Skip to content

DEV_LIGHTWEIGHT.md

Latest commit

 

History

History
77 lines (62 loc) · 3.73 KB

DEV_LIGHTWEIGHT.md

File metadata and controls

77 lines (62 loc) · 3.73 KB

Lightweight development (avoid OOM / machine freeze)

VTL/VSL test suites compile many modules in parallel. On a 32-core machine, v test vsl or v test vtl can spike RAM into the 10–20+ GB range during compilation alone.

Rules of thumb

Do Don't
VJOBS=2 v test vtl/nn/layers/layers_test.v v test vtl (full tree)
VJOBS=2 v test vsl/blas vsl/la vsl/ml v test vsl (82 tests + Vulkan)
v run vtl/examples/nn_cifar10_tiny_synth/main.v v run vtl/examples/nn_cifar10/main.v (huge compile)
Opt-in GPU: VTL_USE_CUDA=1 v -d cuda ... -d cuda on every command by default

Environment variables

Variable Default Meaning
VTL_USE_CUDA off Set to 1 to use CUDA in Linear forward (-d cuda build only).
VTL_GPU_ACTIVATIONS off Phase 2: chain Linear activations on GPU between layers.
VTL_CUDA_BACKWARD off Phase 3: cuBLAS GEMM for Linear gate backward.
VTL_CUDA_OPTIMIZER off Phase 4: cuBLAS moment updates for Adam.
VTL_TEST_CUDA off Set to 1 to run GPU tests (linear_cuda_test.v, device_session_test.v).
VTL_USE_VULKAN off f32 Linear/Conv2D/activations/Adam via Vulkan (-d vulkan). Use v -prod for GPU (debug instance crash on V 0.5.1).
VTL_TEST_VULKAN off Optional Vulkan integration tests (conv2d, activations, Adam, training smoke).
VJOBS (V auto) Cap parallel compile jobs, e.g. VJOBS=2.

CUDA is opt-in so normal CPU work never touches the GPU driver.

Recommended commands

v up
cd ~/.vmodules

# CPU smoke (fast)
VJOBS=2 v test vtl/nn/layers/layers_test.v
VJOBS=2 v test vtl/nn/models/serialization_test.v
v run vtl/examples/nn_cifar10_tiny_synth/main.v
VJOBS=2 v run vtl/examples/nn_cifar10_f32_tiny_synth/main.v
VJOBS=1 v test vtl/nn/f32_training_smoke_test.v vtl/nn/f32_autograd_smoke_test.v
VTL_USE_CUDA=1 VTL_TEST_CUDA=1 VJOBS=1 v -d cuda test vtl/nn/cuda_training_smoke_test.v
VTL_USE_CUDA=1 VJOBS=1 v -d cuda run vtl/examples/nn_cifar10_cuda/main.v
VJOBS=2 v run vtl/examples/nn_cifar10_f32_vulkan_tiny_synth/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

# Single-file CUDA test (only when you want GPU)
VTL_USE_CUDA=1 VTL_TEST_CUDA=1 VJOBS=1 v -d cuda test vtl/nn/layers/linear_cuda_test.v
VJOBS=2 v test vtl/autograd/device_session_test.v
# GPU backward parity (optional)
# VTL_USE_CUDA=1 VTL_TEST_CUDA=1 VTL_CUDA_BACKWARD=1 VJOBS=1 v -d cuda test vtl/autograd/device_session_test.v

# VSL CUDA ops (one file)
VJOBS=1 v -d cuda test vsl/cuda/examples/cuda_ops_test.v

# Vulkan f32 (CPU compile path without SDK)
VJOBS=2 v test vtl/nn/layers/linear_vulkan_integration_test.v
v run vtl/examples/nn_cifar10_vulkan/main.v

# Vulkan f32 GPU full stack (Linear + Conv2D + ReLU + Adam)
VTL_USE_VULKAN=1 VJOBS=1 v -prod -d vulkan run vtl/examples/nn_cifar10_vulkan/main.v
VTL_USE_VULKAN=1 VJOBS=1 v -prod -d vulkan test vtl/nn/f32_vulkan_training_smoke_d_vulkan_test.v
VTL_USE_VULKAN=1 VJOBS=1 v -prod -d vulkan test \
  vtl/nn/internal/conv2d_vulkan_forward_f32_d_vulkan_test.v \
  vtl/nn/internal/conv2d_vulkan_backward_f32_d_vulkan_test.v \
  vtl/nn/layers/activation_vulkan_relu_f32_d_vulkan_test.v \
  vtl/nn/optimizers/adam_f32_vulkan_d_vulkan_test.v
VSL_TEST_VULKAN=1 VJOBS=1 v -prod -d vulkan test vsl/vulkan/compute/adam_step_vulkan_test.v

CI split (implemented, #109)

  • PR default (ci.yml): bin/test — scoped v test, compile examples except nn_cifar10/main.v, and run tiny synthetic CIFAR examples.
  • Label full-ml: workflow ci-full-ml.yml — bin/test --full (weekly schedule + manual dispatch)
  • Local full suite: VJOBS=1 ~/.vmodules/vtl/bin/test --full