catniff 0.8.1 → 0.8.3

package/dist/core.d.ts

@@ -69,6 +69,8 @@ export declare class Tensor {
     cat(other: Tensor | TensorValue, dim?: number): Tensor;
     squeeze(dims?: number[] | number): Tensor;
     unsqueeze(dim: number): Tensor;
+    sort(dim?: number, descending?: boolean): Tensor;
+    topk(k: number, dim?: number, largest?: boolean): Tensor;
     static reduce(tensor: Tensor, dims: number[] | number | undefined, keepDims: boolean, config: {
         identity: number;
         operation: (accumulator: number, value: number) => number;

package/dist/core.js

@@ -208,13 +208,16 @@ class Tensor {
         const outputStrides = Tensor.getStrides(outputShape);
         const outputSize = Tensor.shapeToSize(outputShape);
         const outputValue = new dtype_1.TypedArray[outputDtype](outputSize);
+        // Check fast path conditions of two tensors
+        const aFastPath = tA.isContiguous() && tA.numel === outputSize;
+        const bFastPath = tB.isContiguous() && tB.numel === outputSize;
         for (let i = 0; i < outputSize; i++) {
             // Get coordinates from 1D index
-            const coordsOutput = Tensor.indexToCoords(i, outputStrides);
+            const coordsOutput = aFastPath && bFastPath ? [] : Tensor.indexToCoords(i, outputStrides);
             // Convert the coordinates to 1D index of flattened A with respect to A's shape
-            const indexA = Tensor.coordsToUnbroadcastedIndex(coordsOutput, paddedAShape, paddedAStrides);
+            const indexA = aFastPath ? i : Tensor.coordsToUnbroadcastedIndex(coordsOutput, paddedAShape, paddedAStrides);
             // Convert the coordinates to 1D index of flattened B with respect to B's shape
-            const indexB = Tensor.coordsToUnbroadcastedIndex(coordsOutput, paddedBShape, paddedBStrides);
+            const indexB = bFastPath ? i : Tensor.coordsToUnbroadcastedIndex(coordsOutput, paddedBShape, paddedBStrides);
             // Calculate with op
             outputValue[i] = op(tA.value[indexA + tA.offset], tB.value[indexB + tB.offset]);
         }
@@ -934,6 +937,114 @@ class Tensor {
         }
         return out;
     }
+    // Tensor sort
+    sort(dim = -1, descending = false) {
+        if (dim < 0) {
+            dim += this.shape.length;
+        }
+        // If dimension out of bound, throw error
+        if (dim >= this.shape.length || dim < 0) {
+            throw new Error("Dimension do not exist to sort");
+        }
+        // Copy if not contiguous
+        const outputSize = this.numel;
+        const outputShape = this.shape;
+        let outputValue, outputStrides;
+        if (this.isContiguous()) {
+            outputValue = [...this.value];
+            outputStrides = this.strides;
+        }
+        else {
+            outputValue = new dtype_1.TypedArray[this.dtype](outputSize);
+            outputStrides = Tensor.getStrides(outputShape);
+            for (let flatIndex = 0; flatIndex < outputSize; flatIndex++) {
+                const coords = Tensor.indexToCoords(flatIndex, outputStrides);
+                const originalIndex = Tensor.coordsToIndex(coords, this.strides);
+                outputValue[flatIndex] = this.value[originalIndex + this.offset];
+            }
+        }
+        // Calculate dimensions for gather-scatter
+        const dimSize = outputShape[dim];
+        const outerSize = outputShape.slice(0, dim).reduce((a, b) => a * b, 1);
+        const innerSize = outputShape.slice(dim + 1).reduce((a, b) => a * b, 1);
+        // Store permutation indices for gradient
+        const permutation = new Array(outputSize);
+        // Sort each group independently
+        for (let outer = 0; outer < outerSize; outer++) {
+            for (let inner = 0; inner < innerSize; inner++) {
+                const group = [];
+                for (let i = 0; i < dimSize; i++) {
+                    const flatIdx = outer * (dimSize * innerSize) + i * innerSize + inner;
+                    group.push({
+                        value: outputValue[flatIdx],
+                        dimIdx: i
+                    });
+                }
+                // Sort this group by value
+                group.sort((a, b) => descending ? b.value - a.value : a.value - b.value);
+                // Scatter: write back sorted values and record permutation
+                for (let i = 0; i < dimSize; i++) {
+                    const flatIdx = outer * (dimSize * innerSize) + i * innerSize + inner;
+                    outputValue[flatIdx] = group[i].value;
+                    // Record where this element came from (for gradient)
+                    const originalFlatIdx = outer * (dimSize * innerSize) + group[i].dimIdx * innerSize + inner;
+                    permutation[flatIdx] = originalFlatIdx;
+                }
+            }
+        }
+        const out = new Tensor(outputValue, {
+            shape: outputShape,
+            strides: outputStrides,
+            offset: 0,
+            numel: outputSize,
+            device: this.device,
+            dtype: this.dtype
+        });
+        // Gradient setup
+        if (this.requiresGrad) {
+            out.requiresGrad = true;
+            out.children.push(this);
+            out.gradFn = () => {
+                const outGrad = out.grad;
+                // Scatter output gradients back to original positions
+                const inputGradValue = new dtype_1.TypedArray[this.dtype](outputSize);
+                for (let sortedIdx = 0; sortedIdx < outputSize; sortedIdx++) {
+                    const originalIdx = permutation[sortedIdx];
+                    inputGradValue[originalIdx] = outGrad.value[sortedIdx];
+                }
+                const inputGrad = new Tensor(inputGradValue, {
+                    shape: outputShape,
+                    strides: outputStrides,
+                    offset: 0,
+                    numel: outputSize,
+                    device: this.device,
+                    dtype: this.dtype
+                });
+                Tensor.addGrad(this, inputGrad);
+            };
+        }
+        return out;
+    }
+    // Top-k sampling
+    topk(k, dim = -1, largest = true) {
+        if (dim < 0) {
+            dim += this.shape.length;
+        }
+        // If dimension out of bound, throw error
+        if (dim >= this.shape.length || dim < 0) {
+            throw new Error("Dimension do not exist to get topk");
+        }
+        const dimRanges = new Array(this.shape.length);
+        for (let index = 0; index < dimRanges.length; index++) {
+            if (index === dim) {
+                dimRanges[index] = [0, k];
+            }
+            else {
+                dimRanges[index] = [];
+            }
+        }
+        return this.sort(dim, largest).slice(dimRanges);
+    }
     // Generic reduction operation handler
     static reduce(tensor, dims, keepDims, config) {
         if (tensor.shape.length === 0)

