catniff 0.4.0 → 0.4.2

package/dist/backend.d.ts

@@ -1,83 +1,4 @@
-import { Tensor, TensorValue } from "./core";
+import { Tensor } from "./core";
 export interface Backend {
-    sum?(tensor: Tensor, dims?: number[] | number, keepDims?: boolean): Tensor;
-    prod?(tensor: Tensor, dims?: number[] | number, keepDims?: boolean): Tensor;
-    mean?(tensor: Tensor, dims?: number[] | number, keepDims?: boolean): Tensor;
-    max?(tensor: Tensor, dims?: number[] | number, keepDims?: boolean): Tensor;
-    min?(tensor: Tensor, dims?: number[] | number, keepDims?: boolean): Tensor;
-    softmax?(tensor: Tensor, dims?: number[] | number): Tensor;
-    add?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    sub?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    mul?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    pow?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    div?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    remainder?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    ge?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    le?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    gt?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    lt?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    eq?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    ne?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    logicalAnd?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    logicalOr?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    logicalXor?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    logicalNot?(self: Tensor): Tensor;
-    bitwiseAnd?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    bitwiseOr?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    bitwiseXor?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    bitwiseNot?(self: Tensor): Tensor;
-    bitwiseLeftShift?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    bitwiseRightShift?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    neg?(self: Tensor): Tensor;
-    reciprocal?(self: Tensor): Tensor;
-    square?(self: Tensor): Tensor;
-    abs?(self: Tensor): Tensor;
-    sign?(self: Tensor): Tensor;
-    sin?(self: Tensor): Tensor;
-    cos?(self: Tensor): Tensor;
-    tan?(self: Tensor): Tensor;
-    asin?(self: Tensor): Tensor;
-    acos?(self: Tensor): Tensor;
-    atan?(self: Tensor): Tensor;
-    atan2?(self: Tensor): Tensor;
-    sinh?(self: Tensor): Tensor;
-    cosh?(self: Tensor): Tensor;
-    asinh?(self: Tensor): Tensor;
-    acosh?(self: Tensor): Tensor;
-    atanh?(self: Tensor): Tensor;
-    deg2rad?(self: Tensor): Tensor;
-    rad2deg?(self: Tensor): Tensor;
-    sqrt?(self: Tensor): Tensor;
-    rsqrt?(self: Tensor): Tensor;
-    exp?(self: Tensor): Tensor;
-    exp2?(self: Tensor): Tensor;
-    expm1?(self: Tensor): Tensor;
-    log?(self: Tensor): Tensor;
-    log2?(self: Tensor): Tensor;
-    log10?(self: Tensor): Tensor;
-    log1p?(self: Tensor): Tensor;
-    relu?(self: Tensor): Tensor;
-    sigmoid?(self: Tensor): Tensor;
-    tanh?(self: Tensor): Tensor;
-    softplus?(self: Tensor): Tensor;
-    softsign?(self: Tensor): Tensor;
-    silu?(self: Tensor): Tensor;
-    mish?(self: Tensor): Tensor;
-    maximum?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    minimum?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    round?(self: Tensor): Tensor;
-    floor?(self: Tensor): Tensor;
-    ceil?(self: Tensor): Tensor;
-    trunc?(self: Tensor): Tensor;
-    frac?(self: Tensor): Tensor;
-    clip?(self: Tensor, min: number, max: number): Tensor;
-    erf?(self: Tensor): Tensor;
-    erfc?(self: Tensor): Tensor;
-    erfinv?(self: Tensor): Tensor;
-    dot?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    mm?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    bmm?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    mv?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    matmul?(self: Tensor, other: Tensor | TensorValue): Tensor;
-    to?(tensor: Tensor): Tensor;
+    transfer(tensor: Tensor): Tensor;
 }

package/dist/core.d.ts

@@ -46,6 +46,8 @@ export declare class Tensor {
     mean(dims?: number[] | number, keepDims?: boolean): Tensor;
     max(dims?: number[] | number, keepDims?: boolean): Tensor;
     min(dims?: number[] | number, keepDims?: boolean): Tensor;
+    var(dims?: number[] | number, keepDims?: boolean): Tensor;
+    std(dims?: number[] | number, keepDims?: boolean): Tensor;
     softmax(dims?: number[] | number): Tensor;
     add(other: TensorValue | Tensor): Tensor;
     sub(other: TensorValue | Tensor): Tensor;

package/dist/core.js

@@ -337,21 +337,22 @@ class Tensor {
     }
     // Tensor sum reduction
     sum(dims, keepDims = false) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.sum) {
-            return backend.sum(this, dims, keepDims);
-        }
         if (typeof this.value === "number")
             return this;
-        if (typeof dims === "number") {
-            dims = [dims];
-        }
         if (typeof dims === "undefined") {
             dims = Array.from({ length: this.shape.length }, (_, index) => index);
         }
+        if (Array.isArray(dims)) {
+            // Sort in descending order
+            const sortedDims = dims.sort((a, b) => b - a);
+            let reducedThis = this;
+            for (let i = 0; i < sortedDims.length; i++) {
+                reducedThis = reducedThis.sum(sortedDims[i], true);
+            }
+            return keepDims ? reducedThis : reducedThis.squeeze(dims);
+        }
         // Dims that are reduced now have size-1
-        const outputShape = this.shape.map((dim, i) => dims.includes(i) ? 1 : dim);
+        const outputShape = this.shape.map((dim, i) => dims === i ? 1 : dim);
