feat(src/utils): add GPU acceleration and improve threading for embeddings and distance calculations

- Integrated GPU acceleration via WebGL in `embedding.ts` and `distance.ts` for optimized performance.
- Enhanced fallback mechanisms for CPU processing when GPU is unavailable or fails to initialize.
- Added threading support for embedding and distance calculations in `embedding.ts` and `distance.ts` to improve performance and avoid blocking the main thread.
- Refactored TensorFlow.js imports to dynamically load modules (`@tensorflow/tfjs-core`, `@tensorflow/tfjs-backend-webgl`, `@tensorflow/tfjs-backend-cpu`) for modular dependency usage.
- Introduced GPU-accelerated batch distance calculations in `distance.ts` with appropriate error handling.
- Applied consistent formatting to improve code readability and maintainability while adhering to project style guidelines.
This commit is contained in:
David Snelling 2025-06-26 19:04:37 -07:00
parent bd123c4bb9
commit 6a8d044970
3 changed files with 348 additions and 65 deletions

View file

@ -1,9 +1,12 @@
/**
* Distance functions for vector similarity calculations
* Optimized for Node.js 23.11+ using enhanced array methods
* GPU-accelerated versions available for high-performance computing
*/
import { DistanceFunction, Vector } from '../coreTypes.js'
import { executeInThread } from './workerUtils.js'
import { isThreadingAvailable } from './environment.js'
/**
* Calculates the Euclidean distance between two vectors
@ -41,8 +44,8 @@ export const cosineDistance: DistanceFunction = (a: Vector, b: Vector): number =
dotProduct: acc.dotProduct + (val * b[i]),
normA: acc.normA + (val * val),
normB: acc.normB + (b[i] * b[i])
};
}, { dotProduct: 0, normA: 0, normB: 0 });
}
}, { dotProduct: 0, normA: 0, normB: 0 })
if (normA === 0 || normB === 0) {
return 2 // Maximum distance for zero vectors
@ -84,3 +87,198 @@ export const dotProductDistance: DistanceFunction = (a: Vector, b: Vector): numb
// Convert to a distance metric (lower is better)
return -dotProduct
}
/**
* GPU-accelerated batch distance calculation
* Uses TensorFlow.js with WebGL backend when available for optimal performance
* Falls back to CPU processing when GPU is not available
*
* @param queryVector The query vector to compare against all vectors
* @param vectors Array of vectors to compare against
* @param distanceFunction The distance function to use
* @returns Promise resolving to array of distances
*/
export async function calculateDistancesWithGPU(
queryVector: Vector,
vectors: Vector[],
distanceFunction: DistanceFunction = euclideanDistance
): Promise<number[]> {
// For small batches, use the standard distance function
if (vectors.length < 10) {
return vectors.map(vector => distanceFunction(queryVector, vector))
}
try {
// Function to be executed in a worker thread
const distanceCalculator = async (
args: {
queryVector: Vector,
vectors: Vector[],
distanceFnString: string
}
) => {
const { queryVector, vectors, distanceFnString } = args
// Try to use TensorFlow.js with GPU acceleration if available
const useTensorFlow = async () => {
try {
// TensorFlow.js will use its default EPSILON value
// Dynamically import TensorFlow.js core module and backends
const tf = await import('@tensorflow/tfjs-core')
// Import CPU backend as fallback
await import('@tensorflow/tfjs-backend-cpu')
let usingGPU = false
try {
// Try to import and use WebGL backend (GPU)
await import('@tensorflow/tfjs-backend-webgl')
// Check if WebGL is available and set it as the backend
if (await tf.findBackend('webgl') || await tf.ready().then(() => tf.findBackend('webgl'))) {
await tf.setBackend('webgl')
usingGPU = true
} else {
await tf.setBackend('cpu')
}
} catch (err) {
// If WebGL fails, use CPU
await tf.setBackend('cpu')
}
// Convert vectors to tensors
const queryTensor = tf.tensor2d([queryVector])
const vectorsTensor = tf.tensor2d(vectors)
let distances: number[]
// Calculate distances based on the distance function type
if (distanceFnString.includes('euclideanDistance')) {
