Merge commit 'd803bfe2b1fe7f5e219e50ac20d6801a0a58ac75' as 'vendor/ruvector'
This commit is contained in:
46
vendor/ruvector/crates/ruvector-learning-wasm/src/lib.rs
vendored
Normal file
46
vendor/ruvector/crates/ruvector-learning-wasm/src/lib.rs
vendored
Normal file
@@ -0,0 +1,46 @@
|
||||
//! MicroLoRA WASM - Ultra-fast Low-Rank Adaptation for Edge AI
|
||||
//!
|
||||
//! This crate provides rank-2 LoRA (Low-Rank Adaptation) matrices optimized for
|
||||
//! WASM execution with <100us adaptation latency. Designed for real-time
|
||||
//! per-operator-type learning in query optimization systems.
|
||||
//!
|
||||
//! ## Key Features
|
||||
//!
|
||||
//! - **Rank-2 LoRA**: Minimal parameter count (2d parameters per adapter)
|
||||
//! - **Per-Operator Scoping**: Separate adapters for different operator types
|
||||
//! - **<100us Adaptation**: Instant weight updates for real-time learning
|
||||
//! - **WASM-Optimized**: no_std compatible, minimal allocations
|
||||
//!
|
||||
//! ## Architecture
|
||||
//!
|
||||
//! ```text
|
||||
//! Input Embedding (d-dim)
|
||||
//! |
|
||||
//! v
|
||||
//! +---------+
|
||||
//! | A: d x 2 | Down projection
|
||||
//! +---------+
|
||||
//! |
|
||||
//! v
|
||||
//! +---------+
|
||||
//! | B: 2 x d | Up projection
|
||||
//! +---------+
|
||||
//! |
|
||||
//! v
|
||||
//! Delta W = alpha * (A @ B)
|
||||
//! |
|
||||
//! v
|
||||
//! Output = Input + Delta W
|
||||
//! ```
|
||||
|
||||
mod lora;
|
||||
mod operator_scope;
|
||||
mod trajectory;
|
||||
|
||||
pub use lora::{LoRAConfig, LoRAPair, MicroLoRAEngine};
|
||||
pub use operator_scope::{OperatorScope, ScopedLoRA};
|
||||
pub use trajectory::{Trajectory, TrajectoryBuffer, TrajectoryStats};
|
||||
|
||||
// Re-export core types for JS
|
||||
pub use lora::wasm_exports::*;
|
||||
pub use operator_scope::wasm_exports::*;
|
||||
559
vendor/ruvector/crates/ruvector-learning-wasm/src/lora.rs
vendored
Normal file
559
vendor/ruvector/crates/ruvector-learning-wasm/src/lora.rs
vendored
Normal file
@@ -0,0 +1,559 @@
|
||||
//! MicroLoRA: Rank-2 Low-Rank Adaptation with <100us latency
|
||||
//!
|
||||
//! Implements the core LoRA algorithm: output = input + alpha * (input @ A @ B)
|
||||
//! where A: [d x 2] and B: [2 x d] for rank-2 adaptation.
|
||||
|
||||
use wasm_bindgen::prelude::*;
|
||||
|
||||
/// Configuration for MicroLoRA
|
||||
#[derive(Debug, Clone, Copy)]
|
||||
pub struct LoRAConfig {
|
||||
/// Embedding dimension (typically 256)
|
||||
pub dim: usize,
|
||||
/// LoRA rank (1-2 for micro, default 2)
|
||||
pub rank: usize,
|
||||
/// Scaling factor alpha (default 0.1)
|
||||
pub alpha: f32,
|
||||
/// Learning rate for adaptation (default 0.01)
|
||||
pub learning_rate: f32,
|
||||
/// Dropout rate (0.0 = no dropout)
|
||||
pub dropout: f32,
|
||||
}
|
||||
|
||||
impl Default for LoRAConfig {
|
||||
fn default() -> Self {
|
||||
Self {
|
||||
dim: 256,
|
||||
rank: 2,
|
||||
alpha: 0.1,
|
||||
learning_rate: 0.01,
|
||||
dropout: 0.0,
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/// A single LoRA adapter pair (A and B matrices)
|
||||
///
|
||||
/// For rank-2:
|
||||
/// - A: [dim x 2] - Down projection
|
||||
/// - B: [2 x dim] - Up projection (initialized to zero)
|
||||
///
|
||||
/// Forward: output = input + alpha * (input @ A @ B)
|
||||
#[derive(Clone)]
|
||||
pub struct LoRAPair {
|
||||
/// Down projection matrix A: [dim][rank]
|
||||
/// Stored as Vec<[f32; 2]> for rank-2
|
||||
a: Vec<[f32; 2]>,
|
||||
/// Up projection matrix B: [rank][dim]
|
||||
/// Stored as [[f32; 256]; 2] for fixed 256-dim embeddings
|
||||
b: [[f32; 256]; 2],
|
||||
/// Scaling factor
|
||||
alpha: f32,
|
||||
/// Learning rate
|
||||
lr: f32,
|
||||
/// Embedding dimension
|
||||
dim: usize,
|
||||
/// Adaptation count for statistics
|
||||
adapt_count: u64,
|
||||
}
|
||||
|
||||
impl LoRAPair {
|
||||
/// Create a new LoRA pair with Kaiming initialization for A, zeros for B
|
||||
pub fn new(config: &LoRAConfig) -> Self {
|
||||
let dim = config.dim.min(256); // Cap at 256 for fixed-size B
|
||||
let rank = config.rank.min(2); // Cap at 2 for micro
|
||||
|
||||
// Initialize A with small random values (Kaiming-like)
|
||||
// Using deterministic pseudo-random for reproducibility
|
||||
let mut a = Vec::with_capacity(dim);
|
||||
let scale = (2.0 / dim as f32).sqrt() * 0.1; // Small initialization
|
||||
|
||||
for i in 0..dim {
|
||||
let seed = i as u32;
|
||||
let r0 = pseudo_random(seed) * scale - scale / 2.0;
|
||||
let r1 = if rank > 1 {
|
||||
pseudo_random(seed.wrapping_add(1000)) * scale - scale / 2.0
|
||||
} else {
|
||||
0.0
|
||||
};
|
||||
a.push([r0, r1]);
|
||||
}
|
||||
|
||||
// B initialized to zeros (LoRA standard practice)
|
||||
let b = [[0.0f32; 256]; 2];
|
||||
|
||||
Self {
|
||||
a,
|
||||
b,
|
||||
alpha: config.alpha,
|
||||
lr: config.learning_rate,
|
||||
dim,
|
||||
adapt_count: 0,
|
||||
}
|
||||
}
|
||||
|
||||
/// Forward pass: output = input + alpha * (input @ A @ B)
|
||||
///
|
||||
/// Complexity: O(d * r + r * d) = O(2dr) for rank r
|
||||
/// For rank-2, d=256: ~1024 ops = <100us
|
||||
#[inline]
|
||||
pub fn forward(&self, input: &[f32]) -> Vec<f32> {
|
||||
let n = input.len().min(self.dim);
|
||||
let mut output = input.to_vec();
|
||||
|
||||
// Compute low_rank = input @ A (result: [2])
|
||||
let mut low_rank = [0.0f32; 2];
|
||||
for i in 0..n {
|
||||
low_rank[0] += input[i] * self.a[i][0];
|
||||
low_rank[1] += input[i] * self.a[i][1];
|
||||
}
|
||||
|
||||
// Compute delta = low_rank @ B (result: [dim])
|
||||
// Output = input + alpha * delta
|
||||
for i in 0..n {
|
||||
let delta = low_rank[0] * self.b[0][i] + low_rank[1] * self.b[1][i];
|
||||
output[i] += self.alpha * delta;
|
||||
}
|
||||
|
||||
output
|
||||
}
|
||||
|
||||
/// Forward pass into pre-allocated buffer (zero-allocation hot path)
|
||||
#[inline]
|
||||
pub fn forward_into(&self, input: &[f32], output: &mut [f32]) {
|
||||
let n = input.len().min(self.dim).min(output.len());
|
||||
|
||||
// Copy input to output
|
||||
output[..n].copy_from_slice(&input[..n]);
|
||||
|
||||
// Compute low_rank = input @ A
|
||||
let mut low_rank = [0.0f32; 2];
|
||||
for i in 0..n {
|
||||
low_rank[0] += input[i] * self.a[i][0];
|
||||
low_rank[1] += input[i] * self.a[i][1];
|
||||
}
|
||||
|
||||
// Add delta to output
|
||||
for i in 0..n {
|
||||
let delta = low_rank[0] * self.b[0][i] + low_rank[1] * self.b[1][i];
|
||||
output[i] += self.alpha * delta;
|
||||
}
|
||||
}
|
||||
|
||||
/// Adapt weights based on gradient signal
|
||||
///
|
||||
/// Uses rank-1 outer product update to B matrix for instant adaptation.