         const outputStrides = Tensor.getStrides(outputShape);
         const outputSize = Tensor.shapeToSize(outputShape);
         const outputValue = new Array(outputSize).fill(0);
@@ -368,7 +369,7 @@ class Tensor {
         for (let realFlatIndex = 0; realFlatIndex < originalSize; realFlatIndex++) {
             const coords = Tensor.indexToCoords(realFlatIndex, this.strides);
             // Force 0 on reduced axes to collapse into size-1 dims
-            const outCoords = coords.map((val, i) => dims.includes(i) ? 0 : val);
+            const outCoords = coords.map((val, i) => dims === i ? 0 : val);
             // Convert output coordinates to flat index
             const outFlatIndex = Tensor.coordsToIndex(outCoords, outputStrides);
             // Add into sum
@@ -395,21 +396,22 @@ class Tensor {
     }
     // Tensor product reduction
     prod(dims, keepDims = false) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.prod) {
-            return backend.prod(this, dims, keepDims);
-        }
         if (typeof this.value === "number")
             return this;
-        if (typeof dims === "number") {
-            dims = [dims];
-        }
         if (typeof dims === "undefined") {
             dims = Array.from({ length: this.shape.length }, (_, index) => index);
         }
+        if (Array.isArray(dims)) {
+            // Sort in descending order
+            const sortedDims = dims.sort((a, b) => b - a);
+            let reducedThis = this;
+            for (let i = 0; i < sortedDims.length; i++) {
+                reducedThis = reducedThis.prod(sortedDims[i], true);
+            }
+            return keepDims ? reducedThis : reducedThis.squeeze(dims);
+        }
         // Dims that are reduced now have size-1
-        const outputShape = this.shape.map((dim, i) => dims.includes(i) ? 1 : dim);
+        const outputShape = this.shape.map((dim, i) => dims === i ? 1 : dim);
         const outputStrides = Tensor.getStrides(outputShape);
         const outputSize = Tensor.shapeToSize(outputShape);
         const outputValue = new Array(outputSize).fill(1);
@@ -418,7 +420,7 @@ class Tensor {
         for (let realFlatIndex = 0; realFlatIndex < originalSize; realFlatIndex++) {
             const coords = Tensor.indexToCoords(realFlatIndex, this.strides);
             // Force 0 on reduced axes to collapse into size-1 dims
-            const outCoords = coords.map((val, i) => dims.includes(i) ? 0 : val);
+            const outCoords = coords.map((val, i) => dims === i ? 0 : val);
             // Convert output coordinates to flat index
             const outFlatIndex = Tensor.coordsToIndex(outCoords, outputStrides);
             // Multiply into product
@@ -437,7 +439,7 @@ class Tensor {
                 for (let realFlatIndex = 0; realFlatIndex < originalSize; realFlatIndex++) {
                     const coords = Tensor.indexToCoords(realFlatIndex, this.strides);
                     // Force 0 on reduced axes to collapse into size-1 dims
-                    const outCoords = coords.map((val, i) => dims.includes(i) ? 0 : val);
+                    const outCoords = coords.map((val, i) => dims === i ? 0 : val);
                     // Convert output coordinates to flat index
                     const outFlatIndex = Tensor.coordsToIndex(outCoords, outputStrides);
                     // Grad is the product of other elements of the same axis, which is product of all els divided by the current value
@@ -451,21 +453,22 @@ class Tensor {
     }
     // Tensor mean reduction
     mean(dims, keepDims = false) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.mean) {
-            return backend.mean(this, dims, keepDims);
-        }
         if (typeof this.value === "number")
             return this;
-        if (typeof dims === "number") {
-            dims = [dims];
-        }
         if (typeof dims === "undefined") {
             dims = Array.from({ length: this.shape.length }, (_, index) => index);
         }
+        if (Array.isArray(dims)) {
+            // Sort in descending order
+            const sortedDims = dims.sort((a, b) => b - a);
+            let reducedThis = this;
+            for (let i = 0; i < sortedDims.length; i++) {
+                reducedThis = reducedThis.mean(sortedDims[i], true);
+            }
+            return keepDims ? reducedThis : reducedThis.squeeze(dims);
+        }
         // Dims that are reduced now have size-1
-        const outputShape = this.shape.map((dim, i) => dims.includes(i) ? 1 : dim);
+        const outputShape = this.shape.map((dim, i) => dims === i ? 1 : dim);
         const outputStrides = Tensor.getStrides(outputShape);
         const outputSize = Tensor.shapeToSize(outputShape);
         const outputValue = new Array(outputSize).fill(0);
@@ -475,7 +478,7 @@ class Tensor {
         for (let realFlatIndex = 0; realFlatIndex < originalSize; realFlatIndex++) {
             const coords = Tensor.indexToCoords(realFlatIndex, this.strides);
             // Force 0 on reduced axes to collapse into size-1 dims
-            const outCoords = coords.map((val, i) => dims.includes(i) ? 0 : val);
+            const outCoords = coords.map((val, i) => dims === i ? 0 : val);
             // Convert output coordinates to flat index
             const outFlatIndex = Tensor.coordsToIndex(outCoords, outputStrides);
             // Calculate sum and contributors to the sum
@@ -500,7 +503,7 @@ class Tensor {
                 for (let realFlatIndex = 0; realFlatIndex < originalSize; realFlatIndex++) {
                     const coords = Tensor.indexToCoords(realFlatIndex, this.strides);
                     // Force 0 on reduced axes to collapse into size-1 dims
-                    const outCoords = coords.map((val, i) => dims.includes(i) ? 0 : val);