package/dist/nn.d.ts

@@ -70,6 +70,7 @@ export declare class Embedding {
     constructor(numEmbeddings: number, embeddingDim: number, device?: string, dtype?: dtype);
     forward(input: Tensor | TensorValue): Tensor;
 }
+export declare function scaledDotProductAttention(query: Tensor, key: Tensor, value: Tensor, attnMask?: Tensor, dropout?: number, isCausal?: boolean, scale?: number): Tensor;
 export declare class MultiheadAttention {
     qProjection: Linear;
     kProjection: Linear;
@@ -80,7 +81,7 @@ export declare class MultiheadAttention {
     headDim: number;
     dropout: number;
     constructor(embedDim: number, numHeads: number, dropout?: number, bias?: boolean, device?: string, dtype?: dtype);
-    forward(query: Tensor, key: Tensor, value: Tensor, needWeights?: boolean, attnMask?: Tensor, averageAttnWeights?: boolean): [Tensor, Tensor | undefined];
+    forward(query: Tensor, key: Tensor, value: Tensor, needWeights?: boolean, attnMask?: Tensor, averageAttnWeights?: boolean, isCausal?: boolean): [Tensor, Tensor | undefined];
 }
 export interface StateDict {
     [key: string]: any;
@@ -93,6 +94,7 @@ export declare const nn: {
     LayerNorm: typeof LayerNorm;
     RMSNorm: typeof RMSNorm;
     Embedding: typeof Embedding;
+    scaledDotProductAttention: typeof scaledDotProductAttention;
     MultiheadAttention: typeof MultiheadAttention;
     state: {
         getParameters(model: any, visited?: WeakSet<object>): Tensor[];

package/dist/nn.js

@@ -1,6 +1,7 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.nn = exports.MultiheadAttention = exports.Embedding = exports.RMSNorm = exports.LayerNorm = exports.LSTMCell = exports.GRUCell = exports.RNNCell = exports.Linear = void 0;
+exports.scaledDotProductAttention = scaledDotProductAttention;
 const core_1 = require("./core");
 function linearTransform(input, weight, bias) {
     let output = input.matmul(weight.t());
@@ -240,6 +241,25 @@ class Embedding {
     }
 }
 exports.Embedding = Embedding;
+function scaledDotProductAttention(query, key, value, attnMask, dropout = 0, isCausal = false, scale) {
+    const targetLen = query.shape[query.shape.length - 2];
+    const sourceLen = key.shape[key.shape.length - 2];
+    const dimSize = query.shape[query.shape.length - 1];
+    // Attention scores
+    let scores = query.matmul(key.transpose(-2, -1)).div(scale ?? Math.sqrt(dimSize));
+    // Set attention mask to causal mask if specified
+    if (isCausal) {
+        attnMask = core_1.Tensor.ones([targetLen, sourceLen], { device: query.device }).triu(1);
+    }
+    // Apply attention mask if specified
+    if (attnMask) {
+        scores = scores.maskedFill(attnMask, -Infinity);
+    }
+    // Calculate attention weights
+    let attnWeights = scores.softmax().dropout(dropout);
+    // Apply attention to values
+    return attnWeights.matmul(value);
+}
 class MultiheadAttention {
     qProjection;
     kProjection;
@@ -259,7 +279,7 @@ class MultiheadAttention {
         this.headDim = Math.floor(embedDim / numHeads);
         this.dropout = dropout;
     }
-    forward(query, key, value, needWeights = true, attnMask, averageAttnWeights = true) {
+    forward(query, key, value, needWeights = true, attnMask, averageAttnWeights = true, isCausal = false) {
         // Batch-first
         const [batchSize, targetLen, embedDim] = query.shape;
         const sourceLen = key.shape[1];
@@ -272,6 +292,10 @@ class MultiheadAttention {
         V = V.reshape([batchSize, sourceLen, this.numHeads, this.headDim]).transpose(1, 2);
         // Attention scores
         let scores = Q.matmul(K.transpose(-2, -1)).div(Math.sqrt(this.headDim));
+        // Set attention mask to causal mask if specified
+        if (isCausal) {
+            attnMask = core_1.Tensor.ones([targetLen, sourceLen], { device: this.qProjection.weight.device }).triu(1);
+        }
         // Apply attention mask if specified
         if (attnMask) {
             scores = scores.maskedFill(attnMask, -Infinity);
@@ -362,6 +386,7 @@ exports.nn = {
     LayerNorm,
     RMSNorm,
     Embedding,
+    scaledDotProductAttention,
     MultiheadAttention,
     state
 };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "catniff",
-  "version": "0.8.1",
+  "version": "0.8.3",
   "description": "Torch-like deep learning framework for Javascript",
   "main": "index.js",
   "scripts": {