VTL

Maintainer planning: https://github.com/orgs/vlang/projects/8 (Vlang ML Roadmap — VSL + VTL; may require project access)

Detailed roadmap: ROADMAP.md

GPU memory: DEVICE_MEMORY.md

Done (2026-06-01)

Issue	Topic
#87	Serialization + CIFAR checkpoints
#88	vs NumPy/PyTorch benchmarks + PR comments
#89–#91	CUDA Linear/Conv2D + DeviceSession (Phase 1)
#101/#104	GPU activation chain (Phase 2)
#105	Linear CUDA backward (VTL_CUDA_BACKWARD=1, Phase 3)
#107	Conv2D CUDA backward (cuDNN, same eligibility as forward)
#111/#114	Adam on GPU + persistent DeviceSession slots (#106)
#110	Vulkan f32 Linear in Sequential forward (VTL_USE_VULKAN=1, -d vulkan)
#116	f32 autograd: Sequential + MSE forward/backprop compile
—	f32 tiny training: nn_cifar10_f32_tiny_synth + f32_training_smoke_test
—	CUDA training smoke: nn_cifar10_cuda + nn/cuda_training_smoke_test
—	f32 Vulkan training: nn_cifar10_vulkan + f32_vulkan_training_smoke_d_vulkan_test
—	Vulkan Conv2D f32 forward/backward (same-padding, im2col+GEMM)
—	Vulkan ReLU/Sigmoid f32 (relu_vulkan_f32 / sigmoid_vulkan_f32)
—	Vulkan Adam f32 fused shader (VTL_USE_VULKAN=1, VSL adam_step)
—	Conv2D autograd: register weight/bias parents (conv2d_autograd_smoke_test)
#86	DataLoader
—	from_array clones shape (fixes #41 aliasing)

VSL (downstream): #280–#285, #304 conv2d backward GEMM, #305 Adam shaders.

Beta gate (open)

Priority	Issue	Topic
P1	#41	Windows example crash — shape clone landed; needs Windows CI confirmation

Post-beta tracking

Priority	Issue	Topic
P2	#63	ARM GPU support
P2	—	Vulkan: persistent GPU activation chain between layers (CUDA has VTL_GPU_ACTIVATIONS)

The beta contract is tensors, autograd, high-level layers/losses/optimizers, datasets, and f32 CPU training. CUDA and Vulkan remain opt-in experimental backend paths during beta. Generic f64 training is tracked as post-beta until the autograd Gate[T] interface compiles without mixed Payload[f32] and Payload[f64] specialization.

Local development

Lightweight workflow: DEV_LIGHTWEIGHT.md

v up
cd ~/.vmodules
v test vtl/nn vtl/datasets
v run vtl/examples/nn_cifar10_tiny_synth/main.v
v run vtl/examples/nn_cifar10_f32_tiny_synth/main.v
# Vulkan f32 full stack (use -prod for GPU)
# VTL_USE_VULKAN=1 v -prod -d vulkan run vtl/examples/nn_cifar10_vulkan/main.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

CI

• Default PR: bin/test (scoped modules + nn_cifar10_tiny_synth)
• f32 beta smoke: nn_cifar10_f32_tiny_synth and f32 autograd/training tests
• full-ml label: ci-full-ml.yml runs bin/test --full
• CUDA / Vulkan: local or future labeled workflows

Project board sync

From repo root (requires gh scopes read:project, project):

./.github/scripts/sync-ml-project-8.sh