+                    const outCoords = coords.map((val, i) => dims === i ? 0 : val);
                     // Convert output coordinates to flat index
                     const outFlatIndex = Tensor.coordsToIndex(outCoords, outputStrides);
                     // Mean = 1/n * (el1 + el2 + ... + eln) so grad = 1/n
@@ -514,21 +517,22 @@ class Tensor {
     }
     // Tensor maximum reduction
     max(dims, keepDims = false) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.max) {
-            return backend.max(this, dims, keepDims);
-        }
         if (typeof this.value === "number")
             return this;
-        if (typeof dims === "number") {
-            dims = [dims];
-        }
         if (typeof dims === "undefined") {
             dims = Array.from({ length: this.shape.length }, (_, index) => index);
         }
+        if (Array.isArray(dims)) {
+            // Sort in descending order
+            const sortedDims = dims.sort((a, b) => b - a);
+            let reducedThis = this;
+            for (let i = 0; i < sortedDims.length; i++) {
+                reducedThis = reducedThis.max(sortedDims[i], true);
+            }
+            return keepDims ? reducedThis : reducedThis.squeeze(dims);
+        }
         // Dims that are reduced now have size-1
-        const outputShape = this.shape.map((dim, i) => dims.includes(i) ? 1 : dim);
+        const outputShape = this.shape.map((dim, i) => dims === i ? 1 : dim);
         const outputStrides = Tensor.getStrides(outputShape);
         const outputSize = Tensor.shapeToSize(outputShape);
         const outputValue = new Array(outputSize).fill(-Infinity);
@@ -537,7 +541,7 @@ class Tensor {
         for (let realFlatIndex = 0; realFlatIndex < originalSize; realFlatIndex++) {
             const coords = Tensor.indexToCoords(realFlatIndex, this.strides);
             // Force 0 on reduced axes to collapse into size-1 dims
-            const outCoords = coords.map((val, i) => dims.includes(i) ? 0 : val);
+            const outCoords = coords.map((val, i) => dims === i ? 0 : val);
             // Convert output coordinates to flat index
             const outFlatIndex = Tensor.coordsToIndex(outCoords, outputStrides);
             // Get max over time
@@ -555,14 +559,25 @@ class Tensor {
             out.children.push(this);
             out.gradFn = () => {
                 const gradShape = this.shape, gradStrides = this.strides, gradValue = new Array(originalSize).fill(0);
+                const shareCounts = new Array(outputSize).fill(0);
+                const originalValue = this.value;
                 for (let realFlatIndex = 0; realFlatIndex < originalSize; realFlatIndex++) {
                     const coords = Tensor.indexToCoords(realFlatIndex, this.strides);
                     // Force 0 on reduced axes to collapse into size-1 dims
-                    const outCoords = coords.map((val, i) => dims.includes(i) ? 0 : val);
+                    const outCoords = coords.map((val, i) => dims === i ? 0 : val);
                     // Convert output coordinates to flat index
                     const outFlatIndex = Tensor.coordsToIndex(outCoords, outputStrides);
-                    // Calculate grad by checking if a positon holds a value equal to the max value
-                    gradValue[realFlatIndex] = outputValue[outFlatIndex] === this.value[realFlatIndex] ? 1 : 0;
+                    // We collect how many elements share the same max value first
+                    shareCounts[outFlatIndex] += outputValue[outFlatIndex] === originalValue[realFlatIndex] ? 1 : 0;
+                }
+                for (let realFlatIndex = 0; realFlatIndex < originalSize; realFlatIndex++) {
+                    const coords = Tensor.indexToCoords(realFlatIndex, this.strides);
+                    // Force 0 on reduced axes to collapse into size-1 dims
+                    const outCoords = coords.map((val, i) => dims === i ? 0 : val);
+                    // Convert output coordinates to flat index
+                    const outFlatIndex = Tensor.coordsToIndex(outCoords, outputStrides);
+                    // Here we share the grad between the elements that share the same max value
+                    gradValue[realFlatIndex] = outputValue[outFlatIndex] === originalValue[realFlatIndex] ? 1 / shareCounts[outFlatIndex] : 0;
                 }
                 const localGrad = new Tensor(gradValue, { shape: gradShape, strides: gradStrides });
                 Tensor.addGrad(this, out.grad.withGrad(false).mul(localGrad));
@@ -572,21 +587,22 @@ class Tensor {
     }
     // Tensor minimum reduction
     min(dims, keepDims = false) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.min) {
-            return backend.min(this, dims, keepDims);
-        }
         if (typeof this.value === "number")
             return this;
-        if (typeof dims === "number") {
-            dims = [dims];
-        }
         if (typeof dims === "undefined") {
             dims = Array.from({ length: this.shape.length }, (_, index) => index);
         }
+        if (Array.isArray(dims)) {
+            // Sort in descending order
+            const sortedDims = dims.sort((a, b) => b - a);
+            let reducedThis = this;
+            for (let i = 0; i < sortedDims.length; i++) {
+                reducedThis = reducedThis.min(sortedDims[i], true);
+            }
+            return keepDims ? reducedThis : reducedThis.squeeze(dims);
+        }
         // Dims that are reduced now have size-1
-        const outputShape = this.shape.map((dim, i) => dims.includes(i) ? 1 : dim);
+        const outputShape = this.shape.map((dim, i) => dims === i ? 1 : dim);
         const outputStrides = Tensor.getStrides(outputShape);
