Files
foxhunt/crates/ml-ensemble/src/inference_ensemble.rs
jgrusewski 22004a7368 refactor(cuda): eliminate candle from ml-core, ml-ppo, and 4 thin crates
Hard refactor — no shims, no compat layers. Candle removed from Cargo.toml
and all source files in 6 crates:

- ml-core: MlDevice enum, checkpoint.rs (safetensors direct), cudarc imports
  fixed from candle re-export to direct, AdamWConfig lr_decay, cuda_compat
  gutted. Net -7,341 lines.
- ml-ppo: All 16 files rewritten. LSTM→CudaLSTM, VarMap→GpuVarStore,
  PPOAgent 2306→700 lines, checkpoint→binary format.
- ml-ensemble: GPU-resident sigmoid via custom CUDA kernel.
- ml-explainability: Integrated gradients via GPU finite-difference kernels.
- ml-labeling: Device→MlDevice.
- ml-hyperopt: Cargo.toml only.

Remaining: ml-dqn (24 files), ml-supervised (4 files), ml crate (104 files).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-03-17 22:27:56 +01:00

685 lines
23 KiB
Rust

//! Lightweight synchronous ensemble coordinator
//!
//! Wraps N [`ModelInferenceAdapter`] instances and aggregates their
//! predictions via confidence-weighted voting. Models that are not
//! ready or that fail inference are gracefully skipped with warnings.
use std::collections::HashMap;
use rayon::prelude::*;
use crate::inference_adapter::{
EnsemblePrediction, FeatureVector, ModelInferenceAdapter, PredictionMeta, RawPrediction,
};
use crate::{MLError, MLResult};
/// Synchronous coordinator that aggregates predictions from multiple
/// [`ModelInferenceAdapter`] instances using confidence-weighted voting.
///
/// # Aggregation formula
///
/// ```text
/// weighted_direction = sum(direction_i * weight_i * confidence_i)
/// / sum(weight_i * confidence_i)
/// ```
///
/// where `weight_i` defaults to 1.0 and can be overridden via
/// [`set_weight`](InferenceEnsemble::set_weight).
#[allow(missing_debug_implementations)]
pub struct InferenceEnsemble {
adapters: Vec<Box<dyn ModelInferenceAdapter>>,
weights: HashMap<String, f64>,
}
impl InferenceEnsemble {
/// Create a new ensemble from a vector of adapters.
///
/// All adapters start with a default weight of 1.0.
pub fn new(adapters: Vec<Box<dyn ModelInferenceAdapter>>) -> Self {
Self {
adapters,
weights: HashMap::new(),
}
}
/// Set a custom weight for a model identified by name.
///
/// Models without an explicit weight use 1.0.
pub fn set_weight(&mut self, model_name: &str, weight: f64) {
self.weights.insert(model_name.to_string(), weight);
}
/// Return the number of adapters whose `is_ready()` returns true.
pub fn ready_count(&self) -> usize {
self.adapters.iter().filter(|a| a.is_ready()).count()
}
/// Run inference across all ready adapters and aggregate via
/// confidence-weighted voting.
///
/// Uses [`predict_raw()`](ModelInferenceAdapter::predict_raw) to collect
/// raw logits where available. Models that return logits are
/// aggregated via sigmoid -> weighted-sum. Models without logits
/// fall back to the original scalar-weighted-average path. The two
/// paths are merged by model-count-weighted average.
///
/// Returns [`MLError::InferenceError`] if no models are ready.