|
||||
/// Target latency: <100us
|
||||
#[inline]
|
||||
pub fn adapt(&mut self, gradient: &[f32]) {
|
||||
let n = gradient.len().min(self.dim);
|
||||
|
||||
// Compute gradient norm for normalization
|
||||
let mut grad_norm_sq = 0.0f32;
|
||||
for i in 0..n {
|
||||
grad_norm_sq += gradient[i] * gradient[i];
|
||||
}
|
||||
|
||||
if grad_norm_sq < 1e-16 {
|
||||
return; // Skip if gradient is too small
|
||||
}
|
||||
|
||||
let grad_norm = fast_sqrt(grad_norm_sq);
|
||||
let inv_norm = 1.0 / grad_norm;
|
||||
|
||||
// Compute column sums of A for scaling
|
||||
let mut a_col_sum = [0.0f32; 2];
|
||||
for i in 0..n {
|
||||
a_col_sum[0] += self.a[i][0];
|
||||
a_col_sum[1] += self.a[i][1];
|
||||
}
|
||||
|
||||
// Update B using outer product: B += lr * a_sum * normalized_grad^T
|
||||
for j in 0..n {
|
||||
let normalized_grad = gradient[j] * inv_norm;
|
||||
self.b[0][j] += self.lr * a_col_sum[0] * normalized_grad;
|
||||
self.b[1][j] += self.lr * a_col_sum[1] * normalized_grad;
|
||||
}
|
||||
|
||||
self.adapt_count += 1;
|
||||
}
|
||||
|
||||
/// Adapt with improvement signal (for reinforcement learning)
|
||||
///
|
||||
/// Uses the improvement ratio to scale the update magnitude.
|
||||
#[inline]
|
||||
pub fn adapt_with_reward(&mut self, gradient: &[f32], improvement: f32) {
|
||||
if improvement <= 0.0 {
|
||||
return; // Only learn from positive improvements
|
||||
}
|
||||
|
||||
let n = gradient.len().min(self.dim);
|
||||
|
||||
// Scale learning rate by improvement (clamped)
|
||||
let scaled_lr = self.lr * improvement.min(2.0);
|
||||
|
||||
// Compute gradient norm
|
||||
let mut grad_norm_sq = 0.0f32;
|
||||
for i in 0..n {
|
||||
grad_norm_sq += gradient[i] * gradient[i];
|
||||
}
|
||||
|
||||
if grad_norm_sq < 1e-16 {
|
||||
return;
|
||||
}
|
||||
|
||||
let inv_norm = 1.0 / fast_sqrt(grad_norm_sq);
|
||||
|
||||
// Compute A column sums
|
||||
let mut a_col_sum = [0.0f32; 2];
|
||||
for i in 0..n {
|
||||
a_col_sum[0] += self.a[i][0];
|
||||
a_col_sum[1] += self.a[i][1];
|
||||
}
|
||||
|
||||
// Update B
|
||||
for j in 0..n {
|
||||
let normalized_grad = gradient[j] * inv_norm;
|
||||
self.b[0][j] += scaled_lr * a_col_sum[0] * normalized_grad;
|
||||
self.b[1][j] += scaled_lr * a_col_sum[1] * normalized_grad;
|
||||
}
|
||||
|
||||
self.adapt_count += 1;
|
||||
}
|
||||
|
||||
/// Reset B matrix to zeros (fresh start)
|
||||
pub fn reset(&mut self) {
|
||||
for i in 0..256 {
|
||||
self.b[0][i] = 0.0;
|
||||
self.b[1][i] = 0.0;
|
||||
}
|
||||
self.adapt_count = 0;
|
||||
}
|
||||
|
||||
/// Get the number of adaptations performed
|
||||
pub fn adapt_count(&self) -> u64 {
|
||||
self.adapt_count
|
||||
}
|
||||
|
||||
/// Get the effective weight delta norm (for monitoring)
|
||||
pub fn delta_norm(&self) -> f32 {
|
||||
let mut norm_sq = 0.0f32;
|
||||
for i in 0..self.dim {
|
||||
let delta = self.b[0][i] * self.b[0][i] + self.b[1][i] * self.b[1][i];
|
||||
norm_sq += delta;
|
||||
}
|
||||
fast_sqrt(norm_sq) * self.alpha
|
||||
}
|
||||
|
||||
/// Get parameter count
|
||||
pub fn param_count(&self) -> usize {
|
||||
self.a.len() * 2 + 256 * 2
|
||||
}
|
||||
}
|
||||
|
||||
/// Main MicroLoRA engine managing multiple LoRA pairs
|
||||
pub struct MicroLoRAEngine {
|
||||
/// Default LoRA pair for unscoped operations
|
||||
default_lora: LoRAPair,
|
||||
/// Configuration (kept for potential future use)
|
||||
#[allow(dead_code)]
|
||||
config: LoRAConfig,
|
||||
/// Total forward passes
|
||||
forward_count: u64,
|
||||
/// Total adaptations
|
||||
total_adapt_count: u64,
|
||||
}
|
||||
|
||||
impl MicroLoRAEngine {
|
||||
/// Create a new MicroLoRA engine
|
||||
pub fn new(config: LoRAConfig) -> Self {
|
||||
Self {
|
||||
default_lora: LoRAPair::new(&config),
|
||||
config,
|
||||
forward_count: 0,
|
||||
total_adapt_count: 0,
|
||||
}
|
||||
}
|
||||
|
||||
/// Forward pass through the default LoRA
|
||||
#[inline]
|
||||
pub fn forward(&mut self, input: &[f32]) -> Vec<f32> {
|
||||
self.forward_count += 1;
|
||||
self.default_lora.forward(input)
|
||||
}
|
||||
|
||||
/// Adapt the default LoRA with gradient
|
||||
#[inline]
|
||||
pub fn adapt(&mut self, gradient: &[f32]) {
|
||||
self.default_lora.adapt(gradient);
|
||||
self.total_adapt_count += 1;
|
||||
}
|
||||
|
||||
/// Adapt with improvement reward
|
||||
#[inline]
|
||||
pub fn adapt_with_reward(&mut self, gradient: &[f32], improvement: f32) {
|
||||
self.default_lora.adapt_with_reward(gradient, improvement);
|
||||
self.total_adapt_count += 1;
|
||||
}
|
||||
|
||||
/// Reset the engine
|
||||
pub fn reset(&mut self) {
|
||||
self.default_lora.reset();
|
||||
self.forward_count = 0;
|
||||
self.total_adapt_count = 0;
|
||||
}
|
||||
|
||||
/// Get statistics
|
||||
pub fn stats(&self) -> (u64, u64, f32) {
|
||||
(
|
||||
self.forward_count,
|
||||
self.total_adapt_count,
|
||||
self.default_lora.delta_norm(),
|
||||
)
|
||||
}
|
||||
|
||||
/// Get the underlying LoRA pair for advanced use
|
||||
pub fn lora(&self) -> &LoRAPair {
|
||||
&self.default_lora
|
||||
}
|
||||
|
||||
/// Get mutable reference to underlying LoRA
|
||||
pub fn lora_mut(&mut self) -> &mut LoRAPair {
|
||||
&mut self.default_lora
|
||||
}
|
||||
}
|
||||
|
||||
impl Default for MicroLoRAEngine {
|
||||
fn default() -> Self {
|
||||
Self::new(LoRAConfig::default())
|
||||
}
|
||||
}
|
||||
|
||||
// ============ Helper Functions ============
|
||||
|
||||
/// Fast inverse square root (Quake III style)
|
||||
#[inline(always)]
|
||||
fn fast_sqrt(x: f32) -> f32 {
|
||||
if x <= 0.0 {
|
||||
return 0.0;
|
||||
}
|
||||
let i = 0x5f3759df - (x.to_bits() >> 1);
|
||||
let y = f32::from_bits(i);
|
||||
x * y * (1.5 - 0.5 * x * y * y)
|
||||
}
|
||||
|
||||
/// Deterministic pseudo-random number generator
|
||||
#[inline(always)]
|
||||
fn pseudo_random(seed: u32) -> f32 {
|
||||
// Simple xorshift
|
||||
let mut x = seed;
|
||||
x ^= x << 13;
|
||||
x ^= x >> 17;
|
||||
x ^= x << 5;
|
||||
(x as f32) / (u32::MAX as f32)
|
||||
}
|
||||
|
||||
// ============ WASM Bindings ============
|
||||
|
||||
pub mod wasm_exports {
|
||||
