Add campaign20 head kernel benchmark tasks (20 cases).

[chushic]

Introduce baseline harness tasks under tasks/campaign20/baseline for GEAK head-kernel benchmarking.

[irvineoy]

Thanks for adding these tasks!

Before merging, could you please confirm whether all newly added tasks have been run through the project’s Task Validator Agent, using the latest available Codex or Claude Code, and share the validation results in the PR?

One more question on task organization: these tasks are placed under the new tasks/campaign20/baseline category, while their task_type values are existing categories (triton2triton, hip2hip, torch2hip, etc). I wonder should these tasks live under the existing triton2triton / hip2hip task categories so they follow the current task organization and discovery conventions?

[irvineoy]

Hi, quick heads-up: we are currently working on a larger refactor #48 that moves the benchmark/runtime flow to a Docker-first environment. This touches a lot of shared infrastructure, so reviews and merges may be slower than usual for a bit.

Once that refactor lands, I’ll help rebase open PRs onto that one. Sorry for the delay, and thanks for your patience.

[irvineoy]

Thanks for adding these campaign20 benchmark tasks. I did a quick pass and there are a few blockers that should be addressed before we can merge this.

1. Some of the new tasks appear to be broken or not self-contained:

   • gemm_a8w8_blockscale, moe_stage1, and moe_stage2 have source_file_path entries that do not exist in the task directory, so they will fail source-file validation.
   • Some tasks list external production ops as target_kernel_functions rather than functions defined in the provided source files, e.g. moe_gemm_fp8_blockscale and paged_attention_decode. This means the agent does not really have a local editable kernel target.
   • A few runners/harnesses depend on hardcoded external paths or installed packages, such as /sgl-workspace/aiter / GEAK_REPO_ROOT, which breaks the self-contained task requirement.

2. main has recently gone through a large refactor, so this PR needs to be rebased onto the latest main and adjusted to the current task structure.

3. Please run the task_validator agent for these tasks and attach the validation reports. Only tasks that pass validation should be added to the codebase.

4. These 20 tasks are currently placed under a new category folder: tasks/campaign20/baseline. Based on their configs, 7 are hip2hip tasks and 13 are triton2triton tasks. Please place them under the existing task categories instead of creating a new top-level category. If you want to run exactly this 20-task subset, you can create your own local config that lists these tasks explicitly.