pub fn predict(&self, features: &FeatureVector) -> MLResult<EnsemblePrediction> {
let ready_adapters: Vec<&Box<dyn ModelInferenceAdapter>> =
self.adapters.iter().filter(|a| a.is_ready()).collect();
if ready_adapters.is_empty() {
return Err(MLError::InferenceError(
"No models are ready for inference".to_owned(),
));
}
// Run predict_raw() in parallel across all ready adapters
let raw_predictions: Vec<(String, RawPrediction)> = ready_adapters
.par_iter()
.filter_map(|adapter| {
let model_name = adapter.model_name().to_string();
match adapter.predict_raw(features) {
Ok(pred) => {
if !pred.direction_scalar.is_finite() || !pred.confidence.is_finite() {
tracing::warn!(
model = %model_name,
direction = %pred.direction_scalar,
confidence = %pred.confidence,
"Model returned NaN/Inf prediction, skipping (circuit breaker)"
);
None
} else {
Some((model_name, pred))
}
}
Err(e) => {
tracing::warn!(
model = %model_name,
error = %e,
"Model prediction failed, skipping"
);
None
}
}
})
.collect();
if raw_predictions.is_empty() {
return Err(MLError::InferenceError(
"All ready models failed during inference".to_owned(),
));
}
// Partition into logits-available vs CPU-scalar predictions
let (logits_preds, cpu_preds): (Vec<_>, Vec<_>) = raw_predictions
.into_iter()
.partition(|(_, p)| p.logits.is_some());
// Collect model names from both paths for the ensemble label
let mut model_names: Vec<String> = Vec::new();
let mut total_confidence_sum = 0.0_f64;
let mut total_count: usize = 0;
// --- Logits path: apply sigmoid, weighted-sum ---
let logits_result = if !logits_preds.is_empty() {
let result = self.aggregate_logits(&logits_preds);
for (name, pred) in &logits_preds {
model_names.push(name.clone());
total_confidence_sum += pred.confidence.clamp(0.0, 1.0);
}
total_count += logits_preds.len();
result
} else {
None
};
// --- CPU path: existing scalar weighted average ---
let cpu_result = if !cpu_preds.is_empty() {
let result = self.aggregate_cpu(&cpu_preds);
for (name, pred) in &cpu_preds {
model_names.push(name.clone());
total_confidence_sum += pred.confidence.clamp(0.0, 1.0);
}
total_count += cpu_preds.len();
result
} else {
None
};
// Merge logits and CPU directions by model-count-weighted average
let direction = match (logits_result, cpu_result) {
(Some((logits_dir, logits_n)), Some((cpu_dir, cpu_n))) => {
let total_n = (logits_n + cpu_n) as f64;
(logits_dir * logits_n as f64 + cpu_dir * cpu_n as f64) / total_n
}
(Some((dir, _)), None) => dir,
(None, Some((dir, _))) => dir,
(None, None) => 0.0,
};
let avg_confidence = if total_count > 0 {
total_confidence_sum / total_count as f64
} else {
0.0
};
let ensemble_name = format!("ENSEMBLE({})", model_names.join("+"));
Ok(EnsemblePrediction {
model_name: ensemble_name,
direction,
confidence: avg_confidence,
metadata: PredictionMeta::default(),
})
}
/// Aggregate predictions with logits: apply sigmoid to each logit,
/// compute confidence-weighted sum, remap to [-1,1] direction space.
fn aggregate_logits(
&self,
preds: &[(String, RawPrediction)],
) -> Option<(f64, usize)> {
if preds.is_empty() {
return None;
}
let mut weighted_direction_sum = 0.0_f64;
let mut weight_confidence_sum = 0.0_f64;
for (name, pred) in preds {
let logits = match pred.logits.as_ref() {
Some(l) if !l.is_empty() => l,
_ => continue,
};
// Apply sigmoid to logits and compute mean
let sigmoid_sum: f64 = logits
.iter()
.map(|&x| 1.0 / (1.0 + (-f64::from(x)).exp()))
.sum();
let sigmoid_mean = sigmoid_sum / logits.len() as f64;
// Remap [0,1] sigmoid output to [-1,1] direction space
let direction = sigmoid_mean * 2.0 - 1.0;
let confidence = pred.confidence.clamp(0.0, 1.0);
let w = self.weights.get(name).copied().unwrap_or(1.0);
let wc = w * confidence;
weighted_direction_sum += direction * wc;
weight_confidence_sum += wc;
}
let direction = if weight_confidence_sum.abs() < f64::EPSILON {
0.0
} else {
weighted_direction_sum / weight_confidence_sum
};
Some((direction, preds.len()))
}
/// Batched inference across all ready adapters: processes N feature vectors
/// through each model with a single GPU upload -> forward -> download per model,
/// then aggregates predictions per sample via confidence-weighted voting.