use super::*;
|
||||
#[allow(unused_imports)]
|
||||
use wasm_bindgen::prelude::*;
|
||||
|
||||
/// WASM-exposed MicroLoRA engine
|
||||
#[wasm_bindgen]
|
||||
pub struct WasmMicroLoRA {
|
||||
engine: MicroLoRAEngine,
|
||||
// Pre-allocated buffers for zero-allocation hot paths
|
||||
input_buffer: Vec<f32>,
|
||||
output_buffer: Vec<f32>,
|
||||
}
|
||||
|
||||
#[wasm_bindgen]
|
||||
impl WasmMicroLoRA {
|
||||
/// Create a new MicroLoRA engine
|
||||
///
|
||||
/// @param dim - Embedding dimension (default 256, max 256)
|
||||
/// @param alpha - Scaling factor (default 0.1)
|
||||
/// @param learning_rate - Learning rate (default 0.01)
|
||||
#[wasm_bindgen(constructor)]
|
||||
pub fn new(dim: Option<usize>, alpha: Option<f32>, learning_rate: Option<f32>) -> Self {
|
||||
let config = LoRAConfig {
|
||||
dim: dim.unwrap_or(256).min(256),
|
||||
rank: 2,
|
||||
alpha: alpha.unwrap_or(0.1),
|
||||
learning_rate: learning_rate.unwrap_or(0.01),
|
||||
dropout: 0.0,
|
||||
};
|
||||
|
||||
let actual_dim = config.dim;
|
||||
Self {
|
||||
engine: MicroLoRAEngine::new(config),
|
||||
input_buffer: vec![0.0; actual_dim],
|
||||
output_buffer: vec![0.0; actual_dim],
|
||||
}
|
||||
}
|
||||
|
||||
/// Get pointer to input buffer for direct memory access
|
||||
#[wasm_bindgen]
|
||||
pub fn get_input_ptr(&mut self) -> *mut f32 {
|
||||
self.input_buffer.as_mut_ptr()
|
||||
}
|
||||
|
||||
/// Get pointer to output buffer for direct memory access
|
||||
#[wasm_bindgen]
|
||||
pub fn get_output_ptr(&self) -> *const f32 {
|
||||
self.output_buffer.as_ptr()
|
||||
}
|
||||
|
||||
/// Get embedding dimension
|
||||
#[wasm_bindgen]
|
||||
pub fn dim(&self) -> usize {
|
||||
self.input_buffer.len()
|
||||
}
|
||||
|
||||
/// Forward pass using internal buffers (zero-allocation)
|
||||
///
|
||||
/// Write input to get_input_ptr(), call forward(), read from get_output_ptr()
|
||||
#[wasm_bindgen]
|
||||
pub fn forward(&mut self) {
|
||||
self.engine
|
||||
.default_lora
|
||||
.forward_into(&self.input_buffer, &mut self.output_buffer);
|
||||
self.engine.forward_count += 1;
|
||||
}
|
||||
|
||||
/// Forward pass with typed array input (allocates output)
|
||||
#[wasm_bindgen]
|
||||
pub fn forward_array(&mut self, input: &[f32]) -> Vec<f32> {
|
||||
self.engine.forward(input)
|
||||
}
|
||||
|
||||
/// Adapt using input buffer as gradient
|
||||
#[wasm_bindgen]
|
||||
pub fn adapt(&mut self) {
|
||||
self.engine.adapt(&self.input_buffer.clone());
|
||||
}
|
||||
|
||||
/// Adapt with typed array gradient
|
||||
#[wasm_bindgen]
|
||||
pub fn adapt_array(&mut self, gradient: &[f32]) {
|
||||
self.engine.adapt(gradient);
|
||||
}
|
||||
|
||||
/// Adapt with improvement reward using input buffer as gradient
|
||||
#[wasm_bindgen]
|
||||
pub fn adapt_with_reward(&mut self, improvement: f32) {
|
||||
self.engine
|
||||
.adapt_with_reward(&self.input_buffer.clone(), improvement);
|
||||
}
|
||||
|
||||
/// Reset the engine
|
||||
#[wasm_bindgen]
|
||||
pub fn reset(&mut self) {
|
||||
self.engine.reset();
|
||||
}
|
||||
|
||||
/// Get forward pass count
|
||||
#[wasm_bindgen]
|
||||
pub fn forward_count(&self) -> u64 {
|
||||
self.engine.forward_count
|
||||
}
|
||||
|
||||
/// Get adaptation count
|
||||
#[wasm_bindgen]
|
||||
pub fn adapt_count(&self) -> u64 {
|
||||
self.engine.total_adapt_count
|
||||
}
|
||||
|
||||
/// Get delta norm (weight change magnitude)
|
||||
#[wasm_bindgen]
|
||||
pub fn delta_norm(&self) -> f32 {
|
||||
self.engine.default_lora.delta_norm()
|
||||
}
|
||||
|
||||
/// Get parameter count
|
||||
#[wasm_bindgen]
|
||||
pub fn param_count(&self) -> usize {
|
||||
self.engine.default_lora.param_count()
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
#[cfg(test)]
|
||||
mod tests {
|
||||
use super::*;
|
||||
|
||||
#[test]
|
||||
fn test_lora_pair_creation() {
|
||||
let config = LoRAConfig::default();
|
||||
let lora = LoRAPair::new(&config);
|
||||
assert_eq!(lora.dim, 256);
|
||||
assert_eq!(lora.adapt_count, 0);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn test_lora_forward() {
|
||||
let config = LoRAConfig::default();
|
||||
let lora = LoRAPair::new(&config);
|
||||
|
||||
let input = vec![1.0; 256];
|
||||
let output = lora.forward(&input);
|
||||
|
||||
assert_eq!(output.len(), 256);
|
||||
// Initially B is zeros, so output should equal input
|
||||
for i in 0..256 {
|
||||
assert!((output[i] - input[i]).abs() < 1e-6);
|
||||
}
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn test_lora_adapt() {
|
||||
let config = LoRAConfig::default();
|
||||
let mut lora = LoRAPair::new(&config);
|
||||
|
||||
let gradient = vec![0.1; 256];
|
||||
lora.adapt(&gradient);
|
||||
|
||||
assert_eq!(lora.adapt_count, 1);
|
||||
assert!(lora.delta_norm() > 0.0);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn test_lora_forward_after_adapt() {
|
||||
let config = LoRAConfig::default();
|
||||
let mut lora = LoRAPair::new(&config);
|
||||
|
||||
// Adapt
|
||||
let gradient = vec![0.1; 256];
|
||||
lora.adapt(&gradient);
|
||||
|
||||
// Forward should now produce different output
|
||||
let input = vec![1.0; 256];
|
||||
let output = lora.forward(&input);
|
||||
|
||||
// Output should differ from input after adaptation
|
||||
let mut diff = 0.0f32;
|
||||
for i in 0..256 {
|
||||
diff += (output[i] - input[i]).abs();
|
||||
}
|
||||
assert!(
|
||||
diff > 0.0,
|
||||
"Output should differ from input after adaptation"
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn test_engine_stats() {
|
||||
let mut engine = MicroLoRAEngine::default();
|
||||
|
||||
let input = vec![1.0; 256];
|
||||
let _ = engine.forward(&input);
|
||||
engine.adapt(&input);
|
||||
|
||||
let (forwards, adapts, delta) = engine.stats();
|
||||
assert_eq!(forwards, 1);
|
||||
assert_eq!(adapts, 1);
|
||||
assert!(delta >= 0.0);
|
||||
}
|
||||
}
|
||||
523
vendor/ruvector/crates/ruvector-learning-wasm/src/operator_scope.rs
vendored
Normal file
523
vendor/ruvector/crates/ruvector-learning-wasm/src/operator_scope.rs
vendored
Normal file
@@ -0,0 +1,523 @@
|
||||
//! Per-Operator-Type Scoped LoRA
|
||||
//!