// Euclidean distance using GPU-optimized operations
// Formula: sqrt(sum((a - b)^2))
const expanded = tf.sub((queryTensor as any).expandDims(1), (vectorsTensor as any).expandDims(0))
const squaredDiff = tf.square(expanded)
const sumSquaredDiff = tf.sum(squaredDiff, -1)
const distancesTensor = tf.sqrt(sumSquaredDiff)
distances = await (distancesTensor as any).squeeze().array() as number[]
// Clean up tensors
queryTensor.dispose()
vectorsTensor.dispose()
expanded.dispose()
squaredDiff.dispose()
sumSquaredDiff.dispose()
distancesTensor.dispose()
} else if (distanceFnString.includes('cosineDistance')) {
// Cosine distance using GPU-optimized operations
// Formula: 1 - (a·b / (||a|| * ||b||))
const dotProduct = tf.matMul(queryTensor, (vectorsTensor as any).transpose())
const queryNorm = tf.norm(queryTensor, 2, 1)
const vectorsNorm = tf.norm(vectorsTensor, 2, 1)
const normProduct = tf.outerProduct(queryNorm as any, vectorsNorm as any)
const cosineSimilarity = tf.div(dotProduct, normProduct)
const distancesTensor = tf.sub(tf.scalar(1), cosineSimilarity)
distances = await (distancesTensor as any).squeeze().array() as number[]
// Clean up tensors
queryTensor.dispose()
vectorsTensor.dispose()
dotProduct.dispose()
queryNorm.dispose()
vectorsNorm.dispose()
normProduct.dispose()
cosineSimilarity.dispose()
distancesTensor.dispose()
} else if (distanceFnString.includes('manhattanDistance')) {
// Manhattan distance using GPU-optimized operations
// Formula: sum(|a - b|)
const diff = tf.sub((queryTensor as any).expandDims(1), (vectorsTensor as any).expandDims(0))
const absDiff = tf.abs(diff)
const distancesTensor = tf.sum(absDiff, -1)
distances = await (distancesTensor as any).squeeze().array() as number[]
// Clean up tensors
queryTensor.dispose()
vectorsTensor.dispose()
diff.dispose()
absDiff.dispose()
distancesTensor.dispose()
} else if (distanceFnString.includes('dotProductDistance')) {
// Dot product distance using GPU-optimized operations
// Formula: -sum(a * b)
const dotProduct = tf.matMul(queryTensor, (vectorsTensor as any).transpose())
const distancesTensor = tf.neg(dotProduct)
distances = await (distancesTensor as any).squeeze().array() as number[]
// Clean up tensors
queryTensor.dispose()
vectorsTensor.dispose()
dotProduct.dispose()
distancesTensor.dispose()
} else {
// For unknown distance functions, fall back to CPU implementation
throw new Error('Unsupported distance function for GPU acceleration')
}
return {
distances,
usingGPU
}
} catch (error) {
// If TensorFlow.js fails, fall back to CPU implementation
throw error
}
}
// Try to use TensorFlow.js with GPU acceleration
try {
return await useTensorFlow()
} catch (error) {
// Fall back to CPU implementation if TensorFlow.js fails
// Recreate the distance function from its string representation
const distanceFunction = new Function('return ' + distanceFnString)() as DistanceFunction
// Calculate distances for all vectors
const distances = vectors.map(vector => distanceFunction(queryVector, vector))
return {
distances,
usingGPU: false
}
}
}
// Execute the distance calculation in a separate thread if threading is available
if (isThreadingAvailable()) {
try {
// Convert the distance function to a string for serialization
const distanceFnString = distanceFunction.toString()
// Execute in a separate thread
const result = await executeInThread<{ distances: number[], usingGPU: boolean }>(
distanceCalculator.toString(),
{ queryVector, vectors, distanceFnString }
)
return result.distances
} catch (error) {
// Fall back to main thread if threading fails
console.warn('Threaded distance calculation failed, falling back to main thread:', error)
}
}
// If threading is not available or failed, calculate distances in the main thread
return vectors.map(vector => distanceFunction(queryVector, vector))
} catch (error) {
// If anything fails, fall back to the standard distance function
console.error('GPU-accelerated distance calculation failed:', error)
return vectors.map(vector => distanceFunction(queryVector, vector))
}
}