///
/// Returns one [`EnsemblePrediction`] per input feature vector.
pub fn predict_batch(&self, batch: &[FeatureVector]) -> MLResult<Vec<EnsemblePrediction>> {
if batch.is_empty() {
return Ok(Vec::new());
}
let ready_adapters: Vec<&Box<dyn ModelInferenceAdapter>> =
self.adapters.iter().filter(|a| a.is_ready()).collect();
if ready_adapters.is_empty() {
return Err(MLError::InferenceError(
"No models are ready for inference".to_owned(),
));
}
// Run predict_batch() in parallel across all ready adapters.
// Each adapter does a single batched GPU roundtrip internally.
let all_results: Vec<(String, Vec<EnsemblePrediction>)> = ready_adapters
.par_iter()
.filter_map(|adapter| {
let name = adapter.model_name().to_string();
match adapter.predict_batch(batch) {
Ok(preds) => Some((name, preds)),
Err(e) => {
tracing::warn!(
model = %name,
error = %e,
"Batched prediction failed, skipping model"
);
None
}
}
})
.collect();
if all_results.is_empty() {
return Err(MLError::InferenceError(
"All ready models failed during batched inference".to_owned(),
));
}
// Aggregate per-sample: for each index, collect predictions from all models
let n = batch.len();
let mut ensemble_preds = Vec::with_capacity(n);
for i in 0..n {
let mut weighted_dir_sum = 0.0_f64;
let mut weight_conf_sum = 0.0_f64;
let mut conf_sum = 0.0_f64;
let mut model_names = Vec::new();
let mut count: usize = 0;
for (name, preds) in &all_results {
if let Some(pred) = preds.get(i) {
if !pred.direction.is_finite() || !pred.confidence.is_finite() {
continue;
}
let conf = pred.confidence.clamp(0.0, 1.0);
let w = self.weights.get(name).copied().unwrap_or(1.0);
let wc = w * conf;
weighted_dir_sum += pred.direction * wc;
weight_conf_sum += wc;
conf_sum += conf;
model_names.push(name.as_str());
count += 1;
}
}
let direction = if weight_conf_sum.abs() < f64::EPSILON {
0.0
} else {
weighted_dir_sum / weight_conf_sum
};
let avg_confidence = if count > 0 {
conf_sum / count as f64
} else {
0.0
};
ensemble_preds.push(EnsemblePrediction {
model_name: format!("ENSEMBLE({})", model_names.join("+")),
direction,
confidence: avg_confidence,
metadata: PredictionMeta::default(),
});
}
Ok(ensemble_preds)
}
/// CPU-side scalar weighted average (original algorithm).
fn aggregate_cpu(
&self,
preds: &[(String, RawPrediction)],
) -> Option<(f64, usize)> {
if preds.is_empty() {
return None;
}
let mut weighted_direction_sum = 0.0_f64;
let mut weight_confidence_sum = 0.0_f64;
for (name, pred) in preds {
let confidence = pred.confidence.clamp(0.0, 1.0);
let w = self.weights.get(name).copied().unwrap_or(1.0);
let wc = w * confidence;
weighted_direction_sum += pred.direction_scalar * wc;
weight_confidence_sum += wc;
}
let direction = if weight_confidence_sum.abs() < f64::EPSILON {
0.0
} else {
weighted_direction_sum / weight_confidence_sum
};
Some((direction, preds.len()))
}
}
#[cfg(test)]
mod tests {
use super::*;
/// A simple test adapter with configurable direction, confidence, and readiness.