         const outputSize = Tensor.shapeToSize(outputShape);
         const outputValue = new Array(outputSize).fill(Infinity);
@@ -595,7 +611,7 @@ class Tensor {
         for (let realFlatIndex = 0; realFlatIndex < originalSize; realFlatIndex++) {
             const coords = Tensor.indexToCoords(realFlatIndex, this.strides);
             // Force 0 on reduced axes to collapse into size-1 dims
-            const outCoords = coords.map((val, i) => dims.includes(i) ? 0 : val);
+            const outCoords = coords.map((val, i) => dims === i ? 0 : val);
             // Convert output coordinates to flat index
             const outFlatIndex = Tensor.coordsToIndex(outCoords, outputStrides);
             // Get min over time
@@ -613,14 +629,25 @@ class Tensor {
             out.children.push(this);
             out.gradFn = () => {
                 const gradShape = this.shape, gradStrides = this.strides, gradValue = new Array(originalSize).fill(0);
+                const shareCounts = new Array(outputSize).fill(0);
+                const originalValue = this.value;
+                for (let realFlatIndex = 0; realFlatIndex < originalSize; realFlatIndex++) {
+                    const coords = Tensor.indexToCoords(realFlatIndex, this.strides);
+                    // Force 0 on reduced axes to collapse into size-1 dims
+                    const outCoords = coords.map((val, i) => dims === i ? 0 : val);
+                    // Convert output coordinates to flat index
+                    const outFlatIndex = Tensor.coordsToIndex(outCoords, outputStrides);
+                    // We collect how many elements share the same min value first
+                    shareCounts[outFlatIndex] += outputValue[outFlatIndex] === originalValue[realFlatIndex] ? 1 : 0;
+                }
                 for (let realFlatIndex = 0; realFlatIndex < originalSize; realFlatIndex++) {
                     const coords = Tensor.indexToCoords(realFlatIndex, this.strides);
                     // Force 0 on reduced axes to collapse into size-1 dims
-                    const outCoords = coords.map((val, i) => dims.includes(i) ? 0 : val);
+                    const outCoords = coords.map((val, i) => dims === i ? 0 : val);
                     // Convert output coordinates to flat index
                     const outFlatIndex = Tensor.coordsToIndex(outCoords, outputStrides);
-                    // Calculate grad by checking if a positon holds a value equal to the min value
-                    gradValue[realFlatIndex] = outputValue[outFlatIndex] === this.value[realFlatIndex] ? 1 : 0;
+                    // Here we share the grad between the elements that share the same min value
+                    gradValue[realFlatIndex] = outputValue[outFlatIndex] === originalValue[realFlatIndex] ? 1 / shareCounts[outFlatIndex] : 0;
                 }
                 const localGrad = new Tensor(gradValue, { shape: gradShape, strides: gradStrides });
                 Tensor.addGrad(this, out.grad.withGrad(false).mul(localGrad));
@@ -628,23 +655,34 @@ class Tensor {
         }
         return keepDims ? out : out.squeeze(dims);
     }
+    // Tensor variance reduction
+    var(dims, keepDims = false) {
+        const meanXSquared = this.square().mean(dims, keepDims);
+        const meanXSquaredExpanded = this.mean(dims, keepDims).square();
+        return meanXSquared.sub(meanXSquaredExpanded);
+    }
+    // Tensor standard deviation reduction
+    std(dims, keepDims = false) {
+        return this.var(dims, keepDims).sqrt();
+    }
     // Tensor product reduction
     softmax(dims) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.softmax) {
-            return backend.softmax(this, dims);
-        }
         if (typeof this.value === "number")
             return this;
-        if (typeof dims === "number") {
-            dims = [dims];
-        }
         if (typeof dims === "undefined") {
             dims = Array.from({ length: this.shape.length }, (_, index) => index);
         }
+        if (Array.isArray(dims)) {
+            // Sort in descending order
+            const sortedDims = dims.sort((a, b) => b - a);
+            let reducedThis = this;
+            for (let i = 0; i < sortedDims.length; i++) {
+                reducedThis = reducedThis.softmax(sortedDims[i]);
+            }
+            return reducedThis;
+        }
         // Dims that are reduced now have size-1
-        const expSumShape = this.shape.map((dim, i) => dims.includes(i) ? 1 : dim);
+        const expSumShape = this.shape.map((dim, i) => dims === i ? 1 : dim);
         const expSumStrides = Tensor.getStrides(expSumShape);
         const expSumSize = Tensor.shapeToSize(expSumShape);
         const expSumValue = new Array(expSumSize).fill(0);
@@ -656,7 +694,7 @@ class Tensor {
         for (let realFlatIndex = 0; realFlatIndex < outputSize; realFlatIndex++) {
             const coords = Tensor.indexToCoords(realFlatIndex, outputStrides);
             // Force 0 on reduced axes to collapse into size-1 dims
-            const expSumCoords = coords.map((val, i) => dims.includes(i) ? 0 : val);
+            const expSumCoords = coords.map((val, i) => dims === i ? 0 : val);
             // Convert exp sum coordinates to flat index
             const expSumFlatIndex = Tensor.coordsToIndex(expSumCoords, expSumStrides);
             // Add e^x to the sum cache
@@ -666,7 +704,7 @@ class Tensor {
         for (let realFlatIndex = 0; realFlatIndex < outputSize; realFlatIndex++) {
             const coords = Tensor.indexToCoords(realFlatIndex, outputStrides);
             // Force 0 on reduced axes to collapse into size-1 dims
-            const expSumCoords = coords.map((val, i) => dims.includes(i) ? 0 : val);
