ONNX Runtime improvements

package.json

@@ -27,7 +27,7 @@
     "typegen": "tsc --build",
     "dev": "webpack serve --no-client-overlay",
     "build": "webpack && npm run typegen",
-    "test": "node --experimental-vm-modules --expose-gc node_modules/jest/bin/jest.js --verbose",
+    "test": "node --experimental-vm-modules --expose-gc node_modules/jest/bin/jest.js --verbose --logHeapUsage",
     "readme": "python ./docs/scripts/build_readme.py",
     "docs-api": "node ./docs/scripts/generate.js",
     "docs-preview": "doc-builder preview transformers.js ./docs/source/ --not_python_module",
@@ -55,10 +55,10 @@
   },
   "homepage": "https://github.com/huggingface/transformers.js#readme",
   "dependencies": {
-    "@huggingface/jinja": "^0.4.1",
-    "onnxruntime-node": "1.21.0",
-    "onnxruntime-web": "1.22.0-dev.20250409-89f8206ba4",
-    "sharp": "^0.34.1"
+    "@huggingface/jinja": "^0.5.0",
+    "onnxruntime-node": "1.23.0-dev.20250612-70f14d7670",
+    "onnxruntime-web": "1.23.0-dev.20250612-70f14d7670",
+    "sharp": "^0.34.2"
   },
   "devDependencies": {
     "@types/jest": "^29.5.14",

scripts/quantize.py

@@ -12,7 +12,7 @@
 from onnxruntime.quantization import QuantType, QuantizationMode
 from onnxruntime.quantization.onnx_quantizer import ONNXQuantizer
 from onnxruntime.quantization.registry import IntegerOpsRegistry
-from onnxruntime.quantization.matmul_4bits_quantizer import MatMul4BitsQuantizer
+from onnxruntime.quantization.matmul_nbits_quantizer import MatMulNBitsQuantizer
 from onnxruntime.quantization.matmul_bnb4_quantizer import MatMulBnb4Quantizer
 
 from . import float16
@@ -107,7 +107,7 @@ class QuantizationArguments:
         },
     )
 
-    # MatMul4BitsQuantizer
+    # MatMulNBitsQuantizer
     is_symmetric: bool = field(
         default=True,
         metadata={"help": "Indicate whether to quantize the model symmetrically"},
@@ -234,7 +234,7 @@ def quantize_q4(
     Quantize the weights of the model from float32 to 4-bit int
     """
 
-    quantizer = MatMul4BitsQuantizer(
+    quantizer = MatMulNBitsQuantizer(
         model=model,
         block_size=block_size,
         is_symmetric=is_symmetric,

scripts/requirements.txt

@@ -1,6 +1,6 @@
-transformers[torch]==4.49.0
-onnxruntime==1.20.1
-optimum@git+https://github.com/huggingface/optimum.git@b04feaea78cda58d79b8da67dca3fd0c4ab33435
-onnx==1.17.0
+transformers[torch]==4.52.4
+onnxruntime==1.22.0
+optimum==1.26.1
+onnx==1.18.0
 tqdm==4.67.1
-onnxslim==0.1.48
+onnxslim==0.1.52

src/backends/onnx.js

@@ -38,6 +38,7 @@ const DEVICE_TO_EXECUTION_PROVIDER_MAPPING = Object.freeze({
     webgpu: 'webgpu', // WebGPU
     cuda: 'cuda', // CUDA
     dml: 'dml', // DirectML
+    coreml: 'coreml', // CoreML
 
     webnn: { name: 'webnn', deviceType: 'cpu' }, // WebNN (default)
     'webnn-npu': { name: 'webnn', deviceType: 'npu' }, // WebNN NPU
@@ -63,13 +64,15 @@ if (ORT_SYMBOL in globalThis) {
 } else if (apis.IS_NODE_ENV) {
     ONNX = ONNX_NODE.default ?? ONNX_NODE;
 