struct DummyAdapter {
name: String,
direction: f64,
confidence: f64,
ready: bool,
}
impl ModelInferenceAdapter for DummyAdapter {
fn model_name(&self) -> &str {
&self.name
}
fn predict(&self, _features: &FeatureVector) -> MLResult<EnsemblePrediction> {
Ok(EnsemblePrediction {
model_name: self.name.clone(),
direction: self.direction,
confidence: self.confidence,
metadata: PredictionMeta::default(),
})
}
fn is_ready(&self) -> bool {
self.ready
}
}
fn make_features() -> FeatureVector {
FeatureVector {
values: vec![0.0; 51],
timestamp: 1_700_000_000_000_000,
}
}
#[test]
fn test_ensemble_equal_weight_aggregation() {
// Model A: bullish (dir=1.0, conf=0.8)
// Model B: bearish (dir=-1.0, conf=0.6)
// weighted_direction = (1.0*1.0*0.8 + (-1.0)*1.0*0.6) / (1.0*0.8 + 1.0*0.6)
// = (0.8 - 0.6) / 1.4
// = 0.2 / 1.4
// ~ 0.1429
let adapters: Vec<Box<dyn ModelInferenceAdapter>> = vec![
Box::new(DummyAdapter {
name: "A".to_owned(),
direction: 1.0,
confidence: 0.8,
ready: true,
}),
Box::new(DummyAdapter {
name: "B".to_owned(),
direction: -1.0,
confidence: 0.6,
ready: true,
}),
];
let ensemble = InferenceEnsemble::new(adapters);
let features = make_features();
let pred = ensemble.predict(&features).expect("predict should succeed");
// Net direction should be positive (bullish model has higher confidence)
assert!(
pred.direction > 0.0,
"direction should be positive, got {}",
pred.direction
);
// But not too strong since the bearish model partially cancels
assert!(
pred.direction < 0.5,
"direction should be < 0.5, got {}",
pred.direction
);
// Confidence = average = (0.8 + 0.6) / 2 = 0.7
assert!(
(pred.confidence - 0.7).abs() < 1e-9,
"confidence should be 0.7, got {}",
pred.confidence
);
}
#[test]
fn test_ensemble_skips_unready_models() {
let adapters: Vec<Box<dyn ModelInferenceAdapter>> = vec![
Box::new(DummyAdapter {
name: "Ready".to_owned(),
direction: 1.0,
confidence: 0.9,
ready: true,
}),
Box::new(DummyAdapter {
name: "NotReady".to_owned(),
direction: -1.0,
confidence: 0.9,
ready: false,
}),
];
let ensemble = InferenceEnsemble::new(adapters);
assert_eq!(ensemble.ready_count(), 1);
let features = make_features();
let pred = ensemble.predict(&features).expect("predict should succeed");
// Only the ready (bullish) model contributes
assert!(
pred.direction > 0.5,
"direction should be > 0.5 with only bullish model, got {}",
pred.direction
);
// Model name should only include the ready model
assert!(
pred.model_name.contains("Ready"),
"model_name should contain 'Ready', got {}",
pred.model_name
);
assert!(
!pred.model_name.contains("NotReady"),
"model_name should NOT contain 'NotReady', got {}",
pred.model_name
);
}
#[test]
fn test_ensemble_custom_weights() {
// Model A: bullish, weight 0.75
// Model B: bearish, weight 0.25
// weighted_direction = (1.0*0.75*0.8 + (-1.0)*0.25*0.8) / (0.75*0.8 + 0.25*0.8)
// = (0.6 - 0.2) / (0.6 + 0.2)
// = 0.4 / 0.8
// = 0.5
let adapters: Vec<Box<dyn ModelInferenceAdapter>> = vec![
Box::new(DummyAdapter {
name: "Heavy".to_owned(),
direction: 1.0,
confidence: 0.8,
ready: true,
}),
Box::new(DummyAdapter {
name: "Light".to_owned(),
direction: -1.0,
confidence: 0.8,
ready: true,
}),
];
let mut ensemble = InferenceEnsemble::new(adapters);
ensemble.set_weight("Heavy", 0.75);
ensemble.set_weight("Light", 0.25);
let features = make_features();
let pred = ensemble.predict(&features).expect("predict should succeed");
// The heavier-weighted bullish model should dominate
assert!(
pred.direction > 0.0,
"direction should be positive (heavy bullish dominates), got {}",
pred.direction
);
// With equal confidences, the 3:1 weight ratio strongly favors bullish
// Expected: 0.5, but any positive value > 0.3 confirms dominance
assert!(
pred.direction > 0.3,
"direction should be > 0.3 showing heavy model dominance, got {}",
pred.direction
);
}
#[test]
fn test_ensemble_filters_nan_predictions() {
// NaN model should be skipped; the valid model's prediction stands alone.
