RULE: Avoid copying large tensors
WHY: Memory bandwidth is bottleneck
RUST: References, views, in-place ops

RULE: Batch operations for GPU efficiency
WHY: GPU overhead per kernel launch
RUST: Batch sizes, async data loading

RULE: Use standard model formats
WHY: Train in Python, deploy in Rust
RUST: ONNX via tract or candle

"Need efficient data pipelines"
    ↓ m10-performance: Streaming, batching
    ↓ polars: Lazy evaluation

"Need GPU inference"
    ↓ m07-concurrency: Async data loading
    ↓ candle/tch-rs: CUDA backend

"Need model loading"
    ↓ m12-lifecycle: Lazy init, caching
    ↓ tract: ONNX runtime

use std::sync::OnceLock;
use tract_onnx::prelude::*;

static MODEL: OnceLock<SimplePlan<TypedFact, Box<dyn TypedOp>, Graph<TypedFact, Box<dyn TypedOp>>>> = OnceLock::new();

fn get_model() -> &'static SimplePlan<...> {
    MODEL.get_or_init(|| {
        tract_onnx::onnx()
            .model_for_path("model.onnx")
            .unwrap()
            .into_optimized()
            .unwrap()
            .into_runnable()
            .unwrap()
    })
}

async fn predict(input: Vec<f32>) -> anyhow::Result<Vec<f32>> {
    let model = get_model();
    let input = tract_ndarray::arr1(&input).into_shape((1, input.len()))?;
    let result = model.run(tvec!(input.into()))?;
    Ok(result[0].to_array_view::<f32>()?.iter().copied().collect())
}

async fn batch_predict(inputs: Vec<Vec<f32>>, batch_size: usize) -> Vec<Vec<f32>> {
    let mut results = Vec::with_capacity(inputs.len());

    for batch in inputs.chunks(batch_size) {
        // Stack inputs into batch tensor
        let batch_tensor = stack_inputs(batch);

        // Run inference on batch
        let batch_output = model.run(batch_tensor).await;

        // Unstack results
        results.extend(unstack_outputs(batch_output));
    }

    results
}