Speedtronic
GPU-agnostic PyTorch training, built for speed.
Speedtronic 2.0.0 is an alpha, architecture-neutral Python training framework. A run is assembled from a validated configuration, a registered torch.nn.Module, a data source, AdamW or hybrid Muon, an optional scheduler, precision handling, checkpoints, metrics, and—optionally—a Hub-backed DumbDiLoCo coordinator.
Documentation scope
This site documents every local implementation module, every top-level class and function, every method, the v1 and v2 repository tests, both shipped configurations, both examples, and the packaging surface. The generated source inventory is rebuilt from the Python AST during the documentation build. Documentation coverage is not the same as runtime test coverage; see the test map.
Capability map
| Area | What Speedtronic 2.0.0 provides |
|---|---|
| Devices | CPU, CUDA, and Apple MPS selection |
| Model contract | Any module following Speedtronic's loss/batch protocol |
| Reference model | Decoder-only transformer with RoPE, GQA, SwiGLU, RMSNorm, and tied embeddings |
| Optimization | AdamW, hybrid Muon/Muon+, cautious updates, gradient accumulation, clipping, warmup, cosine or constant schedule |
| Precision | FP32, FP16, BF16, CUDA GradScaler, fused-AdamW probing |
| Performance | Optional torch.compile, opt-in CUDA stream backprop, gradient checkpointing, worker prefetch, pinned memory |
| Data | Deterministic synthetic data, streamed UTF-8 text, serialized datasets, programmatic datasets |
| Persistence | Atomic local checkpoints with model/optimizer/scheduler/RNG/counters |
| Distributed | Asynchronous DumbDiLoCo over a Hugging Face model repository |
| Observability | Text/JSONL metrics plus callable, W&B, and TensorBoard hook adapters |
One-minute path
- Install and run the CPU smoke test.
- Read core concepts.
- Choose a tutorial: custom model, custom data, performance, checkpointing, or observability.
- Use the configuration reference and complete source inventory while integrating.
System shape
Documentation map
Learn the system
- Architecture explains composition, training order, and state ownership.
- Runtime and Trainer documents the public training API and its contracts.
- Configuration lists every field, default, normalization rule, and precedence edge case.
Explore v2
- v2 overview
- Muon, Muon+, and cautious updates
- Out-of-order backprop scheduling
- Shape validation
- Migration and deferred decisions
Build an integration
Operate the framework
Project identity
- Package/distribution name:
speedtronic - Version:
2.0.0 - Python:
>=3.10 - PyTorch:
>=2.1 - License: Apache-2.0
- Default training device:
auto→ CUDA, then MPS, then CPU - Default checkpoint directory:
<run.output_dir>/checkpoints - Distributed transport: Hugging Face model repository files