|
||||
//! Maintains separate LoRA adapters for different operator types,
|
||||
//! enabling specialized learning for each query operator.
|
||||
|
||||
use crate::lora::{LoRAConfig, LoRAPair};
|
||||
use wasm_bindgen::prelude::*;
|
||||
|
||||
/// Operator types for scoping (matches ruvector-dag OperatorType)
|
||||
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
|
||||
#[repr(u8)]
|
||||
pub enum OperatorScope {
|
||||
// Scan operators (0-3)
|
||||
SeqScan = 0,
|
||||
IndexScan = 1,
|
||||
HnswScan = 2,
|
||||
IvfFlatScan = 3,
|
||||
|
||||
// Join operators (4-6)
|
||||
NestedLoopJoin = 4,
|
||||
HashJoin = 5,
|
||||
MergeJoin = 6,
|
||||
|
||||
// Aggregation (7-8)
|
||||
Aggregate = 7,
|
||||
GroupBy = 8,
|
||||
|
||||
// Filter/Project (9-10)
|
||||
Filter = 9,
|
||||
Project = 10,
|
||||
|
||||
// Sort/Limit (11-12)
|
||||
Sort = 11,
|
||||
Limit = 12,
|
||||
|
||||
// Vector operations (13-14)
|
||||
VectorDistance = 13,
|
||||
Rerank = 14,
|
||||
|
||||
// Utility (15-16)
|
||||
Materialize = 15,
|
||||
Result = 16,
|
||||
}
|
||||
|
||||
impl OperatorScope {
|
||||
/// Convert from u8
|
||||
pub fn from_u8(v: u8) -> Option<Self> {
|
||||
match v {
|
||||
0 => Some(Self::SeqScan),
|
||||
1 => Some(Self::IndexScan),
|
||||
2 => Some(Self::HnswScan),
|
||||
3 => Some(Self::IvfFlatScan),
|
||||
4 => Some(Self::NestedLoopJoin),
|
||||
5 => Some(Self::HashJoin),
|
||||
6 => Some(Self::MergeJoin),
|
||||
7 => Some(Self::Aggregate),
|
||||
8 => Some(Self::GroupBy),
|
||||
9 => Some(Self::Filter),
|
||||
10 => Some(Self::Project),
|
||||
11 => Some(Self::Sort),
|
||||
12 => Some(Self::Limit),
|
||||
13 => Some(Self::VectorDistance),
|
||||
14 => Some(Self::Rerank),
|
||||
15 => Some(Self::Materialize),
|
||||
16 => Some(Self::Result),
|
||||
_ => None,
|
||||
}
|
||||
}
|
||||
|
||||
/// Get category for grouped learning
|
||||
pub fn category(&self) -> OperatorCategory {
|
||||
match self {
|
||||
Self::SeqScan | Self::IndexScan | Self::HnswScan | Self::IvfFlatScan => {
|
||||
OperatorCategory::Scan
|
||||
}
|
||||
Self::NestedLoopJoin | Self::HashJoin | Self::MergeJoin => OperatorCategory::Join,
|
||||
Self::Aggregate | Self::GroupBy => OperatorCategory::Aggregation,
|
||||
Self::Filter | Self::Project => OperatorCategory::Transform,
|
||||
Self::Sort | Self::Limit => OperatorCategory::Order,
|
||||
Self::VectorDistance | Self::Rerank => OperatorCategory::Vector,
|
||||
Self::Materialize | Self::Result => OperatorCategory::Utility,
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/// High-level operator categories for shared learning
|
||||
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
|
||||
#[repr(u8)]
|
||||
pub enum OperatorCategory {
|
||||
Scan = 0,
|
||||
Join = 1,
|
||||
Aggregation = 2,
|
||||
Transform = 3,
|
||||
Order = 4,
|
||||
Vector = 5,
|
||||
Utility = 6,
|
||||
}
|
||||
|
||||
/// Scoped LoRA manager with per-operator-type adapters
|
||||
///
|
||||
/// Maintains 17 separate LoRA pairs (one per OperatorScope) for
|
||||
/// specialized learning based on query operator type.
