Sprimo/skills/domain-ml/SKILL.md

name, description, user-invocable

name	description	user-invocable
domain-ml	Use when building ML/AI apps in Rust. Keywords: machine learning, ML, AI, tensor, model, inference, neural network, deep learning, training, prediction, ndarray, tch-rs, burn, candle, 机器学习, 人工智能, 模型推理	false

Machine Learning Domain

Layer 3: Domain Constraints

Domain Constraints → Design Implications

Domain Rule	Design Constraint	Rust Implication
Large data	Efficient memory	Zero-copy, streaming
GPU acceleration	CUDA/Metal support	candle, tch-rs
Model portability	Standard formats	ONNX
Batch processing	Throughput over latency	Batched inference
Numerical precision	Float handling	ndarray, careful f32/f64
Reproducibility	Deterministic	Seeded random, versioning

---

Critical Constraints

Memory Efficiency

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

GPU Utilization

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

Model Portability

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

---

Trace Down ↓

From constraints to design (Layer 2):

"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 Case → Framework

Use Case	Recommended	Why
Inference only	tract (ONNX)	Lightweight, portable
Training + inference	candle, burn	Pure Rust, GPU
PyTorch models	tch-rs	Direct bindings
Data pipelines	polars	Fast, lazy eval

Key Crates

Purpose	Crate
Tensors	ndarray
ONNX inference	tract
ML framework	candle, burn
PyTorch bindings	tch-rs
Data processing	polars
Embeddings	fastembed

Design Patterns

Pattern	Purpose	Implementation
Model loading	Once, reuse	OnceLock<Model>
Batching	Throughput	Collect then process
Streaming	Large data	Iterator-based
GPU async	Parallelism	Data loading parallel to compute

Code Pattern: Inference Server

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())
}

Code Pattern: Batched Inference

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
}

---

Common Mistakes

Mistake	Domain Violation	Fix
Clone tensors	Memory waste	Use views
Single inference	GPU underutilized	Batch processing
Load model per request	Slow	Singleton pattern
Sync data loading	GPU idle	Async pipeline

---

Trace to Layer 1

Constraint	Layer 2 Pattern	Layer 1 Implementation
Memory efficiency	Zero-copy	ndarray views
Model singleton	Lazy init	OnceLock
Batch processing	Chunked iteration	chunks() + parallel
GPU async	Concurrent loading	tokio::spawn + GPU

---

Related Skills

When	See
Performance	m10-performance
Lazy initialization	m12-lifecycle
Async patterns	m07-concurrency
Memory efficiency	m01-ownership