-    // Updated as of ONNX Runtime 1.20.1
+    // Updated as of ONNX Runtime 1.23.0-dev.20250612-70f14d7670
     // The following table lists the supported versions of ONNX Runtime Node.js binding provided with pre-built binaries.
-    // | EPs/Platforms | Windows x64 | Windows arm64 | Linux x64         | Linux arm64 | MacOS x64 | MacOS arm64 |
-    // | ------------- | ----------- | ------------- | ----------------- | ----------- | --------- | ----------- |
-    // | CPU           | ✔️          | ✔️            | ✔️                | ✔️          | ✔️        | ✔️          |
-    // | DirectML      | ✔️          | ✔️            | ❌                | ❌          | ❌        | ❌          |
-    // | CUDA          | ❌          | ❌            | ✔️ (CUDA v11.8)   | ❌          | ❌        | ❌          |
+    // | EPs/Platforms         | Windows x64        | Windows arm64      | Linux x64          | Linux arm64        | MacOS x64          | MacOS arm64        |
+    // | --------------------- | ------------------ | ------------------ | ------------------ | ------------------ | ------------------ | ------------------ |
+    // | CPU                   | ✔️                  | ✔️                  | ✔️                  | ✔️                  | ✔️                  | ✔️                  |
+    // | WebGPU (experimental) | ✔️                  | ✔️                  | ✔️                  | ❌                  | ✔️                  | ✔️                  |
+    // | DirectML              | ✔️                  | ✔️                  | ❌                  | ❌                  | ❌                  | ❌                  |
+    // | CUDA                  | ❌                  | ❌                  | ✔️ (CUDA v12)       | ❌                  | ❌                  | ❌                  |
+    // | CoreML                | ❌                  | ❌                  | ❌                  | ❌                  | ✔️                  | ✔️                  |
     switch (process.platform) {
         case 'win32': // Windows x64 and Windows arm64
             supportedDevices.push('dml');
@@ -80,9 +83,11 @@ if (ORT_SYMBOL in globalThis) {
             }
             break;
         case 'darwin': // MacOS x64 and MacOS arm64
+            supportedDevices.push('coreml');
             break;
     }
 
+    supportedDevices.push('webgpu');
     supportedDevices.push('cpu');
     defaultDevices = ['cpu'];
 } else {
@@ -180,9 +185,16 @@ if (ONNX_ENV?.wasm) {
     if (
         // @ts-ignore Cannot find name 'ServiceWorkerGlobalScope'.ts(2304)
         !(typeof ServiceWorkerGlobalScope !== 'undefined' && self instanceof ServiceWorkerGlobalScope)
+        && env.backends.onnx.versions?.web
         && !ONNX_ENV.wasm.wasmPaths
     ) {
-        ONNX_ENV.wasm.wasmPaths = `https://cdn.jsdelivr.net/npm/@huggingface/transformers@${env.version}/dist/`;
+        const wasmPathPrefix = `https://cdn.jsdelivr.net/npm/onnxruntime-web@${env.backends.onnx.versions.web}/dist/`;
+
+        ONNX_ENV.wasm.wasmPaths = apis.IS_SAFARI ? {
+            "mjs": `${wasmPathPrefix}/ort-wasm-simd-threaded.mjs`,
+            "wasm": `${wasmPathPrefix}/ort-wasm-simd-threaded.wasm`,
+        }
+            : wasmPathPrefix;
     }
 
     // TODO: Add support for loading WASM files from cached buffer when we upgrade to [email protected]

src/env.js

@@ -28,13 +28,6 @@ import url from 'node:url';
 
 const VERSION = '3.5.2';
 
-// 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"  && self.constructor?.name === 'DedicatedWorkerGlobalScope';
-const IS_WEB_CACHE_AVAILABLE = typeof self !== "undefined" && 'caches' in self;
-const IS_WEBGPU_AVAILABLE = typeof navigator !== 'undefined' && 'gpu' in navigator;
-const IS_WEBNN_AVAILABLE = typeof navigator !== 'undefined' && 'ml' in navigator;
-
 const IS_PROCESS_AVAILABLE = typeof process !== 'undefined';
 const IS_NODE_ENV = IS_PROCESS_AVAILABLE && process?.release?.name === 'node';
 const IS_FS_AVAILABLE = !isEmpty(fs);
@@ -44,6 +37,41 @@ const IS_PATH_AVAILABLE = !isEmpty(path);
 const IS_DENO_RUNTIME = typeof globalThis.Deno !== 'undefined';
 const IS_BUN_RUNTIME = typeof globalThis.Bun !== 'undefined';
 
