MagicMindNet

MagicMindNet (pip install magicmindnet) is a beginner-friendly Python AI library — import magicmindnet as ai — backed by a from-scratch Rust core. Train transformer chatbots, classifiers, and toy diffusion models locally; load Hub / GGUF checkpoints; serve an OpenAI-compatible API.

pip install magicmindnet
# optional: Hugging Face / ModelScope / Diffusers backends
pip install "magicmindnet[hub]"

import magicmindnet as ai

data = ai.DatasetQA(data=[
    {"input": "hi", "output": "hello there!"},
    {"input": "how are you?", "output": "doing great, thanks!"},
])
bot = ai.Chatbot(vocab_size=512, n_layer=2, d_model=64, seed=42)
bot.train(data, epochs=5, verbose=True)   # prints per-epoch loss, returns loss history
print(bot.chat("hi"))

bot.save("my_bot.mmn")
bot = ai.load("my_bot.mmn")               # loads any MagicMindNet checkpoint

# Universal hub: HF / ModelScope / Ollama / local
model = ai.from_pretrained("./my_bot.mmn")
print(model.generate("hi", max_new_tokens=8))

CLI after install: mmn-eval smoke · mmn-serve --model my_bot.mmn

Under the hood: custom tensors and autograd, AdamW + Muon hybrid optimizers, KV-cache generation with modern sampling (top-p/top-k/min-p, penalties, Mirostat, typical_p), BPE/unigram tokenizers, RoPE/learned/sinusoidal PE, GQA, vision-prefix multimodal, RL + SPIN, strict checkpoint IO, and global format interop (GGUF / PyTorch / NumPy / TF / ONNX / …).

Docs: getting started · docs site (charts & benchmarks) · hub · feature parity · PyPI

---

Table of contents

1. Quick start
2. Architecture
3. Features at a glance
4. Installation & development
5. Python API overview
6. Checkpoints & strict IO
7. Training & examples
8. Testing & coverage
9. CUDA
10. Project layout
11. Documentation index
12. Milestones
13. License

---

Quick start

Requires: Python 3.12+ (wheels on PyPI). Building from source also needs Rust.

pip install magicmindnet
python -c "import magicmindnet as ai; print(ai.__version__)"
python examples/pip_quickstart.py   # if you cloned the repo
mmn-eval smoke

From source / contributors (editable + full gate):

python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
maturin develop --release
python examples/hello_ai.py
bash scripts/verify_gate.sh        # Windows: .\scripts\verify_gate.ps1

Everything is also available function-style for scripting pipelines:

data = ai.DatasetQA(file="qa.json", user_row="input", ai_row="output")
bot = ai.Chatbot(autoset="sub-100M")
cfg = ai.TrainConfig(epochs=1, batch_size=4, cuda=False, optimizer="hybrid")
losses = ai.Train(bot, data, cfg)          # same engine as bot.train(...)
ai.export(bot, "safetensors", "bot.mmn")

Learned position embeddings (opt-in; adds pos_embed to checkpoints):

bot = ai.Chatbot(vocab_size=512, n_layer=2, d_model=64, use_learned_pos_embed=True, max_seq_len=128)

See docs/position_encoding_coverage.md, examples/learned_pos_embed_roundtrip.py, and examples/rope_roundtrip.py.

Feature parity (llama.cpp sampling, Ollama stream/embed, ChatML, cosine LR): see docs/feature_parity.md.

Classification:

ds = ai.DatasetClassification(data=[
    {"text": "sunny warm", "label": "nice"},
    {"text": "rainy cold", "label": "gloomy"},
])
clf = ai.Classifier.from_classification(ds, input_dim=64, seed=42)
clf.train(ds, epochs=20)
print(clf.predict_label("sunny warm"))    # -> "nice"
clf.save("classifier.mmn")

---

Architecture

flowchart TB
    subgraph python [Python API - magicmindnet]
        Datasets[Datasets QA/Corpus/Classification/Image]
        Models[Chatbot / Classifier / Diffusion]
        Train[Train / TrainClassifier / RL / SPIN]
        IO[export / import / merge / quantize / limit]
    end
    subgraph rust [Rust workspace]
        Core[mmn-core tensor + autograd]
        NN[mmn-nn layers]
        ModelsR[mmn-models]
        TrainR[mmn-train]
        IO_R[mmn-io checkpoints]
        Optim[mmn-optim AdamW + Muon]
        Data[mmn-data + ChatXML]
        CUDA[mmn-cuda optional GEMM]
    end
    python -->|PyO3 mmn-py| rust
    Core --> NN --> ModelsR --> TrainR
    IO --> IO_R
    Train --> TrainR
    Train --> Optim
    Datasets --> Data