|
||||
pub struct ScopedLoRA {
|
||||
/// Per-operator-type LoRA pairs (17 total)
|
||||
adapters: [LoRAPair; 17],
|
||||
/// Per-category LoRA pairs for fallback (7 total)
|
||||
category_adapters: [LoRAPair; 7],
|
||||
/// Configuration (kept for potential future use)
|
||||
#[allow(dead_code)]
|
||||
config: LoRAConfig,
|
||||
/// Whether to use category fallback when operator has no history
|
||||
use_category_fallback: bool,
|
||||
/// Per-operator forward counts
|
||||
forward_counts: [u64; 17],
|
||||
}
|
||||
|
||||
impl ScopedLoRA {
|
||||
/// Create a new scoped LoRA manager
|
||||
pub fn new(config: LoRAConfig) -> Self {
|
||||
// Initialize all 17 operator adapters
|
||||
let adapters = std::array::from_fn(|_| LoRAPair::new(&config));
|
||||
let category_adapters = std::array::from_fn(|_| LoRAPair::new(&config));
|
||||
|
||||
Self {
|
||||
adapters,
|
||||
category_adapters,
|
||||
config,
|
||||
use_category_fallback: true,
|
||||
forward_counts: [0; 17],
|
||||
}
|
||||
}
|
||||
|
||||
/// Forward pass for a specific operator type
|
||||
#[inline]
|
||||
pub fn forward(&mut self, scope: OperatorScope, input: &[f32]) -> Vec<f32> {
|
||||
let idx = scope as usize;
|
||||
self.forward_counts[idx] += 1;
|
||||
|
||||
// Use operator-specific adapter
|
||||
let output = self.adapters[idx].forward(input);
|
||||
|
||||
// If using fallback and this operator has little history,
|
||||
// blend with category adapter
|
||||
if self.use_category_fallback && self.adapters[idx].adapt_count() < 10 {
|
||||
let cat_idx = scope.category() as usize;
|
||||
let cat_output = self.category_adapters[cat_idx].forward(input);
|
||||
|
||||
// Blend based on relative experience
|
||||
let op_exp = self.adapters[idx].adapt_count() as f32;
|
||||
let weight = (op_exp / 10.0).min(1.0);
|
||||
|
||||
let mut blended = output;
|
||||
for i in 0..blended.len().min(cat_output.len()) {
|
||||
blended[i] = blended[i] * weight + cat_output[i] * (1.0 - weight);
|
||||
}
|
||||
return blended;
|
||||
}
|
||||
|
||||
output
|
||||
}
|
||||
|
||||
/// Adapt the adapter for a specific operator type
|
||||
#[inline]
|
||||
pub fn adapt(&mut self, scope: OperatorScope, gradient: &[f32]) {
|
||||
let idx = scope as usize;
|
||||
self.adapters[idx].adapt(gradient);
|
||||
|
||||
// Also update category adapter for transfer learning
|
||||
let cat_idx = scope.category() as usize;
|
||||
self.category_adapters[cat_idx].adapt(gradient);
|
||||
}
|
||||
|
||||
/// Adapt with improvement reward
|
||||
#[inline]
|
||||
pub fn adapt_with_reward(&mut self, scope: OperatorScope, gradient: &[f32], improvement: f32) {
|
||||
let idx = scope as usize;
|
||||
self.adapters[idx].adapt_with_reward(gradient, improvement);
|
||||
|
||||
// Also update category adapter
|
||||
let cat_idx = scope.category() as usize;
|
||||
self.category_adapters[cat_idx].adapt_with_reward(gradient, improvement);
|
||||
}
|
||||
|
||||
/// Reset a specific operator's adapter
|
||||
pub fn reset_scope(&mut self, scope: OperatorScope) {
|
||||
let idx = scope as usize;
|
||||
self.adapters[idx].reset();
|
||||
self.forward_counts[idx] = 0;
|
||||
}
|
||||
|
||||
/// Reset all adapters
|
||||
pub fn reset_all(&mut self) {
|
||||
for adapter in &mut self.adapters {
|
||||
adapter.reset();
|
||||
}
|
||||
for adapter in &mut self.category_adapters {
|
||||
adapter.reset();
|
||||
}
|
||||
self.forward_counts = [0; 17];
|
||||
}
|
||||
|
||||
/// Get statistics for a specific operator
|
||||
pub fn stats(&self, scope: OperatorScope) -> (u64, u64, f32) {
|
||||
let idx = scope as usize;
|
||||
(
|
||||
self.forward_counts[idx],
|
||||
self.adapters[idx].adapt_count(),
|
||||
self.adapters[idx].delta_norm(),
|
||||
)
|
||||
}
|
||||
|
||||
/// Get total statistics across all operators
|
||||
pub fn total_stats(&self) -> (u64, u64, f32) {
|
||||
let total_forwards: u64 = self.forward_counts.iter().sum();
|
||||
let total_adapts: u64 = self.adapters.iter().map(|a| a.adapt_count()).sum();
|
||||
let max_delta: f32 = self
|
||||
.adapters
|
||||
.iter()
|
||||
.map(|a| a.delta_norm())
|
||||
.fold(0.0, f32::max);
|
||||
|
||||
(total_forwards, total_adapts, max_delta)
|
||||
}
|
||||
|
||||
/// Get the most active operator scopes
|
||||
pub fn most_active(&self, top_n: usize) -> Vec<(OperatorScope, u64)> {
|
||||
let mut counts: Vec<(usize, u64)> = self
|
||||
.forward_counts
|
||||
.iter()
|
||||
.enumerate()
|
||||
.map(|(i, &c)| (i, c))
|
||||
.collect();
|
||||
|
||||
counts.sort_by(|a, b| b.1.cmp(&a.1));
|
||||
|
||||
counts
|
||||
.into_iter()
|
||||
.take(top_n)
|
||||
.filter_map(|(idx, count)| {
|
||||
OperatorScope::from_u8(idx as u8).map(|scope| (scope, count))
|
||||
})
|
||||
.collect()
|
||||
}
|
||||
|
||||
/// Set category fallback mode
|
||||
pub fn set_category_fallback(&mut self, enabled: bool) {
|
||||
self.use_category_fallback = enabled;
|
||||
}
|
||||
|
||||
/// Get reference to operator adapter
|
||||
pub fn adapter(&self, scope: OperatorScope) -> &LoRAPair {
|
||||
&self.adapters[scope as usize]
|
||||
}
|
||||
|
||||
/// Get mutable reference to operator adapter
|
||||
pub fn adapter_mut(&mut self, scope: OperatorScope) -> &mut LoRAPair {
|
||||
&mut self.adapters[scope as usize]
|
||||
}
|
||||
}
|
||||
|
||||
impl Default for ScopedLoRA {
|
||||
fn default() -> Self {
|
||||
Self::new(LoRAConfig::default())
|
||||
}
|
||||
}
|
||||
|
||||
// ============ WASM Bindings ============
|
||||
|
||||
pub mod wasm_exports {
|
||||
use super::*;
|
||||
#[allow(unused_imports)]
|
||||
use wasm_bindgen::prelude::*;
|
||||
|
||||
/// WASM-exposed Scoped LoRA manager
|
||||
#[wasm_bindgen]
|
||||
pub struct WasmScopedLoRA {
|
||||
inner: ScopedLoRA,
|
||||
input_buffer: Vec<f32>,
|
||||
output_buffer: Vec<f32>,
|
||||
}
|
||||
|
||||
#[wasm_bindgen]
|
||||
impl WasmScopedLoRA {
|
||||
/// Create a new scoped LoRA manager
|
||||
///
|
||||
/// @param dim - Embedding dimension (max 256)
|
||||
/// @param alpha - Scaling factor (default 0.1)
|
||||
/// @param learning_rate - Learning rate (default 0.01)
|
||||
#[wasm_bindgen(constructor)]
|
||||
pub fn new(dim: Option<usize>, alpha: Option<f32>, learning_rate: Option<f32>) -> Self {
|
||||
let config = LoRAConfig {
|
||||
dim: dim.unwrap_or(256).min(256),
|
||||
rank: 2,
|
||||
alpha: alpha.unwrap_or(0.1),
|
||||
learning_rate: learning_rate.unwrap_or(0.01),
|
||||
dropout: 0.0,
|
||||
};
|
||||
|
||||
let actual_dim = config.dim;
|
||||
Self {
|
||||
inner: ScopedLoRA::new(config),
|
||||
input_buffer: vec![0.0; actual_dim],
|
||||
output_buffer: vec![0.0; actual_dim],
|
||||
}
|
||||
}
|
||||
|
||||
/// Get input buffer pointer
|
||||
#[wasm_bindgen]
|
||||
pub fn get_input_ptr(&mut self) -> *mut f32 {
|
||||
self.input_buffer.as_mut_ptr()
|
||||
}
|
||||
|
||||
/// Get output buffer pointer
|
||||
#[wasm_bindgen]
|
||||
