length);
+        const rope_deltas_list = model_inputs.rope_deltas.map((x) => delta + x);
+        model_inputs.position_ids = stack([rope_deltas_list, rope_deltas_list, rope_deltas_list], 0);
       }
     }
     return model_inputs;
@@ -23144,27 +23228,6 @@ var HunYuanDenseV1Model = class extends HunYuanDenseV1PreTrainedModel {
 var HunYuanDenseV1ForCausalLM = class extends HunYuanDenseV1PreTrainedModel {
 };
 
-// src/models/llava/modeling_llava.js
-var LlavaPreTrainedModel = class extends PreTrainedModel {
-  forward_params = ["input_ids", "attention_mask", "pixel_values", "position_ids", "past_key_values"];
-};
-var LlavaForConditionalGeneration = class extends LlavaPreTrainedModel {
-  _merge_input_ids_with_image_features(kwargs) {
-    const vision_hidden_size = kwargs.image_features.dims.at(-1);
-    const reshaped_image_hidden_states = kwargs.image_features.view(-1, vision_hidden_size);
-    return default_merge_input_ids_with_image_features({
-      // @ts-ignore
-      image_token_id: this.config.image_token_index ?? this.config.image_token_id,
-      ...kwargs,
-      image_features: reshaped_image_hidden_states
-    });
-  }
-};
-var Moondream1ForConditionalGeneration = class extends LlavaForConditionalGeneration {
-};
-var LlavaQwen2ForCausalLM = class extends LlavaForConditionalGeneration {
-};
-
 // src/models/idefics3/modeling_idefics3.js
 var Idefics3ForConditionalGeneration = class extends LlavaForConditionalGeneration {
   forward_params = [
@@ -26062,6 +26125,7 @@ var MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = /* @__PURE__ */ new Map([
   ["qwen3_vl", "Qwen3VLForCausalLM"],
   ["qwen3_vl_moe", "Qwen3VLMoeForCausalLM"],
   ["qwen3_5", "Qwen3_5ForCausalLM"],
+  ["qwen3_5_text", "Qwen3_5ForCausalLM"],
   ["qwen3_5_moe", "Qwen3_5MoeForCausalLM"],
   ["gemma3n", "Gemma3nForCausalLM"],
   ["phi", "PhiForCausalLM"],
@@ -26149,6 +26213,7 @@ var MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES = /* @__PURE__ */ new Map([
   ["smolvlm", "SmolVLMForConditionalGeneration"],
   ["paligemma", "PaliGemmaForConditionalGeneration"],
   ["llava_qwen2", "LlavaQwen2ForCausalLM"],
+  ["gemma3", "Gemma3ForConditionalGeneration"],
   ["gemma3n", "Gemma3nForConditionalGeneration"],
   ["mistral3", "Mistral3ForConditionalGeneration"],
   ["lighton_ocr", "LightOnOcrForConditionalGeneration"],
@@ -26343,13 +26408,6 @@ var CUSTOM_MAPPING = [
   ],
   ["SupertonicForConditionalGeneration", SupertonicForConditionalGeneration, MODEL_TYPES.Supertonic],
   ["ChatterboxModel", ChatterboxModel, MODEL_TYPES.Chatterbox],
-  ["Qwen2VLForCausalLM", Qwen2VLForCausalLM, MODEL_TYPES.MultimodalLanguageModelOnly],
-  ["Qwen2_5_VLForCausalLM", Qwen2_5_VLForCausalLM, MODEL_TYPES.MultimodalLanguageModelOnly],
-  ["Qwen3VLForCausalLM", Qwen3VLForCausalLM, MODEL_TYPES.MultimodalLanguageModelOnly],
-  ["Qwen3VLMoeForCausalLM", Qwen3VLMoeForCausalLM, MODEL_TYPES.MultimodalLanguageModelOnly],
-  ["Qwen3_5ForCausalLM", Qwen3_5ForCausalLM, MODEL_TYPES.MultimodalLanguageModelOnly],
-  ["Qwen3_5MoeForCausalLM", Qwen3_5MoeForCausalLM, MODEL_TYPES.MultimodalLanguageModelOnly],
-  ["Gemma3nForCausalLM", Gemma3nForCausalLM, MODEL_TYPES.MultimodalLanguageModelOnly],
   [
     "VoxtralRealtimeForConditionalGeneration",
     VoxtralRealtimeForConditionalGeneration,
