VTL and VSL architecture
View SVG source · PNG
import vtl
t := vtl.from_array([1.0, 2, 3, 4], [2, 2])!
t.get([1, 1])
// 4.0
Features
- Tensors — create, slice, indexed
take/take_nd/take_flat/put, choose/select,
round_decimals, nan_to_num, compress, indices, pad, set membership, reshape,
transpose, move/roll axes,
N-D diagonal views, NumPy-style Kronecker products, broadcast, map/reduce,
einsum, and complex128 storage with elementwise and matrix products
- Autograd — reverse-mode AD; arbitrary computational graphs
- Neural networks —
Sequential API; Linear, Conv2D, LSTM, Attention, …
- Losses & optimizers — MSE, MAE, BCE, Hinge, Focal, CrossEntropy, Huber;
Adam, AdamW, NAdam, RAdam, SGD, …
- Linear algebra — VSL-backed real matmul and general eigensolvers,
pure-V complex128 matmul, solve, QR, LU, Cholesky, SVD, pinv
- Hardware — zero-copy
Tensor.data for C libs; optional CUDA and Vulkan training paths
ML Release Highlights
The ML beta scope is the high-level VTL API: tensors, autograd, layers, losses,
optimizers, datasets, and CPU training. CUDA and Vulkan paths are available for
opt-in validation and early adopters, but remain experimental backend
accelerators rather than stable user contracts.
| Area |
Status |
| f32 training |
Sequential + MSE + Adam smoke tests |
| CUDA |
Experimental opt-in Linear/Conv2D forward; f64 Linear/Conv2D/Dropout backward; activation chain and Adam slots |
| Vulkan |
Experimental opt-in f32 Linear, Conv2D same-padding, ReLU/Sigmoid/Softplus/SELU/HardSwish, fused Adam shader |
| Datasets |
MNIST, IMDB, CIFAR-10 loaders plus CI-safe synthetic examples |
| Benchmarks |
VTL vs NumPy/PyTorch scripts and PR benchmark workflow |
For memory-safe local commands, see DEV_LIGHTWEIGHT.md.
Quick start
import vtl
import vtl.autograd
import vtl.nn.layers
import vtl.nn.models
import vtl.nn.optimizers
mut ctx := autograd.ctx[f32]()
mut model := models.sequential_from_ctx[f32](ctx)
model.input([784])
model.linear(256)
model.linear(10)
model.mse_loss()
input_tensor := vtl.zeros[f32]([64, 784])
mut x := ctx.variable(input_tensor)
y_pred := model.forward(x)!
target := vtl.zeros[f32]([64, 10])
mut loss_val := model.loss(y_pred, target)!
loss_val.backprop()!
mut opt := optimizers.adam_optimizer[f32](optimizers.AdamOptimizerConfig{
learning_rate: 0.001
})
opt.build_params(model.info.layers)
opt.update()!
Module overview
| Module |
Purpose |
Guide |
vtl |
Tensor creation, slicing, broadcasting, reductions |
First steps |
| Complex tensors |
math.complex.Complex storage and elementwise arithmetic |
Complex tensors |
vtl.fft |
Real and complex FFTs via VSL PocketFFT |
FFT |
vtl.csv |
Numeric CSV tensor input and output |
NumPy I/O |
vtl.npy / vtl.npz |
Typed NumPy array and archive I/O |
NumPy I/O |
vtl.autograd |
Differentiable operations and backpropagation |
Autograd |
vtl.autograd_cuda |
Optional CUDA autograd |
Device memory |
vtl.la |
VSL-backed linear algebra |
Linear algebra |
vtl.nn |
Layers, losses, optimizers, and training |
Neural networks |
vtl.nn.models |
Model construction, training, serialization |
Neural networks |
vtl.nn.layers |
Dense, convolutional, recurrent, and attention layers |
Neural networks |
vtl.nn.loss |
Regression and classification losses |
Neural networks |
vtl.nn.optimizers |
Optimizers and learning-rate schedulers |
Optimizers |
vtl.nn.data |
Neural-network data loaders |
Examples |
vtl.nn.internal |
Internal tensor and activation operations |
Source |
vtl.nn.gates |
Neural-network autograd gates |
Autograd |
vtl.datasets |
MNIST, CIFAR-10, and IMDB loaders |
Datasets |
vtl.stats |
Averages, descriptive statistics, and summaries |
Reductions |
| Core lookup |
Unique values, indexing, and set operations |
Indexing |
vtl.ml.metrics |
Machine-learning metrics and evaluation |
Source |
vtl.storage |
CPU, CUDA, VCL, and Vulkan storage |
Device memory |
Installation
VTL uses VSL for linear algebra. The core vtl
module works without optional system BLAS/LAPACK, but LA features need VSL.
Follow VSL install instructions,
then:
v install vtl
Testing
# Run the full scoped suite from ~/.vmodules; compiler parallelism is disabled
# to keep the local memory ceiling predictable.
cd ~/.vmodules
systemd-run --user --scope --property=MemoryMax=4G --property=MemorySwapMax=0 \
env VJOBS=2 VTL_NO_PARALLEL=1 ./vtl/bin/test
See DEV_LIGHTWEIGHT.md for memory-safe subsets in CI.
Documentation
Start Here
| Tutorial |
Topic |
| TUTORIAL_FIRST_STEPS.md |
Tensor creation, indexing, slicing |
| TUTORIAL_MATH_UFUNCS.md |
NumPy-style elementwise math and binary decomposition |
| TUTORIAL_MAP_REDUCE.md |
map / nmap and reductions |
| TUTORIAL_AUTOGRAD.md |
Variable, gates, backprop |
| TUTORIAL_REDUCTIONS.md |
argmax / argmin / cumsum |
| TUTORIAL_NEURAL_NETWORKS.md |
Layers, losses, Sequential |
| TUTORIAL_OPTIMIZERS.md |
Adam, AdamW, RMSProp, schedulers |
| TUTORIAL_LINEAR_ALGEBRA.md |
LA basics via VSL |
| TUTORIAL_FFT.md |
One-dimensional real Fourier transforms via PocketFFT |
| TUTORIAL_ADVANCED_LA.md |
QR, LU, Cholesky, pinv |
| TUTORIAL_BROADCASTING.md |
Broadcasting rules |
| TUTORIAL_SLICING.md |
Slicing and views |
Full index: docs/README.md.
Contributors
VTL’s early development was inspired by vnum, part of the Vlang Num group,
and the work of Christopher Zimmerman.
The core was reimplemented while keeping that lineage and inspiration. The
original vnum work carried a BSD 3-Clause license with attribution to
vlang-num contributors; the historical license text is preserved in the
VTL repository history.
This acknowledgment records that history and does not change VTL’s current
license, shown below.
Made with contributors-img.
License
MIT
评论