+// 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" && self.constructor?.name === 'DedicatedWorkerGlobalScope';
+const IS_WEB_CACHE_AVAILABLE = typeof self !== "undefined" && 'caches' in self;
+const IS_WEBGPU_AVAILABLE = IS_NODE_ENV || (typeof navigator !== 'undefined' && 'gpu' in navigator);
+const IS_WEBNN_AVAILABLE = typeof navigator !== 'undefined' && 'ml' in navigator;
+
+/**
+ * Check if the current environment is Safari browser.
+ * Works in both browser and web worker contexts.
+ * @returns {boolean} Whether the current environment is Safari.
+ */
+const isSafari = () => {
+    // Check if we're in a browser environment
+    if (typeof navigator === 'undefined') {
+        return false;
+    }
+
+    const userAgent = navigator.userAgent;
+    const vendor = navigator.vendor || '';
+
+    // Safari has "Apple" in vendor string
+    const isAppleVendor = vendor.indexOf('Apple') > -1;
+
+    // Exclude Chrome on iOS (CriOS), Firefox on iOS (FxiOS), 
+    // Edge on iOS (EdgiOS), and other browsers
+    const notOtherBrowser =
+        !userAgent.match(/CriOS|FxiOS|EdgiOS|OPiOS|mercury|brave/i) &&
+        !userAgent.includes('Chrome') &&
+        !userAgent.includes('Android');
+
+    return isAppleVendor && notOtherBrowser;
+};
+const IS_SAFARI = isSafari();
+
 /**
  * A read-only object containing information about the APIs available in the current environment.
  */
@@ -63,6 +91,9 @@ export const apis = Object.freeze({
     /** Whether the WebNN API is available */
     IS_WEBNN_AVAILABLE,
 
+    /** Whether we are running in a Safari browser */
+    IS_SAFARI,
+
     /** Whether the Node.js process API is available */
     IS_PROCESS_AVAILABLE,
 

src/models.js

@@ -154,10 +154,11 @@ const MODEL_CLASS_TO_NAME_MAPPING = new Map();
  * @param {string} pretrained_model_name_or_path The path to the directory containing the model file.
  * @param {string} fileName The name of the model file.
  * @param {import('./utils/hub.js').PretrainedModelOptions} options Additional options for loading the model.
+ * @param {boolean} [is_decoder=false] Whether the model is a decoder model.
  * @returns {Promise<{buffer_or_path: Uint8Array|string, session_options: Object, session_config: Object}>} A Promise that resolves to the data needed to create an InferenceSession object.
  * @private
  */
-async function getSession(pretrained_model_name_or_path, fileName, options) {
+async function getSession(pretrained_model_name_or_path, fileName, options, is_decoder = false) {
     let custom_config = options.config?.['transformers.js_config'] ?? {};
 
     let device = options.device ?? custom_config.device;
@@ -218,7 +219,14 @@ async function getSession(pretrained_model_name_or_path, fileName, options) {
 
     if (!DEFAULT_DTYPE_SUFFIX_MAPPING.hasOwnProperty(selectedDtype)) {
         throw new Error(`Invalid dtype: ${selectedDtype}. Should be one of: ${Object.keys(DATA_TYPES).join(', ')}`);
-    } else if (selectedDtype === DATA_TYPES.fp16 && selectedDevice === 'webgpu' && !(await isWebGpuFp16Supported())) {
+    } else if (
+        selectedDevice === 'webgpu' && (
+            // NOTE: Currently, we assume that the Native WebGPU EP always supports fp16. In future, we will add a check for this.
+            !apis.IS_NODE_ENV
+            &&
+            (selectedDtype === DATA_TYPES.fp16 && !(await isWebGpuFp16Supported()))
+        )
+    ) {
         throw new Error(`The device (${selectedDevice}) does not support fp16.`);
     }
 
@@ -316,7 +324,7 @@ async function getSession(pretrained_model_name_or_path, fileName, options) {
         }
     }
 