@@ -28031,6 +28089,41 @@ var TASK_ALIASES = Object.freeze({
   embeddings: "feature-extraction"
 });
 
+// src/utils/model_registry/resolve_model_type.js
+function resolve_model_type(config, { warn = true } = {}) {
+  const architectures = (
+    /** @type {string[]} */
+    config.architectures || []
+  );
+  for (const arch of architectures) {
+    const mappedType = MODEL_TYPE_MAPPING.get(arch);
+    if (mappedType !== void 0) {
+      return mappedType;
+    }
+  }
+  if (config.model_type) {
+    const mappedType = MODEL_TYPE_MAPPING.get(config.model_type);
+    if (mappedType !== void 0) {
+      return mappedType;
+    }
+    for (const mapping of Object.values(MODEL_MAPPING_NAMES)) {
+      if (mapping.has(config.model_type)) {
+        const resolved = MODEL_TYPE_MAPPING.get(mapping.get(config.model_type));
+        if (resolved !== void 0) {
+          return resolved;
+        }
+      }
+    }
+  }
+  if (warn) {
+    const archList = architectures.length > 0 ? architectures.join(", ") : "(none)";
+    logger.warn(
+      `[resolve_model_type] Architecture(s) not found in MODEL_TYPE_MAPPING: [${archList}] for model type '${config.model_type}'. Falling back to EncoderOnly (single model.onnx file). If you encounter issues, please report at: ${GITHUB_ISSUE_URL}`
+    );
+  }
+  return MODEL_TYPES.EncoderOnly;
+}
+
 // src/utils/model_registry/get_model_files.js
 function get_config(modelId, { config = null, cache_dir = null, local_files_only = false, revision = "main" } = {}) {
   if (config !== null) {
@@ -28053,43 +28146,7 @@ async function get_model_files(modelId, { config = null, dtype: overrideDtype =
   const subfolder = "onnx";
   const rawDevice = overrideDevice ?? custom_config.device;
   let dtype = overrideDtype ?? custom_config.dtype;
-  let modelType;
-  const architectures = (
-    /** @type {string[]} */
-    config.architectures || []
-  );
-  let foundInMapping = false;
-  for (const arch of architectures) {
-    const mappedType = MODEL_TYPE_MAPPING.get(arch);
-    if (mappedType !== void 0) {
-      modelType = mappedType;
-      foundInMapping = true;
-      break;
-    }
-  }
-  if (!foundInMapping && config.model_type) {
-    const mappedType = MODEL_TYPE_MAPPING.get(config.model_type);
-    if (mappedType !== void 0) {
-      modelType = mappedType;
-      foundInMapping = true;
-    }
-    if (!foundInMapping) {
-      for (const mapping of Object.values(MODEL_MAPPING_NAMES)) {
-        if (mapping.has(config.model_type)) {
-          modelType = MODEL_TYPE_MAPPING.get(mapping.get(config.model_type));
-          foundInMapping = true;
-          break;
-        }
-      }
-    }
-  }
-  if (!foundInMapping) {
-    const archList = architectures.length > 0 ? architectures.join(", ") : "(none)";
-    logger.warn(
-      `[get_model_files] Architecture(s) not found in MODEL_TYPE_MAPPING: [${archList}] for model type '${config.model_type}'. Falling back to EncoderOnly (single model.onnx file). If you encounter issues, please report at: ${GITHUB_ISSUE_URL}`
-    );
-    modelType = MODEL_TYPES.EncoderOnly;
-  }
+  const modelType = resolve_model_type(config);
   const add_model_file = (fileName, baseName = null) => {
     baseName = baseName ?? fileName;
     const selectedDevice = selectDevice(rawDevice, fileName);