Crate	Role
mmn-core	Tensor, CPU ops, autograd tape, CE / embedding backward
mmn-nn	Linear, LayerNorm, GELU, transformer block pieces
mmn-models	Chatbot, Classifier, Diffusion (VAE + UNet training/sampling)
mmn-train	LM + classifier training loops, RL, SPIN, loss APIs
mmn-io	Safetensors JSON wrapper, merge, quantize, bin stub
mmn-optim	AdamW, Newton–Schulz Muon, hybrid optimizer, grad accumulation
mmn-data	Dataset loaders, ChatXML formatting
mmn-resource	limit() CPU/memory percent parsing
mmn-cuda	Optional CUDA GEMM parity path
mmn-py	PyO3 module magicmindnet

---

Features at a glance

Area	Capabilities
Datasets	DatasetQA, DatasetCorpus, DatasetClassification, DatasetImageGen, DatasetImageEdit; in-memory data=[...] lists; ChatXML think-tag split
Chatbot	bot.train / bot.chat / bot.save / Chatbot.load; autoset presets (sub-100M, sub-1B, sub-10B), vision flag, seed, shape getters
Classifier	from_classification, with_labels, clf.train, predict probs, predict_label
Training	Train, TrainClassifier, TrainDiffusion all return per-epoch loss history; verbose=True progress; batch accumulation; adamw / muon / hybrid optimizers (typos raise ValueError)
Generation	KV cache, top-k/top-p/min-p, repetition/frequency/presence penalties, stop strings, sliding context
RL / SPIN	Toy alignment loops on small models
IO	Universal ai.load(path) auto-detects model family + format; mmn-safetensors-v1, mmn-hf-safetensors-v1 (binary HF Chatbot), mmn-hf-classifier-v1, mmn-classifier-v1, mmn-bin-v1 stub; strict import
Global formats	Zero format libraries anywhere (safetensors container included — from scratch). GGUF: reads every current GGML tensor type, encodes classic quants byte-identical to the reference plus the complete k-quant family Q2_K–Q6_K + Q8_K (the reference Python package can't encode k-quants at all), cross-validated against llama.cpp's gguf package; embedded SentencePiece and gpt2 byte-BPE tokenizers. PyTorch zip + legacy .pt + sharded *.index.json. NumPy .npy/.npz with from-scratch DEFLATE both ways. TensorFlow: Keras .h5/.keras and checkpoint v2, read AND write — h5py and tf.train.load_checkpoint open our output. ONNX read + write (passes onnx.checker). Flax/JAX msgpack pytrees read + write (from-scratch MessagePack codec). Generic .safetensors arrays in every spec dtype. Official safetensors package opens our containers — docs/interop.md
Merge	Element-wise mean of all weights; vision OR; init_seed from first model
Quantize	int8 / int4 on chatbot + classifier weights
Diffusion	VAE encode/decode, UNet denoise training, inpainting, sampling, checkpoint IO
Errors	Typed Python exceptions: CUDAError, DataMismatchError, ModelMismatchError, …; constructor typos raise ValueError with the valid options
Typing	Ships py.typed + full .pyi stubs — IDE autocomplete for the whole API

Known gaps: see docs/limitations.md (external HF model import, full diffusion backward, etc.).

---

Installation & development

PyPI (recommended)

pip install magicmindnet
pip install "magicmindnet[hub]"   # transformers / diffusers / modelscope

Package metadata: version 0.2.1, console scripts mmn-serve and mmn-eval, typed stubs (py.typed).

Editable / from source

pip install -e ".[dev]"
maturin develop --release -m crates/mmn-py/Cargo.toml

# Linux/macOS gate scripts
bash scripts/ci_local.sh
bash scripts/verify_gate.sh

Windows:

.\scripts\ci_local.ps1
.\scripts\lint.ps1
.\scripts\count_tests.ps1
.\scripts\verify_gate.ps1

Pre-commit (optional): .pre-commit-config.yaml runs ruff on Python sources.