+            const expSumCoords = coords.map((val, i) => dims === i ? 0 : val);
             // Convert exp sum coordinates to flat index
             const expSumFlatIndex = Tensor.coordsToIndex(expSumCoords, expSumStrides);
             // Calculate e^xi / sum
@@ -699,488 +737,228 @@ class Tensor {
     }
     // Tensor element-wise addition
     add(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.add) {
-            return backend.add(this, other);
-        }
         return this.elementWiseABDAG(other, (a, b) => a + b, (self, other, outGrad) => outGrad, (self, other, outGrad) => outGrad);
     }
     // Tensor element-wise subtraction
     sub(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.sub) {
-            return backend.sub(this, other);
-        }
         return this.elementWiseABDAG(other, (a, b) => a - b, (self, other, outGrad) => outGrad, (self, other, outGrad) => outGrad.neg());
     }
     subtract = this.sub;
     // Tensor element-wise multiplication
     mul(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.mul) {
-            return backend.mul(this, other);
-        }
         return this.elementWiseABDAG(other, (a, b) => a * b, (self, other, outGrad) => outGrad.mul(other), (self, other, outGrad) => outGrad.mul(self));
     }
     multiply = this.mul;
     // Tensor element-wise power
     pow(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.pow) {
-            return backend.pow(this, other);
-        }
         return this.elementWiseABDAG(other, (a, b) => a ** b, (self, other, outGrad) => outGrad.mul(other.mul(self.pow(other.sub(1)))), (self, other, outGrad) => outGrad.mul(self.pow(other).mul(self.log())));
     }
     // Tensor element-wise division
     div(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.div) {
-            return backend.div(this, other);
-        }
         return this.elementWiseABDAG(other, (a, b) => a / b, (self, other, outGrad) => outGrad.div(other), (self, other, outGrad) => outGrad.mul(self.neg().div(other.square())));
     }
     divide = this.div;
     // Tensor element-wise modulo
     remainder(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.remainder) {
-            return backend.remainder(this, other);
-        }
         return this.elementWiseABDAG(other, (a, b) => a % b);
     }
     // Tensor element-wise greater or equal comparison
     ge(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.ge) {
-            return backend.ge(this, other);
-        }
         return this.elementWiseABDAG(other, (a, b) => a >= b ? 1 : 0);
     }
     greaterEqual = this.ge;
     // Tensor element-wise less or equal comparison
     le(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.le) {
-            return backend.le(this, other);
-        }
         return this.elementWiseABDAG(other, (a, b) => a <= b ? 1 : 0);
     }
     lessEqual = this.le;
     // Tensor element-wise greater-than comparison
     gt(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.gt) {
-            return backend.gt(this, other);
-        }
         return this.elementWiseABDAG(other, (a, b) => a > b ? 1 : 0);
     }
     greater = this.gt;
     // Tensor element-wise less-than comparison
     lt(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.lt) {
-            return backend.lt(this, other);
-        }
         return this.elementWiseABDAG(other, (a, b) => a < b ? 1 : 0);
     }
     less = this.lt;
     // Tensor element-wise equality comparison
     eq(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.eq) {
-            return backend.eq(this, other);
-        }
         return this.elementWiseABDAG(other, (a, b) => a === b ? 1 : 0);
     }
     equal = this.eq;
     // Tensor element-wise not equality comparison
     ne(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.ne) {
-            return backend.ne(this, other);
-        }
         return this.elementWiseABDAG(other, (a, b) => a !== b ? 1 : 0);
     }
     notEqual = this.ne;
     // Tensor element-wise logical and
     logicalAnd(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.logicalAnd) {
-            return backend.logicalAnd(this, other);
-        }
         return this.elementWiseABDAG(other, (a, b) => a === 1 && b === 1 ? 1 : 0);
     }
     // Tensor element-wise logical or
     logicalOr(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.logicalOr) {
-            return backend.logicalOr(this, other);
-        }
         return this.elementWiseABDAG(other, (a, b) => a === 1 || b === 1 ? 1 : 0);
     }
     // Tensor element-wise logical xor
     logicalXor(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.logicalXor) {
-            return backend.logicalXor(this, other);
-        }
         return this.elementWiseABDAG(other, (a, b) => (a === 1 || b === 1) && a !== b ? 1 : 0);
     }
     // Tensor element-wise logical not
     logicalNot() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.logicalNot) {
-            return backend.logicalNot(this);
-        }
         return this.elementWiseSelfDAG((a) => a === 1 ? 0 : 1);
     }
     // Tensor element-wise bitwise and
     bitwiseAnd(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.bitwiseAnd) {
-            return backend.bitwiseAnd(this, other);
-        }
         return this.elementWiseABDAG(other, (a, b) => a & b);
     }
     // Tensor element-wise bitwise or
     bitwiseOr(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.bitwiseOr) {
-            return backend.bitwiseOr(this, other);
-        }
         return this.elementWiseABDAG(other, (a, b) => a | b);
     }
     // Tensor element-wise bitwise xor
     bitwiseXor(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.bitwiseXor) {
-            return backend.bitwiseXor(this, other);
-        }
         return this.elementWiseABDAG(other, (a, b) => a ^ b);
