Add easy KernelBench level2 problem ports

[saarang123]

Summary

• add four exact-port KernelBench Level 2 problems
• add deterministic seeded testcase generation and small debug-friendly samples
• add a local validator script for sample/checker-style sanity on these new ports

Included problems

• gemm-relu-divide
• conv2d-divide-leaky-relu
• conv2d-hardswish-relu
• matmul-mish-mish

Implementation notes

• each problem materializes learned module state as explicit deterministic inputs so the Tensara runtime contract stays simple
• ports preserve the original op ordering from KernelBench
• verifiers report concrete mismatches instead of only returning a boolean
• samples are intentionally small for debugger/sample-path use
• testcase generation is seeded with Problem.get_seed(...)

Validation

• py_compile passed for all new def.py files and the validator script
• all four problem classes import and instantiate successfully against the Tensara Problem base
• sample/testcase metadata generation was checked in a local Torch environment

Environment limitation

• full CUDA execution was not run in this shell because the Torch environments available here were CPU-only (torch.cuda.is_available() == False)

Extra file

• staging/validate_kernelbench_level2_ports.py provides a local CUDA-side validation path for these ports once run in a GPU-capable environment

[saarang123]

Validated on Together H100 on March 30, 2026.

Ran the new KernelBench Level 2 ports on a free production node after checking that nvidia-smi showed no active compute apps. Remote setup was:

• source /data/srng/.bashrc
• conda activate qwen-inference
• run from /home/srng/repos/problems

Command:

python staging/validate_kernelbench_level2_ports.py --all

Result:

- gemm-relu-divide: sample + all generated tests passed
- conv2d-divide-leaky-relu: sample + all generated tests passed
- conv2d-hardswish-relu: sample + all generated tests passed
- matmul-mish-mish: sample + all generated tests passed