Product docs with charts: open docs/site/index.html in a browser.

---

Python API overview

Datasets

ai.DatasetQA(file="qa.json", user_row="input", ai_row="output")
ai.DatasetQA(data=[{"input": "hi", "output": "hello"}])       # no file needed
ai.DatasetCorpus(rowfile="rows.json", txtfile="corpus.txt")
ai.DatasetCorpus(data=["a chunk of text", "another chunk"])
ai.DatasetClassification(file="labels.json", text_col="text", tags_col="tag")
ai.DatasetClassification(data=[{"text": "yay", "label": "pos"}])
ai.DatasetImageGen("manifest.json")
ai.DatasetImageEdit("edit_manifest.json")

Models

bot = ai.Chatbot(vocab_size=..., n_layer=..., d_model=..., seed=..., autoset="sub-100M")
clf = ai.Classifier.from_classification(dataset, input_dim=..., seed=...)
diff = ai.Diffusion()

# Every model has the same four core methods:
model.train(dataset, epochs=..., learning_rate=..., verbose=True)  # -> list of epoch losses
model.save("model.mmn")
Model.load("model.mmn")     # or the universal ai.load("model.mmn")
bot.chat("hello")           # chatbot only; clf.predict_label(text) for classifiers

Training config

cfg = ai.TrainConfig(
    epochs=3,
    batch_size=8,
    learning_rate=0.001,
    cuda=False,
    optimizer="hybrid",  # "adamw" | "muon" | "hybrid" — anything else raises ValueError
    verbose=False,       # True prints per-epoch mean loss
)
losses = ai.Train(bot, dataset, cfg)             # returns per-epoch mean losses
losses = ai.TrainClassifier(clf, dataset, cfg)
losses = ai.TrainDiffusion(diff, dataset, cfg)
ai.RL(bot, dataset, cfg, reward_amount=1.0, punishment_amount=0.5)
ai.SPIN(bot, selfplay_epochs=2, dataset=dataset)

Checkpoint IO

model = ai.load(path)      # universal: Chatbot / Classifier / Diffusion, any format
                           # (.mmn / .safetensors / .gguf / .pt / .npz — detected by content)

ai.export(bot, "safetensors", path)
bot2 = ai.import_model("safetensors", [path])  # first path only
bot.save("bot.gguf", format="gguf")            # also "gguf-q8_0", "npz", "pt"
merged = ai.merge(bot_a, bot_b)
ai.quantize(bot, "int8")  # or "int4"
ai.export_classifier(clf, "safetensors", path)
clf2 = ai.import_classifier("safetensors", [path])
ai.merge_classifier(clf_a, clf_b)
ai.quantize_classifier(clf, "int4")
ai.limit("50%")  # resource cap helper

# Generic array IO — no numpy/torch/h5py required (but their arrays are accepted)
ai.save_npz("arrays.npz", {"w": [[1.0, 2.0]]})
ai.save_pt("arrays.pt", {"w": [[1.0, 2.0]]})   # torch.load-compatible
weights = ai.load_h5("model.weights.h5")       # HDF5 / Keras without h5py
info = ai.gguf_info("model.gguf")              # GGUF metadata, header-only read
tok = ai.load_gguf_tokenizer("model.gguf")     # embedded SentencePiece vocab

Full API reference: docs/API.md. Beginner tutorial: docs/getting_started.md.

---

Checkpoints & strict IO

MagicMindNet checkpoints: JSON mmn-safetensors-v1 / mmn-classifier-v1 (default) or binary mmn-hf-safetensors-v1 / mmn-hf-classifier-v1 (Hugging Face tooling compatible).

Strict import guarantees:

• Required meta: vocab_size, n_layer, d_model (chatbot); input_dim, non-empty labels (classifier).
• Every exported tensor key must be present; missing keys fail (no silent partial load).
• Tensor shapes must match meta-derived expectations (Linear layout [out, in]).
• Tampered shape tests keep element count equal to data length so validation reaches shape checks.

100% tensor-key coverage (missing / shape / merge / quantize for all 12 chatbot keys):

→ docs/checkpoint_coverage.md

Format details: docs/checkpoints.md.