     }
     // Tensor element-wise bitwise not
     bitwiseNot() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.bitwiseNot) {
-            return backend.bitwiseNot(this);
-        }
         return this.elementWiseSelfDAG((a) => ~a);
     }
     // Tensor element-wise left shift
     bitwiseLeftShift(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.bitwiseLeftShift) {
-            return backend.bitwiseLeftShift(this, other);
-        }
         return this.elementWiseABDAG(other, (a, b) => a << b);
     }
     // Tensor element-wise right shift
     bitwiseRightShift(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.bitwiseRightShift) {
-            return backend.bitwiseRightShift(this, other);
-        }
         return this.elementWiseABDAG(other, (a, b) => a >> b);
     }
     // Tensor element-wise negation
     neg() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.neg) {
-            return backend.neg(this);
-        }
         return this.elementWiseSelfDAG((a) => -a, (self, outGrad) => outGrad.mul(-1));
     }
     negative = this.neg;
     // Tensor element-wise reciprocal
     reciprocal() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.reciprocal) {
-            return backend.reciprocal(this);
-        }
         return this.elementWiseSelfDAG((a) => 1 / a, (self, outGrad) => outGrad.mul(self.pow(-2).neg()));
     }
     // Tensor element-wise square
     square() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.square) {
-            return backend.square(this);
-        }
         return this.elementWiseSelfDAG((a) => a * a, (self, outGrad) => outGrad.mul(self.mul(2)));
     }
     // Tensor element-wise absolute
     abs() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.abs) {
-            return backend.abs(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.abs(a), (self, outGrad) => outGrad.mul(self.sign()));
     }
     absolute = this.abs;
     // Tensor element-wise sign function
     sign() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.sign) {
-            return backend.sign(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.sign(a));
     }
     // Tensor element-wise sin
     sin() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.sin) {
-            return backend.sin(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.sin(a), (self, outGrad) => outGrad.mul(self.cos()));
     }
     // Tensor element-wise cos
     cos() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.cos) {
-            return backend.cos(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.cos(a), (self, outGrad) => outGrad.mul(self.sin().neg()));
     }
     // Tensor element-wise tan
     tan() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.tan) {
-            return backend.tan(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.tan(a), (self, outGrad) => outGrad.mul(self.tan().square().add(1)));
     }
     // Tensor element-wise asin
     asin() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.asin) {
-            return backend.asin(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.asin(a), (self, outGrad) => outGrad.div(self.square().neg().add(1).sqrt()));
     }
     arcsin = this.asin;
     // Tensor element-wise acos
     acos() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.acos) {
-            return backend.acos(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.acos(a), (self, outGrad) => outGrad.div(self.square().neg().add(1).sqrt()).neg());
     }
     arccos = this.acos;
     // Tensor element-wise atan
     atan() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.atan) {
-            return backend.atan(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.atan(a), (self, outGrad) => outGrad.div(self.square().add(1)));
     }
     arctan = this.atan;
     // Tensor element-wise atan2
     atan2(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.atan2) {
-            return backend.atan2(this);
-        }
         return this.elementWiseABDAG(other, (a, b) => Math.atan2(a, b), (self, other, outGrad) => outGrad.mul(other.div(self.square().add(other.square()))), (self, other, outGrad) => outGrad.mul(self.neg().div(self.square().add(other.square()))));
     }
     arctan2 = this.atan2;
     // Tensor element-wise sinh
     sinh() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.sinh) {
-            return backend.sinh(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.sinh(a), (self, outGrad) => outGrad.mul(self.cosh()));
     }
     // Tensor element-wise cosh
     cosh() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.cosh) {
-            return backend.cosh(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.cosh(a), (self, outGrad) => outGrad.mul(self.sinh()));
     }
     // Tensor element-wise asinh
     asinh() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.asinh) {
-            return backend.asinh(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.asinh(a), (self, outGrad) => outGrad.div(self.square().add(1).sqrt()));
     }
     arcsinh = this.asinh;
     // Tensor element-wise acosh
     acosh() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.acosh) {
-            return backend.acosh(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.acosh(a), (self, outGrad) => outGrad.div(self.add(1).sqrt().mul(self.sub(1).sqrt())));
     }
     arccosh = this.acosh;
     // Tensor element-wise atanh
     atanh() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.atanh) {
-            return backend.atanh(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.atanh(a), (self, outGrad) => outGrad.div(self.square().neg().add(1)));
     }
     arctanh = this.atanh;
     // Tensor element-wise degree to radian
