Release History

All vectorphile releases — changelog, highlights, and breaking changes.

v0.3.0

Latest

November 2026

Production completeness — 41 modules, 50+ DL architectures, 5-tier GPU dispatch, full tokenizer suite, pretrained hub, RL algorithms, GGUF export, Lightning-style Trainer, and advanced optimizers. Replaces PyTorch + HuggingFace Transformers + scikit-learn under a single original API.

Highlights

→vp.backend — 5-tier GPU dispatch: Triton > PyTorch CUDA > Numba CUDA > im2col+BLAS > NumPy
→vp.nn.architectures — CNN zoo (LeNet5, VGG/11-19, ResNet/18-152, DenseNet, MobileNetV2, EfficientNet B0-B7)
→vp.nn.architectures — SSM: S4 (HiPPO-LegT + bilinear ZOH + FFT), Mamba (selective scan)
→vp.nn.architectures — ViT (tiny/small/base/large/huge), DeiT (distillation token, dual heads)
→vp.nn.architectures — Generative: VAE (ELBO), DCGAN (gradient penalty), DDPM (linear/cosine schedule)
→vp.nn.architectures — UNet, ResUNet, AttentionUNet (gated attention gates)
→vp.nn.architectures — GCN, GAT (multi-head), GraphSAGE (mean/max), GIN (WL-test expressive)
→vp.nn.flash_attention — O(N) memory tiled attention, causal masking
→vp.nn.qat — FakeQuantize, QATLinear/Conv2d, prepare_qat, convert_qat (straight-through estimator)
→vp.nn.optimizers_advanced — Lion, Sophia, Adan, Prodigy, ScheduleFreeAdamW
→vp.nn.lr_schedulers — 10 LR schedulers including WarmupCosineDecay, get_scheduler()
→vp.nn.einops — rearrange, reduce, repeat, pack, unpack, EinMix
→vp.nn.hooks — FeatureExtractor, GradientRecorder, register_forward_hook, register_backward_hook
→vp.tokenizers — BytePairEncoding, WordPieceTokenizer, SentencePieceTokenizer (all from scratch)
→vp.text — beam_search (n-gram blocking), nucleus_sampling, speculative_decoding, TextGenerationPipeline
→vp.vision — Compose, ColorJitter, AugMix, RandomResizedCrop, extract_features, make_grid
→vp.pretrained — load(), list_models() for 20+ models, safetensors + .bin loader, HF Hub compatible
→vp.trainer — Lightning-style Trainer, TrainableModel, automatic device placement, mixed precision
→vp.callbacks — EarlyStopping, ModelCheckpoint, WandBLogger, TensorBoardLogger, CSVLogger, LRFinder
→vp.distributed — DataParallel, DDP, GradientAccumulator, MixedPrecisionTrainer, launch()
→vp.metrics — accuracy, F1, AUC, BLEU, ROUGE, WER, perplexity, NDCG, ECE (all from scratch)
→vp.rl — DQN/DoubleDQN/DuelingDQN, REINFORCE, A2C, PPO, SAC, TD3, PrioritizedReplayBuffer
→vp.export — to_onnx, ONNXRuntime, GGUFReader/Writer, to_gguf/from_gguf (Q4_0/Q8_0)
→vp.data.datasets — MNIST, CIFAR10, CIFAR100, TinyShakespeare, IMDb, WeightedRandomSampler
→vp.autograd — jacobian, hessian, vmap, gradient checkpointing
→vp.cuda — Numba CUDA kernels (tiled matmul, conv, softmax, layernorm) + Triton fused kernels

Breaking changes:None — all v0.2.0 public API preserved.

v0.2.0

Archived

November 2025

View Docs →

Universal ML/DL Engine — full capability parity with NumPy, SciPy, Scikit-learn, XGBoost, LightGBM, CatBoost, PyTorch, JAX, Matplotlib, and Seaborn under a single original API.

Highlights

→vp.nn — Pooling, RNN/LSTM/GRU, Attention, full Transformer, GroupNorm, RMSNorm, InstanceNorm, ModuleList/Dict
→vp.nn.F — 50+ functional ops: activations, conv, pool, norm, dropout, all losses, fold/unfold
→vp.nn.init — xavier, kaiming, orthogonal, sparse, trunc_normal, dirac
→vp.transforms — JAX-style grad, vmap, jit, pmap, checkpoint, jacobian, hessian, jvp, vjp
→vp.ml — LGBMClassifier/Regressor/Ranker, CatBoostClassifier/Regressor/Ranker
→vp.ml — IsolationForest, LOF, EllipticEnvelope, OneClassSVM
→vp.ml — Bagging, AdaBoost, ExtraTrees, Voting, Stacking ensembles
→vp.ml — Pipeline, ColumnTransformer, FeatureUnion, RFE, RFECV, SelectFromModel
→vp.ml — TSNE (Barnes-Hut), UMAP, MDS, Isomap, LLE, SpectralEmbedding
→vp.integrate — quad, dblquad, nquad, solve_ivp, odeint, trapezoid, simpson
→vp.interpolate — CubicSpline, BSpline, griddata, RBFInterpolator, Akima1D, PCHIP
→vp.spatial — KDTree, cKDTree, ConvexHull, Delaunay, Voronoi, cdist, pdist
→vp.ndimage — gaussian/median filter, sobel, morphological ops, label, zoom, rotate
→vp.polynomial — Polynomial, Chebyshev, Legendre, Laguerre, Hermite + utilities
→vp.ma — Full masked array support
→vp.io — MAT, WAV, Parquet, Feather, HDF5, NetCDF4, Matrix Market, NumPy native
→vp.viz — 20+ plot types with dark/light/paper themes and ML-specific charts
→vp.hub — save_model, load_model, Checkpointer, model_card
→vp.quantization — INT8/INT4/FP16 PTQ, MinMaxCalibrator, magnitude pruning
→scipy >= 1.10 added as required dependency

Breaking changes:None — all v0.1.0 public API preserved.

v0.1.0

Initial

October 2025

View Docs →

Initial public release. Full tensor API, reverse-mode autograd, neural network layers, classical ML (40+ estimators), DataFrames, and scientific computing primitives.

Highlights

→vp.Tensor — full ndarray API with broadcasting, slicing, reshaping, type casting
→vp.autograd — dynamic reverse-mode computation graph, no_grad context
→vp.nn — Linear, Conv2d, BatchNorm, LayerNorm, Dropout, Embedding, all activations, all loss functions, Sequential
→vp.optim — SGD, Adam, AdamW, RMSprop, Adagrad + StepLR, CosineAnnealing, OneCycle, ReduceLROnPlateau
→vp.ml — LinearRegression, Ridge, Lasso, ElasticNet, LogisticRegression, Decision Trees, Random Forest, GradientBoosting, XGBoost, KMeans, DBSCAN, PCA, SVC, SVR, KNN, Naive Bayes
→vp.data — DataFrame, Series, GroupBy, merge, pivot, melt, read_csv, read_json
→vp.linalg — svd, eig, solve, lstsq, cholesky, qr, inv, pinv, det, norm
→vp.stats — descriptive stats, correlation, t-test, chi2, ANOVA
→vp.fft, vp.signal, vp.sparse, vp.optimize, vp.amp
→vp.utils.data — Dataset, TensorDataset, DataLoader

Breaking changes:N/A — initial release.