View file

@ -9,6 +9,9 @@ import { executeInThread } from './workerUtils.js'
* TensorFlow Universal Sentence Encoder embedding model
* This model provides high-quality text embeddings using TensorFlow.js
* The required TensorFlow.js dependencies are automatically installed with this package
*
* This implementation will use GPU acceleration via WebGL when available,
* falling back to CPU processing when GPU is not available or fails to initialize.
*/
export class UniversalSentenceEncoder implements EmbeddingModel {
private model: any = null
@ -38,31 +41,38 @@ export class UniversalSentenceEncoder implements EmbeddingModel {
originalWarn(message, ...optionalParams)
}
// Define EPSILON flag before TensorFlow.js is loaded
// This prevents the "Cannot evaluate flag 'EPSILON': no evaluation function found" error
if (typeof window !== 'undefined') {
;(window as any).EPSILON = 1e-7
// Define the flag with an evaluation function for TensorFlow.js
;(window as any).ENV = (window as any).ENV || {}
;(window as any).ENV.flagRegistry =
(window as any).ENV.flagRegistry || {}
;(window as any).ENV.flagRegistry.EPSILON = {
evaluationFn: () => 1e-7
}
} else if (typeof global !== 'undefined') {
;(global as any).EPSILON = 1e-7
// Define the flag with an evaluation function for TensorFlow.js
;(global as any).ENV = (global as any).ENV || {}
;(global as any).ENV.flagRegistry =
(global as any).ENV.flagRegistry || {}
;(global as any).ENV.flagRegistry.EPSILON = {
evaluationFn: () => 1e-7
// TensorFlow.js will use its default EPSILON value
// Dynamically import TensorFlow.js core module and backends
// Use type assertions to tell TypeScript these modules exist
this.tf = await import('@tensorflow/tfjs-core')
// Import CPU and WebGL backends
await import('@tensorflow/tfjs-backend-cpu')
try {
// Try to import and use WebGL backend first (GPU)
await import('@tensorflow/tfjs-backend-webgl')
// Check if WebGL is available and set it as the backend
if (
(await this.tf.findBackend('webgl')) ||
(await this.tf.ready().then(() => this.tf.findBackend('webgl')))
) {
console.log('Using WebGL backend (GPU acceleration)')
await this.tf.setBackend('webgl')
} else {
console.log('WebGL backend not available, falling back to CPU')
await this.tf.setBackend('cpu')
}
} catch (err) {
console.warn(
'WebGL backend failed to initialize, using CPU backend:',
err
)
await this.tf.setBackend('cpu')
}
// Dynamically import TensorFlow.js and Universal Sentence Encoder
// Use type assertions to tell TypeScript these modules exist
this.tf = await import('@tensorflow/tfjs')
this.use = await import('@tensorflow-models/universal-sentence-encoder')
// Load the model
@ -193,9 +203,13 @@ export function createTensorFlowEmbeddingFunction(): EmbeddingFunction {
/**
* Creates a TensorFlow-based Universal Sentence Encoder embedding function that runs in a separate thread
* This provides better performance for CPU-intensive embedding operations
* This provides better performance for embedding operations by:
* 1. Using GPU acceleration via WebGL when available
* 2. Running in a separate thread to avoid blocking the main thread
* 3. Falling back to CPU processing when GPU is not available
*
* @param options Configuration options
* @returns An embedding function that runs in a separate thread
* @returns An embedding function that runs in a separate thread with GPU acceleration when available
*/
export function createThreadedEmbeddingFunction(
options: { fallbackToMain?: boolean } = {}
@ -223,56 +237,113 @@ export function createThreadedEmbeddingFunction(
// Worker implementation function that will be stringified and run in the worker
const workerImplementation = async ({ data }: { data: any }) => {
// We need to dynamically import TensorFlow.js and USE in the worker
const tf = await import('@tensorflow/tfjs')
const use = await import('@tensorflow-models/universal-sentence-encoder')
try {
// We need to dynamically import TensorFlow.js core module and USE in the worker
// Use a variable name that won't conflict with any minified variables
// Load the model
const model = await use.load()
// TensorFlow.js will use its default EPSILON value
// Handle different input types
let textToEmbed: string[]
if (typeof data === 'string') {
if (data.trim() === '') {
return new Array(512).fill(0)
// Import TensorFlow.js modules
const tf = await import('@tensorflow/tfjs-core')
