SOL-ExecBench B200 kernels

CUDA C++ and CuTe DSL solutions for:

• 029_mamba_conv1d_with_gating
• 038_flux_multi_head_rmsnorm_qk
• 003_fp8_mlp_gate_up_projection (SOL problem 179)

Each source submission lives under its matching kernels/ subdirectory.

Requirements

• Git and Make
• Docker with NVIDIA Container Toolkit
• NVIDIA driver 580 or newer
• NVIDIA GPU

Setup

git submodule update --init --recursive
make setup

make setup builds the pinned evaluator image and downloads problem #38. The benchmark data stays under .work/ and is not committed.

Local

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

Local timings are not official B200 SOL Scores.

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.