---

Training & examples

Script	Purpose
examples/hello_ai.py	Start here: in-memory data, bot.train, bot.chat, save/ai.load
examples/quickstart.py	QA load, train, export
examples/benchmark_train.py	Mean loss before/after train
examples/rl_spin.py	RL + SPIN on fixture QA
examples/classification_benchmark.py	Classifier benchmark
examples/classification.py	Train classifier end-to-end
examples/eval_mean_loss.py	Mean QA / classification loss
examples/eval_harness.py	Unified eval suites (smoke / all / …)
examples/checkpoint_roundtrip.py	Chatbot export → import
examples/learned_pos_embed_roundtrip.py	Learned pos_embed export → import + loss parity
examples/rope_roundtrip.py	RoPE chatbot export → import + loss parity
examples/classifier_roundtrip.py	Classifier export → import
examples/train_glint2.py	Glint-2 exact training (looped, RMS, SwiGLU, coda, window, BPE)
examples/generate_glint2.py	Generate with a trained Glint-2 MagicMindNet checkpoint

Full catalog: examples/README.md.

Training notes: docs/training.md.

---

Testing & coverage

After pip install -e ".[dev]" and maturin develop --release:

Command	Purpose
cargo test --workspace	Rust unit tests
pytest -q	Python integration tests
pytest tests/test_io_checkpoint_matrix_py.py -q	Full chatbot IO contract matrix
.\scripts\ci_local.ps1	Full local CI gate
.\scripts\verify_gate.ps1	CI + test count sanity check

Current counts (run .\scripts\count_tests.ps1 after changes):

• Rust #[test]: 580
• pytest: 1088

Test area map: docs/testing.md.

Deep review artifacts: docs/reviews/.

Subagent for IO gap scans: .cursor/agents/magicmindnet-checkpoint-strict.md.

---

CUDA

Build with CUDA when toolkit is installed:

maturin develop --features cuda -m crates/mmn-py/Cargo.toml

Without CUDA, TrainConfig(cuda=True) raises CUDAError with fix guidance. CPU path uses reference GEMM in mmn-core.

---

Project layout

MagicMindNet/
├── crates/           # Rust workspace (mmn-core … mmn-py)
├── magicmindnet/     # Python package stub / re-exports
├── tests/            # pytest integration + IO matrix
├── examples/         # Runnable demos
├── docs/             # API, training, checkpoints, coverage, reviews
├── scripts/          # ci_local, verify_gate, count_tests, lint
├── .cursor/agents/   # Project subagents (CI, IO, train, …)
├── pyproject.toml
├── CONTRIBUTING.md
├── AGENTS.md
└── CHANGELOG.md

---

Documentation index

Doc	Contents
docs/site/index.html	Product docs site — install, charts, hub, parity
docs/getting_started.md	Beginner tutorial — install to first trained model
docs/API.md	Public Python surface
docs/feature_parity.md	PyTorch / llama.cpp / Ollama parity matrix
docs/hub.md	Universal from_pretrained routing
docs/benchmarks.md	Eval harness suites
docs/training.md	Losses, optimizers, batching
docs/training_coverage.md	Training regression matrix
docs/vision_coverage.md	Vision-flag chatbot IO/train path
docs/examples_coverage.md	Examples smoke matrix
docs/attention_coverage.md	Attention forward/train scope (alpha)
docs/layernorm_coverage.md	LayerNorm forward/train scope (alpha)
docs/nn_coverage.md	mmn-nn block unit tests
docs/quantize_coverage.md	int8/int4 quantize regression matrix
docs/dataset_coverage.md	Dataset loader regression matrix
docs/checkpoints.md	Formats, merge, quantize
docs/checkpoint_coverage.md	100% IO regression matrix
docs/limitations.md	Known gaps
docs/testing.md	How to run tests
CONTRIBUTING.md	PR / gate expectations
AGENTS.md	Agent workflow notes
CHANGELOG.md	Release history

---

Milestones

Milestone	Status
M0 Bootstrap	Done
M1 Tensor / autograd / CUDA hooks	Done
M2 AdamW + Muon	Done
M3 Datasets + ChatXML	Done
M4 Chatbot + Train	Done
M5 Classification + vision flag	Done
M6 IO / merge / quantize / limit	Done
M7 RL + SPIN	Done
M8–M9 Diffusion / latent pipeline	TrainDiffusion, inpainting, sample/export/merge/quantize
M10 Release hardening	CI + strict IO tests + coverage matrix

---

License

GPL-3.0-or-later — see LICENSE.