-    if (selectedDevice === 'webgpu') {
+    if (is_decoder && selectedDevice === 'webgpu') {
         const shapes = getKeyValueShapes(options.config, {
             prefix: 'present',
         });
@@ -342,13 +350,14 @@ async function getSession(pretrained_model_name_or_path, fileName, options) {
  * @param {string} pretrained_model_name_or_path The path to the directory containing the model file.
  * @param {Record<string, string>} names The names of the model files to load.
  * @param {import('./utils/hub.js').PretrainedModelOptions} options Additional options for loading the model.
+ * @param {string} [decoder_name] The name of the decoder model, if any.
  * @returns {Promise<Record<string, any>>} A Promise that resolves to a dictionary of InferenceSession objects.
  * @private
  */
-async function constructSessions(pretrained_model_name_or_path, names, options) {
+async function constructSessions(pretrained_model_name_or_path, names, options, decoder_name = undefined) {
     return Object.fromEntries(await Promise.all(
         Object.keys(names).map(async (name) => {
-            const { buffer_or_path, session_options, session_config } = await getSession(pretrained_model_name_or_path, names[name], options);
+            const { buffer_or_path, session_options, session_config } = await getSession(pretrained_model_name_or_path, names[name], options, name === decoder_name);
             const session = await createInferenceSession(buffer_or_path, session_options, session_config);
             return [name, session];
         })
@@ -1148,7 +1157,7 @@ export class PreTrainedModel extends Callable {
             info = await Promise.all([
                 constructSessions(pretrained_model_name_or_path, {
                     model: options.model_file_name ?? 'model',
-                }, options),
+                }, options, 'model'),
                 getOptionalConfigs(pretrained_model_name_or_path, {
                     generation_config: 'generation_config.json',
                 }, options),
@@ -1159,7 +1168,7 @@ export class PreTrainedModel extends Callable {
                 constructSessions(pretrained_model_name_or_path, {
                     model: 'encoder_model',
                     decoder_model_merged: 'decoder_model_merged',
-                }, options),
+                }, options, 'decoder_model_merged'),
                 getOptionalConfigs(pretrained_model_name_or_path, {
                     generation_config: 'generation_config.json',
                 }, options),
@@ -1178,7 +1187,7 @@ export class PreTrainedModel extends Callable {
                 constructSessions(pretrained_model_name_or_path, {
                     model: 'encoder_model',
                     decoder_model_merged: 'decoder_model_merged',
-                }, options),
+                }, options, 'decoder_model_merged'),
             ]);
 
         } else if (modelType === MODEL_TYPES.ImageTextToText) {
@@ -1191,7 +1200,7 @@ export class PreTrainedModel extends Callable {
                 sessions['model'] = 'encoder_model';
             }
             info = await Promise.all([
-                constructSessions(pretrained_model_name_or_path, sessions, options),
+                constructSessions(pretrained_model_name_or_path, sessions, options, 'decoder_model_merged'),
                 getOptionalConfigs(pretrained_model_name_or_path, {
                     generation_config: 'generation_config.json',
                 }, options),
@@ -1204,7 +1213,7 @@ export class PreTrainedModel extends Callable {
                 decoder_model_merged: 'decoder_model_merged',
             }
             info = await Promise.all([
-                constructSessions(pretrained_model_name_or_path, sessions, options),
+                constructSessions(pretrained_model_name_or_path, sessions, options, 'decoder_model_merged'),
                 getOptionalConfigs(pretrained_model_name_or_path, {
                     generation_config: 'generation_config.json',
                 }, options),
@@ -1216,7 +1225,7 @@ export class PreTrainedModel extends Callable {
                     model: 'text_encoder',
                     decoder_model_merged: 'decoder_model_merged',
                     encodec_decode: 'encodec_decode',
-                }, options),
+                }, options, 'decoder_model_merged'),
                 getOptionalConfigs(pretrained_model_name_or_path, {
                     generation_config: 'generation_config.json',
                 }, options),
@@ -1231,7 +1240,7 @@ export class PreTrainedModel extends Callable {
                     gen_head: 'gen_head',
                     gen_img_embeds: 'gen_img_embeds',
                     image_decode: 'image_decode',
-                }, options),
+                }, options, 'model'),
                 getOptionalConfigs(pretrained_model_name_or_path, {
                     generation_config: 'generation_config.json',
                 }, options),
@@ -1243,7 +1252,7 @@ export class PreTrainedModel extends Callable {
                     prepare_inputs_embeds: 'prepare_inputs_embeds',
                     model: 'model',
                     vision_encoder: 'vision_encoder',
-                }, options),
+                }, options, 'model'),
                 getOptionalConfigs(pretrained_model_name_or_path, {
                     generation_config: 'generation_config.json',
                 }, options),