pub fn get_output_ptr(&self) -> *const f32 {
|
||||
self.output_buffer.as_ptr()
|
||||
}
|
||||
|
||||
/// Forward pass for operator type (uses internal buffers)
|
||||
///
|
||||
/// @param op_type - Operator type (0-16)
|
||||
#[wasm_bindgen]
|
||||
pub fn forward(&mut self, op_type: u8) {
|
||||
if let Some(scope) = OperatorScope::from_u8(op_type) {
|
||||
let output = self.inner.forward(scope, &self.input_buffer);
|
||||
let n = output.len().min(self.output_buffer.len());
|
||||
self.output_buffer[..n].copy_from_slice(&output[..n]);
|
||||
}
|
||||
}
|
||||
|
||||
/// Forward pass with typed array
|
||||
#[wasm_bindgen]
|
||||
pub fn forward_array(&mut self, op_type: u8, input: &[f32]) -> Vec<f32> {
|
||||
if let Some(scope) = OperatorScope::from_u8(op_type) {
|
||||
self.inner.forward(scope, input)
|
||||
} else {
|
||||
input.to_vec()
|
||||
}
|
||||
}
|
||||
|
||||
/// Adapt for operator type using input buffer as gradient
|
||||
#[wasm_bindgen]
|
||||
pub fn adapt(&mut self, op_type: u8) {
|
||||
if let Some(scope) = OperatorScope::from_u8(op_type) {
|
||||
self.inner.adapt(scope, &self.input_buffer.clone());
|
||||
}
|
||||
}
|
||||
|
||||
/// Adapt with typed array
|
||||
#[wasm_bindgen]
|
||||
pub fn adapt_array(&mut self, op_type: u8, gradient: &[f32]) {
|
||||
if let Some(scope) = OperatorScope::from_u8(op_type) {
|
||||
self.inner.adapt(scope, gradient);
|
||||
}
|
||||
}
|
||||
|
||||
/// Adapt with improvement reward
|
||||
#[wasm_bindgen]
|
||||
pub fn adapt_with_reward(&mut self, op_type: u8, improvement: f32) {
|
||||
if let Some(scope) = OperatorScope::from_u8(op_type) {
|
||||
self.inner
|
||||
.adapt_with_reward(scope, &self.input_buffer.clone(), improvement);
|
||||
}
|
||||
}
|
||||
|
||||
/// Reset specific operator adapter
|
||||
#[wasm_bindgen]
|
||||
pub fn reset_scope(&mut self, op_type: u8) {
|
||||
if let Some(scope) = OperatorScope::from_u8(op_type) {
|
||||
self.inner.reset_scope(scope);
|
||||
}
|
||||
}
|
||||
|
||||
/// Reset all adapters
|
||||
#[wasm_bindgen]
|
||||
pub fn reset_all(&mut self) {
|
||||
self.inner.reset_all();
|
||||
}
|
||||
|
||||
/// Get forward count for operator
|
||||
#[wasm_bindgen]
|
||||
pub fn forward_count(&self, op_type: u8) -> u64 {
|
||||
if let Some(scope) = OperatorScope::from_u8(op_type) {
|
||||
self.inner.stats(scope).0
|
||||
} else {
|
||||
0
|
||||
}
|
||||
}
|
||||
|
||||
/// Get adapt count for operator
|
||||
#[wasm_bindgen]
|
||||
pub fn adapt_count(&self, op_type: u8) -> u64 {
|
||||
if let Some(scope) = OperatorScope::from_u8(op_type) {
|
||||
self.inner.stats(scope).1
|
||||
} else {
|
||||
0
|
||||
}
|
||||
}
|
||||
|
||||
/// Get delta norm for operator
|
||||
#[wasm_bindgen]
|
||||
pub fn delta_norm(&self, op_type: u8) -> f32 {
|
||||
if let Some(scope) = OperatorScope::from_u8(op_type) {
|
||||
self.inner.stats(scope).2
|
||||
} else {
|
||||
0.0
|
||||
}
|
||||
}
|
||||
|
||||
/// Get total forward count
|
||||
#[wasm_bindgen]
|
||||
pub fn total_forward_count(&self) -> u64 {
|
||||
self.inner.total_stats().0
|
||||
}
|
||||
|
||||
/// Get total adapt count
|
||||
#[wasm_bindgen]
|
||||
pub fn total_adapt_count(&self) -> u64 {
|
||||
self.inner.total_stats().1
|
||||
}
|
||||
|
||||
/// Enable/disable category fallback
|
||||
#[wasm_bindgen]
|
||||
pub fn set_category_fallback(&mut self, enabled: bool) {
|
||||
self.inner.set_category_fallback(enabled);
|
||||
}
|
||||
|
||||
/// Get operator scope name
|
||||
#[wasm_bindgen]
|
||||
pub fn scope_name(op_type: u8) -> String {
|
||||
match op_type {
|
||||
0 => "SeqScan".to_string(),
|
||||
1 => "IndexScan".to_string(),
|
||||
2 => "HnswScan".to_string(),
|
||||
3 => "IvfFlatScan".to_string(),
|
||||
4 => "NestedLoopJoin".to_string(),
|
||||
5 => "HashJoin".to_string(),
|
||||
6 => "MergeJoin".to_string(),
|
||||
7 => "Aggregate".to_string(),
|
||||
8 => "GroupBy".to_string(),
|
||||
9 => "Filter".to_string(),
|
||||
10 => "Project".to_string(),
|
||||
11 => "Sort".to_string(),
|
||||
12 => "Limit".to_string(),
|
||||
13 => "VectorDistance".to_string(),
|
||||
14 => "Rerank".to_string(),
|
||||
15 => "Materialize".to_string(),
|
||||
16 => "Result".to_string(),
|
||||
_ => "Unknown".to_string(),
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
#[cfg(test)]
|
||||
mod tests {
|
||||
use super::*;
|
||||
|
||||
#[test]
|
||||
fn test_scoped_lora_creation() {
|
||||
let lora = ScopedLoRA::default();
|
||||
let (forwards, adapts, delta) = lora.total_stats();
|
||||
assert_eq!(forwards, 0);
|
||||
assert_eq!(adapts, 0);
|
||||
assert_eq!(delta, 0.0);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn test_scoped_forward() {
|
||||
let mut lora = ScopedLoRA::default();
|
||||
let input = vec![1.0; 256];
|
||||
|
||||
let output = lora.forward(OperatorScope::HnswScan, &input);
|
||||
assert_eq!(output.len(), 256);
|
||||
|
||||
let (forwards, _, _) = lora.stats(OperatorScope::HnswScan);
|
||||
assert_eq!(forwards, 1);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn test_scoped_adapt() {
|
||||
let mut lora = ScopedLoRA::default();
|
||||
let gradient = vec![0.1; 256];
|
||||
|
||||
lora.adapt(OperatorScope::Filter, &gradient);
|
||||
|
||||
let (_, adapts, delta) = lora.stats(OperatorScope::Filter);
|
||||
assert_eq!(adapts, 1);
|
||||
assert!(delta > 0.0);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn test_category_transfer() {
|
||||
let mut lora = ScopedLoRA::default();
|
||||
let gradient = vec![0.1; 256];
|
||||
|
||||
// Adapt HnswScan (category: Scan)
|
||||
lora.adapt(OperatorScope::HnswScan, &gradient);
|
||||
|
||||
// SeqScan should benefit from category adapter via fallback
|
||||
let input = vec![1.0; 256];
|
||||
let output = lora.forward(OperatorScope::SeqScan, &input);
|
||||
|
||||
// With fallback enabled and SeqScan having no history,
|
||||
// it should use the category adapter which was updated
|
||||
// This is a behavioral test - output should differ from input
|
||||
let mut diff = 0.0f32;
|
||||
for i in 0..256 {
|
||||
diff += (output[i] - input[i]).abs();
|
||||
}
|
||||
// Due to category transfer, there should be some difference
|
||||
assert!(diff > 0.0, "Category transfer should affect output");
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn test_operator_scope_conversion() {
|
||||
for i in 0..=16u8 {
|
||||
let scope = OperatorScope::from_u8(i);
|
||||
assert!(scope.is_some(), "Scope {} should be valid", i);
|
||||
}
|
||||
assert!(OperatorScope::from_u8(17).is_none());
|
||||
}
|
||||
}
|
||||
428
vendor/ruvector/crates/ruvector-learning-wasm/src/trajectory.rs
vendored
Normal file
428
vendor/ruvector/crates/ruvector-learning-wasm/src/trajectory.rs
vendored
Normal file
@@ -0,0 +1,428 @@
|
||||
//! Trajectory tracking for reinforcement learning
|
||||
//!
|
||||
//! Records execution trajectories for post-hoc learning and pattern analysis.