// Import CPU backend in the worker
await import('@tensorflow/tfjs-backend-cpu')
try {
// Try to import and use WebGL backend first (GPU)
await import('@tensorflow/tfjs-backend-webgl')
// Check if WebGL is available and set it as the backend
if (
(tf.findBackend && (await tf.findBackend('webgl'))) ||
(tf.ready &&
(await tf
.ready()
.then(() => tf.findBackend && tf.findBackend('webgl'))))
) {
console.log('Worker: Using WebGL backend (GPU acceleration)')
if (tf.setBackend) {
await tf.setBackend('webgl')
}
} else {
console.log(
'Worker: WebGL backend not available, falling back to CPU'
)
if (tf.setBackend) {
await tf.setBackend('cpu')
}
}
} catch (err) {
console.warn(
'Worker: WebGL backend failed to initialize, using CPU backend:',
err
)
if (tf.setBackend) {
await tf.setBackend('cpu')
}
}
textToEmbed = [data]
} else if (
Array.isArray(data) &&
data.every((item) => typeof item === 'string')
) {
if (data.length === 0 || data.every((item) => item.trim() === '')) {
return new Array(512).fill(0)
}
textToEmbed = data.filter((item) => item.trim() !== '')
if (textToEmbed.length === 0) {
return new Array(512).fill(0)
}
} else {
throw new Error(
'UniversalSentenceEncoder only supports string or string[] data'
// Import the Universal Sentence Encoder
const sentenceEncoderModule = await import(
'@tensorflow-models/universal-sentence-encoder'
)
// Load the model directly from the module
const model = await sentenceEncoderModule.load()
// Handle different input types
let textToEmbed: string[]
if (typeof data === 'string') {
if (data.trim() === '') {
return new Array(512).fill(0)
}
textToEmbed = [data]
} else if (
Array.isArray(data) &&
data.every((item) => typeof item === 'string')
) {
if (data.length === 0 || data.every((item) => item.trim() === '')) {
return new Array(512).fill(0)
}
textToEmbed = data.filter((item) => item.trim() !== '')
if (textToEmbed.length === 0) {
return new Array(512).fill(0)
}
} else {
throw new Error(
'UniversalSentenceEncoder only supports string or string[] data'
)
}
// Get embeddings
const embeddings = await model.embed(textToEmbed)
// Convert to array and return the first embedding
const embeddingArray = await embeddings.array()
// Dispose of the tensor to free memory
embeddings.dispose()
return embeddingArray[0]
} catch (error) {
console.error('Worker error:', error)
throw error
}
// Get embeddings
const embeddings = await model.embed(textToEmbed)
// Convert to array and return the first embedding
const embeddingArray = await embeddings.array()
// Dispose of the tensor to free memory
embeddings.dispose()
return embeddingArray[0]
}
// Execute the embedding function in a separate thread
// Pass the worker implementation as a string to avoid Promise cloning issues
return await executeInThread<Vector>(workerImplementation.toString(), embedInWorker(data))
return await executeInThread<Vector>(
workerImplementation.toString(),
embedInWorker(data)
)
} catch (error) {
// If threading fails and fallback is enabled, use the standard embedding function
if (options.fallbackToMain) {
console.warn('Threaded embedding failed, falling back to main thread:', error)
console.warn(
'Threaded embedding failed, falling back to main thread:',
error
)
useFallback = true
return standardEmbedding(data)
}
@ -287,7 +358,9 @@ export function createThreadedEmbeddingFunction(
* Default embedding function
* Uses UniversalSentenceEncoder for all text embeddings
* TensorFlow.js is required for this to work
* Uses threading when available for better performance
* Uses GPU acceleration via WebGL when available for optimal performance
* Uses threading when available to avoid blocking the main thread
* Falls back to CPU processing when GPU is not available
*/
export const defaultEmbeddingFunction: EmbeddingFunction =
createThreadedEmbeddingFunction({ fallbackToMain: true })

View file

@ -20,6 +20,18 @@ export function executeInWebWorker<T>(fnString: string, args: any): Promise<T> {
self.index = {}
self.index$1 = {}
self.index$2 = {}
self.universalSentenceEncoder_esm = {}
self.universalSentenceEncoder = {}
self.tfjs = {}
self.tfjs_core = {}
self.tfjs_backend_cpu = {}
self.tfjs_backend_webgl = {}
// Additional variables that might be generated by bundlers
self.use = {}
self.tf = {}
self.sentenceEncoder = {}
self.sentenceEncoderModule = {}
self.onmessage = async function(e) {
try {