@@ -28676,6 +28733,31 @@ async function clear_pipeline_cache(task, modelId, options = {}) {
   return await clear_files_from_cache(modelId, files, options);
 }
 
+// src/utils/model_registry/get_available_dtypes.js
+var CONCRETE_DTYPES = Object.keys(DEFAULT_DTYPE_SUFFIX_MAPPING);
+async function get_available_dtypes(modelId, { config = null, model_file_name = null, revision = "main", cache_dir = null, local_files_only = false } = {}) {
+  config = await get_config(modelId, { config, cache_dir, local_files_only, revision });
+  const subfolder = "onnx";
+  const modelType = resolve_model_type(config);
+  const { sessions } = getSessionsConfig(modelType, config, { model_file_name });
+  const baseNames = Object.values(sessions);
+  const metadataOptions = { revision, cache_dir, local_files_only };
+  const probeResults = await Promise.all(
+    CONCRETE_DTYPES.map(async (dtype) => {
+      const suffix = DEFAULT_DTYPE_SUFFIX_MAPPING[dtype] ?? "";
+      const allExist = await Promise.all(
+        baseNames.map(async (baseName) => {
+          const filename = `${subfolder}/${baseName}${suffix}.onnx`;
+          const metadata = await get_file_metadata(modelId, filename, metadataOptions);
+          return metadata.exists;
+        })
+      );
+      return { dtype, available: allExist.every(Boolean) };
+    })
+  );
+  return probeResults.filter((r) => r.available).map((r) => r.dtype);
+}
+
 // src/utils/model_registry/ModelRegistry.js
 var ModelRegistry = class {
   /**
@@ -28762,6 +28844,29 @@ var ModelRegistry = class {
   static async get_processor_files(modelId) {
     return get_processor_files(modelId);
   }
+  /**
+   * Detects which quantization levels (dtypes) are available for a model
+   * by checking which ONNX files exist on the hub or locally.
+   *
+   * A dtype is considered available if all required model session files
+   * exist for that dtype.
+   *
+   * @param {string} modelId - The model id (e.g., "onnx-community/all-MiniLM-L6-v2-ONNX")
+   * @param {Object} [options] - Optional parameters
+   * @param {import('../../configs.js').PretrainedConfig} [options.config=null] - Pre-loaded config
+   * @param {string} [options.model_file_name=null] - Override the model file name (excluding .onnx suffix)
+   * @param {string} [options.revision='main'] - Model revision
+   * @param {string} [options.cache_dir=null] - Custom cache directory
+   * @param {boolean} [options.local_files_only=false] - Only check local files
+   * @returns {Promise<string[]>} Array of available dtype strings (e.g., ['fp32', 'fp16', 'q4', 'q8'])
+   *
+   * @example
+   * const dtypes = await ModelRegistry.get_available_dtypes('onnx-community/all-MiniLM-L6-v2-ONNX');
+   * console.log(dtypes); // ['fp32', 'fp16', 'int8', 'uint8', 'q8', 'q4']
+   */
+  static async get_available_dtypes(modelId, options = {}) {
+    return get_available_dtypes(modelId, options);
+  }
   /**
    * Quickly checks if a model is fully cached by verifying `config.json` is present,
    * then confirming all required files are cached.
@@ -29208,8 +29313,11 @@ export {
   Gemma2Model,
   Gemma2PreTrainedModel,
   Gemma3ForCausalLM,
+  Gemma3ForConditionalGeneration,
+  Gemma3ImageProcessor,
   Gemma3Model,
   Gemma3PreTrainedModel,
+  Gemma3Processor,
   Gemma3nAudioFeatureExtractor,
   Gemma3nForCausalLM,
   Gemma3nForConditionalGeneration,