Skip to content

gpu-devcontainer.md

Latest commit

 

History

History
60 lines (43 loc) · 1.88 KB

gpu-devcontainer.md

File metadata and controls

60 lines (43 loc) · 1.88 KB
Raw

NVIDIA GPU Devcontainer Workflow

GPU benchmarks use the CUDA devcontainer. They require an NVIDIA GPU and the NVIDIA Container Toolkit on the host.

Target profile:

export BENCHMARK_TARGET_PROFILE=nvidia-gpu

Start the CUDA container:

devcontainer up --workspace-folder . --config .devcontainer/cuda/devcontainer.json

Install optional vendor libraries:

devcontainer exec --workspace-folder . --config .devcontainer/cuda/devcontainer.json \
  bash -lc './scripts/setup_gpu_vendors.sh all'

Run all GPU suites:

devcontainer exec --workspace-folder . --config .devcontainer/cuda/devcontainer.json \
  bash -lc 'BENCHMARK_TARGET_PROFILE=nvidia-gpu ./scripts/run_gpu_suite.sh'

Run selected backends:

devcontainer exec --workspace-folder . --config .devcontainer/cuda/devcontainer.json \
  bash -lc 'BENCHMARK_TARGET_PROFILE=nvidia-gpu GPU_BENCH_BACKENDS=pytorch-cuda ./scripts/run_gpu_suite.sh'

Standard GPU collection requires both tenferro-rs and PyTorch. JAX is not part of maintained GPU reports: during the 2026-07-28 A100 80GB permutation run, the JAX/XLA backend failed to complete after more than 20 minutes while using about 61 GiB of device memory, and earlier rank-24 compilation exceeded 40 minutes. Keeping it in every refresh made the full suite operationally unreliable. Vendor backends remain optional references.

Latest reports:

  • result/nvidia-gpu/gpu/dense.md
  • result/nvidia-gpu/gpu/einsum.md
  • result/nvidia-gpu/gpu/sparse.md
  • result/nvidia-gpu/gpu/tensornetwork.md

Raw run data:

  • data/results/nvidia-gpu/gpu/<suite>/<timestamp>/run.yaml
  • data/results/nvidia-gpu/gpu/<suite>/<timestamp>/records.jsonl
  • data/results/nvidia-gpu/gpu/<suite>/<timestamp>/report.md

GPU timings must include backend-native device synchronization, but not output downloads. Output downloads belong to verification outside the timed region.