     deg2rad() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.deg2rad) {
-            return backend.deg2rad(this);
-        }
         return this.elementWiseSelfDAG((a) => a * (Math.PI / 180), (self, outGrad) => outGrad.mul(Math.PI / 180));
     }
     // Tensor element-wise radian to degree
     rad2deg() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.rad2deg) {
-            return backend.rad2deg(this);
-        }
         return this.elementWiseSelfDAG((a) => a / (Math.PI / 180), (self, outGrad) => outGrad.div(Math.PI / 180));
     }
     // Tensor element-wise square root
     sqrt() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.sqrt) {
-            return backend.sqrt(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.sqrt(a), (self, outGrad) => outGrad.div(self.sqrt().mul(2)));
     }
     // Tensor element-wise reciprocal of square root
     rsqrt() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.rsqrt) {
-            return backend.rsqrt(this);
-        }
         return this.elementWiseSelfDAG((a) => 1 / Math.sqrt(a), (self, outGrad) => outGrad.mul(self.pow(-1.5).mul(-0.5)));
     }
     // Tensor element-wise e^x
     exp() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.exp) {
-            return backend.exp(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.exp(a), (self, outGrad) => outGrad.mul(self.exp()));
     }
     // Tensor element-wise 2^x
     exp2() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.exp2) {
-            return backend.exp2(this);
-        }
         return this.elementWiseSelfDAG((a) => 2 ** a, (self, outGrad) => outGrad.mul(self.exp2().mul(Math.log(2))));
     }
     // Tensor element-wise e^x - 1
     expm1() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.expm1) {
-            return backend.expm1(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.expm1(a), (self, outGrad) => outGrad.mul(self.exp()));
     }
     // Tensor element-wise natural log
     log() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.log) {
-            return backend.log(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.log(a), (self, outGrad) => outGrad.div(self));
     }
     // Tensor element-wise log2
     log2() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.log2) {
-            return backend.log2(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.log2(a), (self, outGrad) => outGrad.div(self.mul(Math.log(2))));
     }
     // Tensor element-wise log10
     log10() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.log10) {
-            return backend.log10(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.log10(a), (self, outGrad) => outGrad.div(self.mul(Math.log(10))));
     }
     // Tensor element-wise log(1+x)
     log1p() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.log1p) {
-            return backend.log1p(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.log1p(a), (self, outGrad) => outGrad.div(self.add(1)));
     }
     // Tensor element-wise relu
     relu() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.relu) {
-            return backend.relu(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.max(a, 0), (self, outGrad) => outGrad.mul(self.gt(0)));
     }
     // Tensor element-wise sigmoid
     sigmoid() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.sigmoid) {
-            return backend.sigmoid(this);
-        }
         return this.elementWiseSelfDAG((a) => 1 / (1 + Math.exp(-a)), (self, outGrad) => {
             const sig = self.sigmoid();
             return outGrad.mul(sig).mul(sig.neg().add(1));
@@ -1188,38 +966,18 @@ class Tensor {
     }
     // Tensor element-wise tanh
     tanh() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.tanh) {
-            return backend.tanh(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.tanh(a), (self, outGrad) => outGrad.mul(self.tanh().square().neg().add(1)));
     }
     // Tensor element-wise softplus
     softplus() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.softplus) {
-            return backend.softplus(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.log1p(Math.exp(a)), (self, outGrad) => outGrad.mul(self.sigmoid()));
     }
     // Tensor element-wise softsign
     softsign() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.softsign) {
-            return backend.softsign(this);
-        }
         return this.elementWiseSelfDAG((a) => a / (1 + Math.abs(a)), (self, outGrad) => outGrad.div(self.abs().add(1).square()));
     }
     // Tensor element-wise silu (swish)
     silu() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.silu) {
-            return backend.silu(this);
-        }
         return this.elementWiseSelfDAG((a) => a / (1 + Math.exp(-a)), (self, outGrad) => {
             const sig = self.sigmoid();
             return outGrad.mul(sig.add(self.mul(sig).mul(sig.neg().add(1))));
@@ -1227,11 +985,6 @@ class Tensor {
     }
     // Tensor element-wise mish
     mish() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.mish) {
-            return backend.mish(this);
-        }
         return this.elementWiseSelfDAG((a) => a * Math.tanh(Math.log1p(Math.exp(a))), (self, outGrad) => {
             const tanhSoftPlus = self.exp().add(1).log().tanh();
             // tanh(softplus(x)) + x * (1 - tanh²(softplus(x))) * sigmoid(x)
@@ -1241,103 +994,48 @@ class Tensor {
     }
     // Tensor element-wise maximum
     maximum(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.maximum) {
-            return backend.maximum(this, other);
-        }
         return this.elementWiseABDAG(other, (a, b) => Math.max(a, b), (self, other, outGrad) => outGrad.mul(self.gt(other).add(self.eq(other).mul(0.5))), (self, other, outGrad) => outGrad.mul(other.gt(self).add(other.eq(self).mul(0.5))));