let adapters: Vec<Box<dyn ModelInferenceAdapter>> = vec![
Box::new(DummyAdapter {
name: "Valid".to_owned(),
direction: 0.8,
confidence: 0.7,
ready: true,
}),
Box::new(DummyAdapter {
name: "NaN_Model".to_owned(),
direction: f64::NAN,
confidence: 0.9,
ready: true,
}),
];
let ensemble = InferenceEnsemble::new(adapters);
let features = make_features();
let pred = ensemble.predict(&features).expect("predict should succeed");
// Only the Valid model should contribute
assert!(
pred.model_name.contains("Valid"),
"model_name should contain 'Valid', got {}",
pred.model_name
);
assert!(
!pred.model_name.contains("NaN_Model"),
"model_name should NOT contain 'NaN_Model', got {}",
pred.model_name
);
// direction should come entirely from the Valid model
assert!(
(pred.direction - 0.8).abs() < 1e-9,
"direction should be 0.8, got {}",
pred.direction
);
// confidence should be from the single valid model
assert!(
(pred.confidence - 0.7).abs() < 1e-9,
"confidence should be 0.7, got {}",
pred.confidence
);
}
#[test]
fn test_predict_batch_aggregates_correctly() {
let adapters: Vec<Box<dyn ModelInferenceAdapter>> = vec![
Box::new(DummyAdapter {
name: "Bull".to_owned(),
direction: 1.0,
confidence: 0.8,
ready: true,
}),
Box::new(DummyAdapter {
name: "Bear".to_owned(),
direction: -1.0,
confidence: 0.6,
ready: true,
}),
];
let ensemble = InferenceEnsemble::new(adapters);
let batch = vec![make_features(), make_features(), make_features()];
let preds = ensemble.predict_batch(&batch).expect("batch should succeed");
assert_eq!(preds.len(), 3, "should produce one prediction per input");
for pred in &preds {
// Same adapters, same features -> same result for each sample
assert!(pred.direction > 0.0, "net direction should be bullish");
assert!(
pred.model_name.contains("Bull"),
"should include Bull model"
);
assert!(
pred.model_name.contains("Bear"),
"should include Bear model"
);
}
}
#[test]
fn test_predict_batch_empty_input() {
let adapters: Vec<Box<dyn ModelInferenceAdapter>> = vec![Box::new(DummyAdapter {
name: "A".to_owned(),
direction: 0.5,
confidence: 0.9,
ready: true,
})];
let ensemble = InferenceEnsemble::new(adapters);
let preds = ensemble.predict_batch(&[]).expect("empty batch should succeed");
assert!(preds.is_empty(), "empty input should produce empty output");
}
#[test]
fn test_predict_batch_skips_nan() {
let adapters: Vec<Box<dyn ModelInferenceAdapter>> = vec![
Box::new(DummyAdapter {
name: "Valid".to_owned(),
direction: 0.5,
confidence: 0.8,
ready: true,
}),
Box::new(DummyAdapter {
name: "NaN".to_owned(),
direction: f64::NAN,
confidence: 0.9,
ready: true,
}),
];
let ensemble = InferenceEnsemble::new(adapters);
let batch = vec![make_features()];
let preds = ensemble.predict_batch(&batch).expect("should succeed");
assert_eq!(preds.len(), 1);
assert!(
(preds[0].direction - 0.5).abs() < 1e-9,
"should use only valid model, got {}",
preds[0].direction
);
}
#[test]
fn test_ensemble_filters_extreme_confidence() {
// Model with confidence 5.0 should be clamped to 1.0
let adapters: Vec<Box<dyn ModelInferenceAdapter>> = vec![
Box::new(DummyAdapter {
name: "Normal".to_owned(),
direction: 1.0,
confidence: 0.6,
ready: true,
}),
Box::new(DummyAdapter {
name: "Extreme".to_owned(),
direction: 1.0,
confidence: 5.0,
ready: true,
}),
];
let ensemble = InferenceEnsemble::new(adapters);
let features = make_features();
let pred = ensemble.predict(&features).expect("predict should succeed");
// Both models should contribute (extreme confidence is clamped, not skipped)
assert!(
pred.model_name.contains("Normal"),
"model_name should contain 'Normal', got {}",
pred.model_name
);
assert!(
pred.model_name.contains("Extreme"),
"model_name should contain 'Extreme', got {}",
pred.model_name
);
// Confidence should be clamped: avg of 0.6 and 1.0 = 0.8
assert!(
(pred.confidence - 0.8).abs() < 1e-9,
"confidence should be 0.8 (avg of 0.6 and clamped 1.0), got {}",
pred.confidence
);
}
}