src/utils/devices.js

@@ -10,6 +10,7 @@ export const DEVICE_TYPES = Object.freeze({
     webgpu: 'webgpu', // WebGPU
     cuda: 'cuda', // CUDA
     dml: 'dml', // DirectML
+    coreml: 'coreml', // CoreML
 
     webnn: 'webnn', // WebNN (default)
     'webnn-npu': 'webnn-npu', // WebNN NPU

tests/models/florence2/test_modeling_florence2.js

@@ -46,7 +46,7 @@ export default () => {
         {
           const inputs = await processor(image, texts[0]);
           const generate_ids = await model.generate({ ...inputs, max_new_tokens: 10 });
-          expect(generate_ids.tolist()).toEqual([[2n, 0n, 48n, 48n, 48n, 48n, 48n, 48n, 48n, 48n, 2n]]);
+          expect(generate_ids.tolist()).toEqual([[2n, 0n, 0n, 0n, 1n, 0n, 0n, 2n]]);
         }
       },
       MAX_TEST_EXECUTION_TIME,
@@ -68,8 +68,8 @@ export default () => {
 
           const generate_ids = await model.generate({ ...inputs, max_new_tokens: 10 });
           expect(generate_ids.tolist()).toEqual([
-            [2n, 0n, 48n, 48n, 48n, 48n, 48n, 48n, 48n, 48n, 2n],
-            [2n, 0n, 48n, 48n, 48n, 48n, 48n, 48n, 48n, 48n, 2n],
+            [2n, 0n, 0n, 0n, 1n, 0n, 0n, 2n],
+            [2n, 0n, 0n, 0n, 1n, 0n, 0n, 2n],
           ]);
         }
       },

tests/models/grounding_dino/test_modeling_grounding_dino.js

@@ -32,7 +32,7 @@ export default () => {
         expect(pred_boxes.dims).toEqual([1, num_queries, 4]);
         expect(logits.max().item()).toBeCloseTo(56.237613677978516, 2);
         expect(logits.min().item()).toEqual(-Infinity);
-        expect(pred_boxes.mean().item()).toEqual(0.2500016987323761);
+        expect(pred_boxes.mean().item()).toBeCloseTo(0.2500016987323761, 4);
       },
       MAX_TEST_EXECUTION_TIME,
     );

tests/pipelines/test_pipelines_depth_estimation.js

@@ -26,7 +26,7 @@ export default () => {
         async () => {
           const output = await pipe(images[0]);
           expect(output.predicted_depth.dims).toEqual([224, 224]);
-          expect(output.predicted_depth.mean().item()).toBeCloseTo(0.000006106501587055391, 6);
+          expect(output.predicted_depth.mean().item()).toBeCloseTo(0.000006106501587055391, 4);
           expect(output.depth.size).toEqual(images[0].size);
         },
         MAX_TEST_EXECUTION_TIME,
@@ -40,10 +40,10 @@ export default () => {
           const output = await pipe(images);
           expect(output).toHaveLength(images.length);
           expect(output[0].predicted_depth.dims).toEqual([224, 224]);
-          expect(output[0].predicted_depth.mean().item()).toBeCloseTo(0.000006106501587055391, 6);
+          expect(output[0].predicted_depth.mean().item()).toBeCloseTo(0.000006106501587055391, 4);
           expect(output[0].depth.size).toEqual(images[0].size);
           expect(output[1].predicted_depth.dims).toEqual([224, 224]);
-          expect(output[1].predicted_depth.mean().item()).toBeCloseTo(0.0000014548650142387487, 6);
+          expect(output[1].predicted_depth.mean().item()).toBeCloseTo(0.0000014548650142387487, 4);
           expect(output[1].depth.size).toEqual(images[1].size);
         },
         MAX_TEST_EXECUTION_TIME,