|
||||
|
||||
use wasm_bindgen::prelude::*;
|
||||
|
||||
/// A single trajectory recording
|
||||
#[derive(Clone)]
|
||||
pub struct Trajectory {
|
||||
/// Embedding at query start
|
||||
pub embedding: Vec<f32>,
|
||||
/// Operator type that was executed (0-16)
|
||||
pub operator_type: u8,
|
||||
/// Attention mechanism used
|
||||
pub attention_type: u8,
|
||||
/// Execution time in milliseconds
|
||||
pub execution_ms: f32,
|
||||
/// Baseline execution time (for comparison)
|
||||
pub baseline_ms: f32,
|
||||
/// Improvement ratio (baseline / actual - 1.0)
|
||||
pub improvement: f32,
|
||||
/// Timestamp (simulation time or wall clock)
|
||||
pub timestamp: u64,
|
||||
}
|
||||
|
||||
impl Trajectory {
|
||||
/// Create a new trajectory
|
||||
pub fn new(
|
||||
embedding: Vec<f32>,
|
||||
operator_type: u8,
|
||||
attention_type: u8,
|
||||
execution_ms: f32,
|
||||
baseline_ms: f32,
|
||||
) -> Self {
|
||||
let improvement = if execution_ms > 0.0 {
|
||||
(baseline_ms / execution_ms) - 1.0
|
||||
} else {
|
||||
0.0
|
||||
};
|
||||
|
||||
Self {
|
||||
embedding,
|
||||
operator_type,
|
||||
attention_type,
|
||||
execution_ms,
|
||||
baseline_ms,
|
||||
improvement,
|
||||
timestamp: 0,
|
||||
}
|
||||
}
|
||||
|
||||
/// Get quality score (0.0 - 1.0)
|
||||
pub fn quality(&self) -> f32 {
|
||||
// Quality based on improvement, saturating at 2x speedup
|
||||
((self.improvement + 1.0) / 2.0).clamp(0.0, 1.0)
|
||||
}
|
||||
|
||||
/// Check if this trajectory represents a success
|
||||
pub fn is_success(&self) -> bool {
|
||||
self.improvement > 0.0
|
||||
}
|
||||
|
||||
/// Get the gradient direction for learning
|
||||
pub fn gradient(&self) -> Vec<f32> {
|
||||
if self.is_success() {
|
||||
// Positive improvement: reinforce this direction
|
||||
self.embedding.clone()
|
||||
} else {
|
||||
// Negative improvement: push away from this direction
|
||||
self.embedding.iter().map(|x| -x).collect()
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/// Statistics for a collection of trajectories
|
||||
#[derive(Clone, Default)]
|
||||
pub struct TrajectoryStats {
|
||||
/// Total trajectory count
|
||||
pub count: u64,
|
||||
/// Mean improvement ratio
|
||||
pub mean_improvement: f32,
|
||||
/// Variance of improvement
|
||||
pub variance: f32,
|
||||
/// Best improvement seen
|
||||
pub best_improvement: f32,
|
||||
/// Success rate (positive improvement)
|
||||
pub success_rate: f32,
|
||||
/// Most common attention type
|
||||
pub best_attention: u8,
|
||||
}
|
||||
|
||||
impl TrajectoryStats {
|
||||
/// Update stats with a new trajectory
|
||||
pub fn update(&mut self, trajectory: &Trajectory) {
|
||||
let n = self.count as f32;
|
||||
let new_n = n + 1.0;
|
||||
|
||||
// Welford's online algorithm for mean and variance
|
||||
let delta = trajectory.improvement - self.mean_improvement;
|
||||
self.mean_improvement += delta / new_n;
|
||||
let delta2 = trajectory.improvement - self.mean_improvement;
|
||||
self.variance += delta * delta2;
|
||||
|
||||
// Update best
|
||||
if trajectory.improvement > self.best_improvement {
|
||||
self.best_improvement = trajectory.improvement;
|
||||
self.best_attention = trajectory.attention_type;
|
||||
}
|
||||
|
||||
// Update success rate
|
||||
let successes = self.success_rate * n;
|
||||
let new_successes = if trajectory.is_success() {
|
||||
successes + 1.0
|
||||
} else {
|
||||
successes
|
||||
};
|
||||
self.success_rate = new_successes / new_n;
|
||||
|
||||
self.count += 1;
|
||||
}
|
||||
|
||||
/// Get variance (finalized)
|
||||
pub fn final_variance(&self) -> f32 {
|
||||
if self.count > 1 {
|
||||
self.variance / (self.count - 1) as f32
|
||||
} else {
|
||||
0.0
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/// Ring buffer for trajectory storage
|
||||
pub struct TrajectoryBuffer {
|
||||
/// Trajectories storage
|
||||
trajectories: Vec<Trajectory>,
|
||||
/// Maximum capacity
|
||||
capacity: usize,
|
||||
/// Write position
|
||||
write_pos: usize,
|
||||
/// Total count (may exceed capacity)
|
||||
total_count: u64,
|
||||
/// Running stats
|
||||
stats: TrajectoryStats,
|
||||
}
|
||||
|
||||
impl TrajectoryBuffer {
|
||||
/// Create a new trajectory buffer
|
||||
pub fn new(capacity: usize) -> Self {
|
||||
Self {
|
||||
trajectories: Vec::with_capacity(capacity),
|
||||
capacity,
|
||||
write_pos: 0,
|
||||
total_count: 0,
|
||||
stats: TrajectoryStats::default(),
|
||||
}
|
||||
}
|
||||
|
||||
/// Push a new trajectory
|
||||
pub fn push(&mut self, trajectory: Trajectory) {
|
||||
self.stats.update(&trajectory);
|
||||
|
||||
if self.trajectories.len() < self.capacity {
|
||||
self.trajectories.push(trajectory);
|
||||
} else {
|
||||
self.trajectories[self.write_pos] = trajectory;
|
||||
}
|
||||
|
||||
self.write_pos = (self.write_pos + 1) % self.capacity;
|
||||
self.total_count += 1;
|
||||
}
|
||||
|
||||
/// Get current buffer contents
|
||||
pub fn trajectories(&self) -> &[Trajectory] {
|
||||
&self.trajectories
|
||||
}
|
||||
|
||||
/// Drain all trajectories (returns ownership, clears buffer)
|
||||
pub fn drain(&mut self) -> Vec<Trajectory> {
|
||||
let result = std::mem::take(&mut self.trajectories);
|
||||
self.write_pos = 0;
|
||||
result
|
||||
}
|
||||
|
||||
/// Get statistics
|
||||
pub fn stats(&self) -> &TrajectoryStats {
|
||||
&self.stats
|
||||
}
|
||||
|
||||
/// Get total count (may exceed capacity)
|
||||
pub fn total_count(&self) -> u64 {
|
||||
self.total_count
|
||||
}
|
||||
|
||||
/// Get current buffer size
|
||||
pub fn len(&self) -> usize {
|
||||
self.trajectories.len()
|
||||
}
|
||||
|
||||
/// Check if empty
|
||||
pub fn is_empty(&self) -> bool {
|
||||
self.trajectories.is_empty()
|
||||
}
|
||||
|
||||
/// Get high-quality trajectories (quality > threshold)
|
||||
pub fn high_quality(&self, threshold: f32) -> Vec<&Trajectory> {
|
||||
self.trajectories
|
||||
.iter()
|
||||
.filter(|t| t.quality() > threshold)
|
||||
.collect()
|
||||
}
|
||||
|
||||
/// Get trajectories for a specific operator type
|
||||
pub fn by_operator(&self, op_type: u8) -> Vec<&Trajectory> {
|
||||
self.trajectories
|
||||
.iter()
|
||||
.filter(|t| t.operator_type == op_type)
|
||||
