VSL

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

Repo roadmap: ROADMAP.md · CUDA: cuda/README.md

Done (2026-06-01)

Issue	Topic
#236	Multi-backend GPU architecture
#237–#239	Vulkan / CUDA / OpenCL foundations
#280	cuBLAS/cuDNN kernels
#281	GPU numerical validation tests
#282	vs NumPy benchmarks + PR comments
#283–#285	Vulkan gating, conv2d, ComputeContext tests
#304	Vulkan Conv2D backward d_weight GEMM layout fix
#305	Vulkan vector_mul, vector_sqrt, fused f32 adam_step shader

VTL (downstream): CUDA Phases 1–4, f32 autograd/training, Vulkan Linear, Conv2D, ReLU/Sigmoid, and Adam are wired into the nn_cifar10_vulkan smoke.

Beta gate (open)

Priority	Issue	Topic
P1	#225	Windows build
P2	#91	BLAS on macOS validation

Post-beta tracking

Priority	Issue	Topic
P1	Phase H	Multi-GPU (device_id, data parallelism)
P1	Phase I	GPU memory pool / zero-copy
P2	#226	VCL examples
P2	—	CUDA variants of benchmarks/vs_numpy/ in CI
P2	—	Extended Vulkan↔CUDA numerical cross-check
P2	—	Vulkan persistent memory / reduced host sync for VTL training

The ML beta uses VSL as a scientific and compute foundation. The stable beta contract is the portable CPU/scientific surface plus vsl.compute; CUDA, Vulkan, and VCL are opt-in experimental accelerators.

Local development

v up
cd ~/.vmodules
v test vsl/blas vsl/la vsl/compute
# CUDA smoke
v -d cuda test vsl/cuda/examples/cuda_ops_test.v
# Vulkan (opt-in; avoid full `v test vsl/vulkan` on low-RAM hosts)
cd vsl && ./bin/test --use-vulkan
VSL_TEST_VULKAN=1 VJOBS=1 v -prod -d vulkan test vulkan/compute/adam_step_vulkan_test.v

Project board sync

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