SOL-ExecBench B200 kernels

CUDA C++, CuTe DSL, and Triton solutions for SOL-ExecBench. Sources, manifests, and validation reports live in kernels/.

Native Runpod setup

The checkout, environment, toolchain, datasets, and caches live under /workspace. Run bash tools/bootstrap_native.sh, then source tools/native_env.sh in each shell. See native setup, strategy, and local validation results. The active ten-leaderboard goal tracks local leads and confirmed hosted results separately.

python tools/campaign.py fetch 25 53 84 85 88
python tools/campaign.py bench 84 --solution kernels/084_silu_activation_backward/solution.json
python tools/campaign.py package 84 --solution kernels/084_silu_activation_backward/solution.json

GPU work uses a shared file lock. Local scores are estimates; this Runpod host denies clock locking, and official evaluations fix B200 SM clocks at 1500 MHz.

Docker setup

Requires Git, Make, Docker with NVIDIA Container Toolkit, and driver 580 or newer.

git submodule update --init --recursive
make setup

make setup builds the pinned evaluator image and downloads problem #38 into .work/.

make info
make lint
make test
make bench
make profile WORKLOAD=0

B200

make test TARGET=b200
make bench TARGET=b200
make status

Problems 29 and 179 use the extra-* targets with KERNEL_ID=29 or KERNEL_ID=179, for example:

make extra-test KERNEL_ID=179
make extra-compile KERNEL_ID=179 TARGET=b200
make extra-package

Package

make package
make verify-package

Run make verify-package on B200. Kernel #38 is generated under dist/; problems 29 and 179 are generated under .work/extra-kernels/submissions/. Upload the matching JSON to NVIDIA. Submission and publication are manual.

Apache-2.0. See LICENSE.