.collect()
|
||||
}
|
||||
|
||||
/// Reset buffer and stats
|
||||
pub fn reset(&mut self) {
|
||||
self.trajectories.clear();
|
||||
self.write_pos = 0;
|
||||
self.total_count = 0;
|
||||
self.stats = TrajectoryStats::default();
|
||||
}
|
||||
}
|
||||
|
||||
impl Default for TrajectoryBuffer {
|
||||
fn default() -> Self {
|
||||
Self::new(1000)
|
||||
}
|
||||
}
|
||||
|
||||
// ============ WASM Bindings ============
|
||||
|
||||
/// WASM-exposed trajectory buffer
|
||||
#[wasm_bindgen]
|
||||
pub struct WasmTrajectoryBuffer {
|
||||
buffer: TrajectoryBuffer,
|
||||
#[allow(dead_code)]
|
||||
embedding_dim: usize,
|
||||
}
|
||||
|
||||
#[wasm_bindgen]
|
||||
impl WasmTrajectoryBuffer {
|
||||
/// Create a new trajectory buffer
|
||||
///
|
||||
/// @param capacity - Maximum number of trajectories to store
|
||||
/// @param embedding_dim - Dimension of embeddings (default 256)
|
||||
#[wasm_bindgen(constructor)]
|
||||
pub fn new(capacity: Option<usize>, embedding_dim: Option<usize>) -> Self {
|
||||
Self {
|
||||
buffer: TrajectoryBuffer::new(capacity.unwrap_or(1000)),
|
||||
embedding_dim: embedding_dim.unwrap_or(256),
|
||||
}
|
||||
}
|
||||
|
||||
/// Record a trajectory
|
||||
///
|
||||
/// @param embedding - Embedding vector (Float32Array)
|
||||
/// @param op_type - Operator type (0-16)
|
||||
/// @param attention_type - Attention mechanism used
|
||||
/// @param execution_ms - Actual execution time
|
||||
/// @param baseline_ms - Baseline execution time
|
||||
#[wasm_bindgen]
|
||||
pub fn record(
|
||||
&mut self,
|
||||
embedding: &[f32],
|
||||
op_type: u8,
|
||||
attention_type: u8,
|
||||
execution_ms: f32,
|
||||
baseline_ms: f32,
|
||||
) {
|
||||
let traj = Trajectory::new(
|
||||
embedding.to_vec(),
|
||||
op_type,
|
||||
attention_type,
|
||||
execution_ms,
|
||||
baseline_ms,
|
||||
);
|
||||
self.buffer.push(traj);
|
||||
}
|
||||
|
||||
/// Get total count
|
||||
#[wasm_bindgen]
|
||||
pub fn total_count(&self) -> u64 {
|
||||
self.buffer.total_count()
|
||||
}
|
||||
|
||||
/// Get buffer length
|
||||
#[wasm_bindgen]
|
||||
pub fn len(&self) -> usize {
|
||||
self.buffer.len()
|
||||
}
|
||||
|
||||
/// Check if empty
|
||||
#[wasm_bindgen]
|
||||
pub fn is_empty(&self) -> bool {
|
||||
self.buffer.is_empty()
|
||||
}
|
||||
|
||||
/// Get mean improvement
|
||||
#[wasm_bindgen]
|
||||
pub fn mean_improvement(&self) -> f32 {
|
||||
self.buffer.stats().mean_improvement
|
||||
}
|
||||
|
||||
/// Get best improvement
|
||||
#[wasm_bindgen]
|
||||
pub fn best_improvement(&self) -> f32 {
|
||||
self.buffer.stats().best_improvement
|
||||
}
|
||||
|
||||
/// Get success rate
|
||||
#[wasm_bindgen]
|
||||
pub fn success_rate(&self) -> f32 {
|
||||
self.buffer.stats().success_rate
|
||||
}
|
||||
|
||||
/// Get best attention type
|
||||
#[wasm_bindgen]
|
||||
pub fn best_attention(&self) -> u8 {
|
||||
self.buffer.stats().best_attention
|
||||
}
|
||||
|
||||
/// Get variance
|
||||
#[wasm_bindgen]
|
||||
pub fn variance(&self) -> f32 {
|
||||
self.buffer.stats().final_variance()
|
||||
}
|
||||
|
||||
/// Reset buffer
|
||||
#[wasm_bindgen]
|
||||
pub fn reset(&mut self) {
|
||||
self.buffer.reset();
|
||||
}
|
||||
|
||||
/// Get high quality trajectory count
|
||||
#[wasm_bindgen]
|
||||
pub fn high_quality_count(&self, threshold: f32) -> usize {
|
||||
self.buffer.high_quality(threshold).len()
|
||||
}
|
||||
|
||||
/// Get trajectory count for operator
|
||||
#[wasm_bindgen]
|
||||
pub fn count_by_operator(&self, op_type: u8) -> usize {
|
||||
self.buffer.by_operator(op_type).len()
|
||||
}
|
||||
}
|
||||
|
||||
#[cfg(test)]
|
||||
mod tests {
|
||||
use super::*;
|
||||
|
||||
#[test]
|
||||
fn test_trajectory_creation() {
|
||||
let embedding = vec![1.0; 256];
|
||||
let traj = Trajectory::new(embedding, 2, 0, 100.0, 150.0);
|
||||
|
||||
assert_eq!(traj.operator_type, 2);
|
||||
assert!(traj.improvement > 0.0); // 150/100 - 1 = 0.5
|
||||
assert!(traj.is_success());
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn test_trajectory_quality() {
|
||||
let embedding = vec![1.0; 256];
|
||||
|
||||
// 2x speedup should give quality close to 1.0
|
||||
let fast = Trajectory::new(embedding.clone(), 0, 0, 50.0, 100.0);
|
||||
assert!(fast.quality() > 0.5);
|
||||
|
||||
// Slowdown should give lower quality
|
||||
let slow = Trajectory::new(embedding, 0, 0, 150.0, 100.0);
|
||||
assert!(slow.quality() < 0.5);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn test_buffer_push() {
|
||||
let mut buffer = TrajectoryBuffer::new(10);
|
||||
let embedding = vec![1.0; 256];
|
||||
|
||||
for i in 0..15 {
|
||||
let traj = Trajectory::new(embedding.clone(), 0, 0, 100.0, 100.0 + i as f32);
|
||||
buffer.push(traj);
|
||||
}
|
||||
|
||||
// Buffer should be at capacity
|
||||
assert_eq!(buffer.len(), 10);
|
||||
// Total count should include all pushes
|
||||
assert_eq!(buffer.total_count(), 15);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn test_stats_update() {
|
||||
let mut stats = TrajectoryStats::default();
|
||||
let embedding = vec![1.0; 256];
|
||||
|
||||
let traj1 = Trajectory::new(embedding.clone(), 0, 0, 100.0, 150.0); // 50% improvement
|
||||
let traj2 = Trajectory::new(embedding.clone(), 0, 1, 100.0, 200.0); // 100% improvement
|
||||
let traj3 = Trajectory::new(embedding, 0, 0, 150.0, 100.0); // -33% (failure)
|
||||
|
||||
stats.update(&traj1);
|
||||
stats.update(&traj2);
|
||||
stats.update(&traj3);
|
||||
|
||||
assert_eq!(stats.count, 3);
|
||||
assert!(stats.success_rate > 0.6); // 2/3 success
|
||||
assert_eq!(stats.best_attention, 1); // Best was attention type 1
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn test_high_quality_filter() {
|
||||
let mut buffer = TrajectoryBuffer::new(100);
|
||||
let embedding = vec![1.0; 256];
|
||||
|
||||
// Add some trajectories with varying quality
|
||||
for i in 0..10 {
|
||||
let baseline = 100.0 + (i as f32) * 20.0;
|
||||
let traj = Trajectory::new(embedding.clone(), 0, 0, 100.0, baseline);
|
||||
buffer.push(traj);
|
||||
}
|
||||
|
||||
let high_quality = buffer.high_quality(0.5);
|
||||
assert!(!high_quality.is_empty());
|
||||
}
|
||||
}
|
||||
Reference in New Issue
Block a user