fixes for ort-1.17

package.json

@@ -38,12 +38,12 @@
   },
   "homepage": "https://github.com/xenova/transformers.js#readme",
   "dependencies": {
-    "onnxruntime-web": "1.14.0",
+    "onnxruntime-web": "1.17.1",
     "sharp": "^0.32.0",
     "@huggingface/jinja": "^0.2.1"
   },
   "optionalDependencies": {
-    "onnxruntime-node": "1.14.0"
+    "onnxruntime-node": "1.17.1"
   },
   "devDependencies": {
     "@types/jest": "^29.5.1",

src/models.js

@@ -203,9 +203,16 @@ function validateInputs(session, inputs) {
 async function sessionRun(session, inputs) {
     const checkedInputs = validateInputs(session, inputs);
     try {
-        // @ts-ignore
-        let output = await session.run(checkedInputs);
+        // pass the original ort tensor
+        const ortFeed = Object.fromEntries(Object.entries(checkedInputs).map(([k, v]) => [k, v.ort_tensor]));
+        let output = await session.run(ortFeed);    
         output = replaceTensors(output);
+        for (const [name, t] of Object.entries(checkedInputs)) {
+            // if we use gpu buffers for kv_caches, we own them and need to dispose()
+            if (name.startsWith('past_key_values')) {
+                t.dispose();
+            };
+        }
         return output;
     } catch (e) {
         // This usually occurs when the inputs are of the wrong type.

src/utils/tensor.js

@@ -17,6 +17,7 @@ import {
 
 const DataTypeMap = Object.freeze({
     float32: Float32Array,
+    float16: Uint16Array,
     float64: Float64Array,
     string: Array, // string[]
     int8: Int8Array,
@@ -39,33 +40,48 @@ const ONNXTensor = ONNX.Tensor;
 
 export class Tensor {
     /** @type {number[]} Dimensions of the tensor. */
-    dims;
+    get dims() {
+        // @ts-ignore
+        return this.ort_tensor.dims;
+    }
+    set dims(value) {
+        // FIXME: ONNXTensor declares dims as readonly so one needs to use the constructor() if dims change.
+        // @ts-ignore
+        this.ort_tensor.dims = value;
+    }
 
     /** @type {DataType} Type of the tensor. */
-    type;
+    get type() {
+        return this.ort_tensor.type;
+    };
 
     /** @type {DataArray} The data stored in the tensor. */
-    data;
+    get data() {
+        return this.ort_tensor.data;
+    }
 
     /** @type {number} The number of elements in the tensor. */
-    size;
+    get size() {
+        return this.ort_tensor.size;
+    };
+
+    ort_tensor;
 
     /**
      * Create a new Tensor or copy an existing Tensor.
      * @param {[DataType, DataArray, number[]]|[import('onnxruntime-common').Tensor]} args
      */
     constructor(...args) {
         if (args[0] instanceof ONNXTensor) {
-            // Create shallow copy
-            Object.assign(this, args[0]);
-
+            this.ort_tensor = args[0];
         } else {
             // Create new tensor
-            Object.assign(this, new ONNXTensor(
+            const t = new ONNXTensor(
                 /** @type {DataType} */(args[0]),
                 /** @type {Exclude<import('./maths.js').AnyTypedArray, Uint8ClampedArray>} */(args[1]),
                 args[2]
-            ));
+            );
+            this.ort_tensor = t;
         }
 
         return new Proxy(this, {
@@ -89,6 +105,11 @@ export class Tensor {
         });
     }
 
+    dispose() {
+        this.ort_tensor.dispose();
+        // this.ort_tensor = undefined;
+    }
+
     /**
      * Returns an iterator object for iterating over the tensor data in row-major order.
      * If the tensor has more than one dimension, the iterator will yield subarrays.