v0.2.0-beta.1 — ML Beta

VTL v0.2.0-beta.1 — ML Beta: f32 training, autograd, and opt-in GPU paths

Paired release: VSL v0.2.0-beta.1 — required dependency.

Overview

First tagged release of VTL as the V machine learning library: n-dimensional tensors, reverse-mode autograd, Sequential neural networks, losses, optimizers, datasets, and f32 CPU training as the stable beta path. CUDA and Vulkan acceleration hooks are included for early adopters but remain experimental.

Depends on: vsl@0.2.0-beta.1

Highlights

Core ML stack (stable beta)

• Tensor[T] — creation, slicing, reshape, transpose, broadcast, reductions, stacking.
• vtl.autograd — Context, Variable, gates, backprop(); f32 Sequential + MSE training.
• vtl.nn.models.Sequential — layer composition, forward, loss attachment.
• Layers — Linear, Conv2D, activations, pooling, normalization, LSTM, attention.
• Losses — MSE, BCE, CrossEntropy, Huber, KL, NLL, and more.
• Optimizers — Adam, AdamW, SGD, RMSProp, AdaGrad + schedulers; f32 CPU Adam for training smokes.
• Datasets — MNIST, IMDB, CIFAR-10 + DataLoader[T].
• Serialization — model checkpoint round-trip.
• from_array shape clone — fixes tensor aliasing (#41).

f32 training path (beta flagship)

• Examples: nn_cifar10_f32_tiny_synth, nn_cifar10_tiny_synth (f64).
• CI smokes: f32 autograd, f32 training tests, scoped bin/test.
• See docs/DEV_LIGHTWEIGHT.md for memory-safe local commands.

CUDA integration (experimental, opt-in)

Phase	Feature	Env / flag
1	Linear + Conv2D forward, DeviceSession	-d cuda, VTL_USE_CUDA=1
2	GPU activation chain	-d cuda
3	Linear/Conv2D backward	VTL_CUDA_BACKWARD=1
4	Adam GPU moment updates + persistent slots	VTL_CUDA_OPTIMIZER=1

Examples/tests: nn_cifar10_cuda, cuda_training_smoke_test.

Vulkan integration (experimental, opt-in)

• Linear forward in Sequential (VTL_USE_VULKAN=1, -d vulkan)
• Conv2D same-padding forward/backward (im2col + GEMM)
• ReLU/Sigmoid f32 unary ops
• Fused Adam f32 shader (VSL adam_step)
• Examples: nn_cifar10_vulkan, nn_cifar10_f32_vulkan_tiny_synth

Stable vs experimental (beta contract)

Tier	Surface
Stable	vtl, vtl.autograd, vtl.nn.models, vtl.nn.layers, vtl.nn.loss, vtl.nn.optimizers (CPU f32), vtl.datasets, vtl.storage (CPU)
Experimental	*_cuda*, *_vulkan*, autograd_cuda/**, GPU optimizer hooks
Internal	vtl.nn.internal/**

Primary path: f32 CPU training with Sequential + Adam + MSE.

Platform support (validated)

Platform	Status	Notes
Linux (ubuntu 20.04/22.04)	Supported	Primary CI
Windows	Smoke validated	Tensor creation smoke + setup-v@v1.6
macOS	Known issue	Inherited VSL LAPACKE issue (#91)
CUDA / Vulkan	Opt-in	Local hardware + env flags

Installation

v install vsl@0.2.0-beta.1
v install vtl@0.2.0-beta.1

vtl/v.mod for this release:

Module {
    name: 'vtl'
    version: '0.2.0-beta.1'
    dependencies: ['vsl@0.2.0-beta.1']
}

Quick start (CPU f32)

v run vtl/examples/nn_cifar10_f32_tiny_synth/main.v
cd vtl && ./bin/test

Opt-in GPU (local)

VTL_USE_CUDA=1 VTL_TEST_CUDA=1 VJOBS=1 v -d cuda test vtl/nn/cuda_training_smoke_test.v
VTL_USE_VULKAN=1 VTL_TEST_VULKAN=1 VJOBS=1 v -prod -d vulkan test vtl/nn/f32_vulkan_training_smoke_d_vulkan_test.v

Examples (beta-relevant)

Example	Purpose
nn_cifar10_f32_tiny_synth	Beta smoke — f32 tiny training
nn_cifar10_tiny_synth	f64 tiny training smoke
nn_cifar10_cuda	CUDA training smoke
nn_cifar10_vulkan	Full Vulkan f32 stack
nn_mnist, nn_xor	Learning path

Known issues (beta)

1. macOS — VSL LAPACKE path may break LA-heavy tests
2. Windows — smoke validated; full matrix not on Windows
3. f64 training — post-beta primary path; use f32
4. Vulkan — no persistent GPU activation chain yet (post-beta)

Upstream: VSL

Requires VSL v0.2.0-beta.1. Do not mix beta tags across repos.

Links

• Documentation
• Tutorial
• ML Roadmap
• Dev lightweight guide
• VSL v0.2.0-beta.1