cargo install aprender
apr pull qwen2.5-coder-1.5b
apr run qwen2.5-coder-1.5b "What is 2+2?"Aprender is a complete ML framework built from scratch in Rust. One cargo install,
one apr binary, 57 commands covering the full ML lifecycle:
| Stage | Commands | What it does |
|---|---|---|
| Inference | apr run, apr chat, apr serve |
Run models locally (GGUF, SafeTensors, APR) |
| Training | apr finetune, apr train, apr distill |
LoRA/QLoRA fine-tuning, knowledge distillation |
| Model Ops | apr convert, apr quantize, apr merge, apr export |
Format conversion, quantization, model merging |
| Inspection | apr inspect, apr validate, apr tensors, apr diff |
Model debugging, validation, comparison |
| Profiling | apr profile, apr bench, apr qa |
Roofline analysis, benchmarks, QA gates |
| Registry | apr pull, apr list, apr rm, apr publish |
HuggingFace Hub integration, model cache |
| GPU | apr gpu, apr parity, apr ptx |
GPU status, CPU/GPU parity checks, PTX analysis |
| Monitoring | apr tui, apr monitor, apr cbtop |
Terminal UI, training monitor, ComputeBrick pipeline |
- 70 workspace crates (was 20 separate repos)
- 25,391 tests, all passing
- 405 provable contracts (equation-based verification)
- 57 CLI commands with contract coverage
- 0
[patch.crates-io]— clean workspace deps
# Install the `apr` binary
cargo install aprender
# Verify
apr --version# Run inference
apr run hf://Qwen/Qwen2.5-Coder-1.5B-Instruct-GGUF "Explain quicksort"
apr chat hf://meta-llama/Llama-3-8B-Instruct-GGUF
# Serve model as API
apr serve model.gguf --port 8080
# Inspect model
apr inspect model.gguf
apr validate model.apr --quality --strict
apr tensors model.gguf | head -20
# Fine-tune with LoRA
apr finetune model.gguf --adapter lora --rank 64 --data train.jsonl
# Convert formats
apr convert model.safetensors --quantize q4_k -o model.gguf
apr export model.apr --format gguf -o model.gguf
# Profile
apr profile model.gguf --roofline
apr bench model.gguf --assert-tps 100The ML library is available as aprender on crates.io:
[dependencies]
aprender-core = "0.29"use aprender::linear_regression::LinearRegression;
use aprender::traits::Estimator;
let model = LinearRegression::new();
model.fit(&x_train, &y_train)?;
let predictions = model.predict(&x_test)?;Algorithms: Linear/Logistic Regression, Decision Trees, Random Forest, GBM, Naive Bayes, KNN, SVM, K-Means, PCA, ARIMA, ICA, GLMs, Graph algorithms, Bayesian inference, text processing, audio processing.
Monorepo with 70 crates in flat crates/aprender-* layout
(same pattern as Polars,
Burn,
Nushell):
paiml/aprender/
├── Cargo.toml # Workspace root + `cargo install aprender`
├── crates/
│ ├── aprender-core/ # ML library (use aprender::*)
│ ├── apr-cli/ # CLI logic (57 commands)
│ ├── aprender-compute/ # SIMD/GPU compute (was: trueno)
│ ├── aprender-gpu/ # CUDA PTX kernels (was: trueno-gpu)
│ ├── aprender-serve/ # Inference server (was: realizar)
│ ├── aprender-train/ # Training loops (was: entrenar)
│ ├── aprender-orchestrate/ # Agents, RAG (was: batuta)
│ ├── aprender-contracts/ # Provable contracts (was: provable-contracts)
│ ├── aprender-profile/ # Profiling (was: renacer)
│ ├── aprender-present-*/ # TUI framework (was: presentar)
│ ├── aprender-db/ # Embedded analytics DB
│ ├── aprender-graph/ # Graph database
│ ├── aprender-rag/ # RAG pipeline
│ └── ... (70 crates total)
├── contracts/ # 405 provable YAML contracts
└── book/ # mdBook documentation
| Model | Format | Speed | Hardware |
|---|---|---|---|
| Qwen2.5-Coder 1.5B | Q4_K | 40+ tok/s | CPU (AVX2) |
| Qwen2.5-Coder 7B | Q4_K | 225+ tok/s | RTX 4090 |
| TinyLlama 1.1B | Q4_0 | 17 tok/s | CPU (APR format) |
Benchmarked against real inference engines on Qwen2.5-Coder 7B Q4_K (RTX 4090). Data from candle-vs-apr proof-of-concept:
| Engine | tok/s | vs Candle | Architecture |
|---|---|---|---|
| llama.cpp b7746 | 443.6 | 1.95x | C++, Flash Attention |
| aprender (realizr) | 369.9 | 1.63x | Rust, CUDA graph + Flash Decoding |
| Candle | 227.4 | 1.00x | Rust, per-op dispatch |
| Concurrency | Agg tok/s | Scaling | Method |
|---|---|---|---|
| 1 | 367 | 1.0x | Single request |
| 8 | 954 | 2.6x | Continuous batching |
| 32 | 3,220 | 8.8x | Orca-style iteration scheduling |
| aprender advantage | Candle limitation | Impact |
|---|---|---|
| CUDA graph (647 kernels, 1 launch) | Per-op dispatch (~640 launches) | +26% |
| Flash Decoding (chunked KV) | Standard SDPA | +15% long ctx |
| Fused DP4A GEMV (4-bit native) | Separate dequant + matmul | ~10% |
| Continuous batching server | CLI only, no server | 8.8x at c=32 |
From ground-truth-apr-ludwig (21 recipes, Popperian falsification methodology):
| Capability | aprender | Ludwig | Notes |
|---|---|---|---|
| Classification (Iris, Wine) | apr finetune |
ludwig train |
Both achieve >95% accuracy |
| LoRA fine-tuning | apr finetune --lora |
Not native | apr: rank-64 in minutes |
| Quantization (INT8/INT4) | apr quantize |
Not supported | apr-native capability |
| Model merging | apr merge --strategy ties |
Not supported | TIES/DARE/SLERP |
| Provable contracts | 405 YAML contracts | None | Equation-based verification |
| Single binary | cargo install aprender |
pip install ludwig |
Rust vs Python |
All benchmarks reproducible from linked repos with cargo test.
Every CLI command and kernel has a provable contract (contracts/*.yaml)
with equations, preconditions, postconditions, and falsification tests:
equations:
validate_exit_code:
formula: exit_code = if score < 50 then 5 else 0
invariants:
- score < 50 implies exit_code != 0
falsification_tests:
- id: FALSIFY-CLI-001
prediction: apr validate bad-model.apr exits non-zero405 contracts across inference, training, quantization, attention, FFN, tokenization, model formats, and CLI safety.
All old crate names still work via backward-compatible shim crates:
| Old | New | Status |
|---|---|---|
trueno = "0.18" |
aprender-compute = "0.29" |
Shim available |
entrenar = "0.7" |
aprender-train = "0.29" |
Shim available |
realizar = "0.8" |
aprender-serve = "0.29" |
Shim available |
batuta = "0.7" |
aprender-orchestrate = "0.29" |
Shim available |
Old repositories are archived and read-only. All development happens here.
git clone https://github.com/paiml/aprender
cd aprender
cargo test --workspace --lib # 25,391 tests
cargo check --workspace # 70 crates
apr --help # 57 commandsMIT