     }
     // Tensor element-wise minimum
     minimum(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.minimum) {
-            return backend.minimum(this, other);
-        }
         return this.elementWiseABDAG(other, (a, b) => Math.min(a, b), (self, other, outGrad) => outGrad.mul(self.lt(other).add(self.eq(other).mul(0.5))), (self, other, outGrad) => outGrad.mul(other.lt(self).add(other.eq(self).mul(0.5))));
     }
     // Tensor element-wise round
     round() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.round) {
-            return backend.round(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.round(a));
     }
     // Tensor element-wise floor
     floor() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.floor) {
-            return backend.floor(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.floor(a));
     }
     // Tensor element-wise ceil
     ceil() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.ceil) {
-            return backend.ceil(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.ceil(a));
     }
     // Tensor element-wise truncation
     trunc() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.trunc) {
-            return backend.trunc(this);
-        }
         return this.elementWiseSelfDAG((a) => Math.trunc(a));
     }
     fix = this.trunc;
     // Tensor element-wise fraction portion
     frac() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.frac) {
-            return backend.frac(this);
-        }
         return this.elementWiseSelfDAG((a) => a - Math.floor(a));
     }
     // Tensor element-wise clip and clamp
     clip(min, max) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.clip) {
-            return backend.clip(this, min, max);
-        }
         return this.elementWiseSelfDAG((a) => Math.max(min, Math.min(max, a)), (self, outGrad) => outGrad.mul(self.ge(min).mul(self.le(max))));
     }
     clamp = this.clip;
     // Tensor element-wise error function
     erf() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.erf) {
-            return backend.erf(this);
-        }
         return this.elementWiseSelfDAG((a) => (0, utils_1.erf)(a), (self, outGrad) => outGrad.mul(self.square().neg().exp().mul(2 / Math.sqrt(Math.PI))));
     }
     // Tensor element-wise complementary error function
     erfc() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.erfc) {
-            return backend.erfc(this);
-        }
         return this.elementWiseSelfDAG((a) => (0, utils_1.erfc)(a), (self, outGrad) => outGrad.mul(self.square().neg().exp().mul(2 / Math.sqrt(Math.PI)).neg()));
     }
     // Tensor element-wise inverse error function
     erfinv() {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.erfinv) {
-            return backend.erfinv(this);
-        }
         return this.elementWiseSelfDAG((a) => (0, utils_1.erfinv)(a), (self, outGrad) => outGrad.mul(self.erfinv().square().exp().mul(Math.sqrt(Math.PI) / 2)));
     }
     // Transpose
@@ -1379,11 +1077,6 @@ class Tensor {
     }
     // 1D tensor dot product
     dot(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.dot) {
-            return backend.dot(this, other);
-        }
         other = Tensor.forceTensor(other);
         // Verify 1D shape
         if (this.shape.length !== 1 || other.shape.length !== 1) {
@@ -1422,11 +1115,6 @@ class Tensor {
     }
     // Matrix multiplication
     mm(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.mm) {
-            return backend.mm(this, other);
-        }
         other = Tensor.forceTensor(other);
         // Verify 2D shape
         if (this.shape.length !== 2 || other.shape.length !== 2) {
@@ -1482,11 +1170,6 @@ class Tensor {
     }
     // Batched 3D tensor matmul
     bmm(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.bmm) {
-            return backend.bmm(this, other);
-        }
         other = Tensor.forceTensor(other);
         // Verify 3D shape
         if (this.shape.length !== 3 || other.shape.length !== 3 || this.shape[0] !== other.shape[0]) {
@@ -1545,11 +1228,6 @@ class Tensor {
     }
     // Convert right-side 1D tensor to a vector (nx1 tensor) to do matmul
     mv(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.mv) {
-            return backend.mv(this, other);
-        }
         other = Tensor.forceTensor(other);
         // Verify 2D shape
         if (this.shape.length !== 2 || other.shape.length !== 1) {
@@ -1559,11 +1237,6 @@ class Tensor {
     }
     // General matrix multiplication with different shapes
     matmul(other) {
-        // Use backend of tensor's device if available, or else fallback to cpu
-        const backend = Tensor.backends.get(this.device);
-        if (backend && backend.matmul) {
-            return backend.matmul(this, other);
-        }
         other = Tensor.forceTensor(other);
         const isThis1D = this.shape.length === 1;
         const isOther1D = other.shape.length === 1;
@@ -1875,10 +1548,10 @@ class Tensor {
     // Op to transfer tensor to another device
     to(device) {
         const backend = Tensor.backends.get(device);
-        if (backend && backend.to) {
-            return backend.to(this);
+        if (backend && backend.transfer) {
+            return backend.transfer(this);
         }
-        throw new Error(`No device found to transfer tensor to or "to" is not implemented for device.`);
+        throw new Error(`No device found to transfer tensor to or a handler is not implemented for device.`);
     }
 }
 exports.Tensor = Tensor;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "catniff",
-  "version": "0.4.0",
+  "version": "0.4.2",
   "description": "A small Torch-like deep learning framework for Javascript with tensor and autograd support",
   "main": "index.js",
   "scripts": {