cleanup(tools): delete finished tools/rdna2_028 campaign drivers

[BlivionIaG]

Purpose

Delete the finished one-off experiment drivers under tools/rdna2_028/. The live TunableOp row pipeline stays. This is subtraction only: 44 files removed, and two doc lines that were only how-to commands for deleted scripts. No edits under vllm/, csrc/, tools/rdna/, tools/rdna2/, recipes, AGENTS.md, or bench_results/.

This PR supersedes the tools/rdna2_028 deletions in draft #42. #42 targets the stale base rdna_extra/v0.28.0 and cannot land on rdna_extras. #42 itself was not modified. No other open PR deletes these drivers on rdna_extras.

AI assistance was used (Cursor Grok 4.7). No model-output, accuracy, or serving change, so no model eval. Every listed file was dead; none was kept as still live.

Kept (live row pipeline)

docs/rdna2/tuning.md names these, and tools/rdna/serve_rdna.sh plus tools/rdna/rdna_launcher_common.sh source tunableop_env.sh:

• tools/rdna2_028/tunableop_env.sh
• tools/rdna2_028/tunableop_rows_pipeline.sh (calls only tune_prod_shapes.py)
• tools/rdna2_028/tune_prod_shapes.py
• tools/rdna2_028/curate_tunableop_rows.py
• tools/rdna2_028/verify_tunableop_lookup.py
• tools/rdna2_028/verify_lookup_hit.sh (calls only verify_tunableop_lookup.py)
• tools/rdna2_028/probe_record_untuned.py

Deleted

Repo-wide search before deletion (vllm/, csrc/, tools/, tests/, benchmarks/, CMakeLists.txt, .github/, tools/rdna/recipes) found no live source, import, or exec. The seven kept files do not call any of them. Callers that did exist were other files in this delete set. After deletion, a repeat of that search is empty.

Group	Files	Remaining mentions (not callers)
Campaign drivers	campaign_3a_capture.sh, campaign_3bcd_arm.sh, campaign_3bcd_curate.sh, campaign_3d_freeze.sh, campaign_matrix.sh	bench_results/2026-09-30_campaign-matrix/SUMMARY.md, bench_results/2026-10-01_exl3-tunableop/repo-rows-premerge/provenance.json
PR #30 harness	pr30_b1_probe.sh, pr30_b2_rocprof.sh, pr30_b3_driver.sh	none
Split-K audit	audit_w4a16.sh, audit_split/ (audit_split_swap.py, split_old.txt, split_fixed.txt)	none
Step-0 profiles	step0_parse.py, step0_profile.sh, step0_torchprof.sh, step0_tp_parse.py	bench_results/2026-10-02_sweep-gaps/SUMMARY.md
EXL3 campaign	exl3_capture.sh, exl3_validate.sh, exl3_mtp2_validate.sh, exl3_analyze.py, exl3_freezer.py, exl3_tunableop_campaign.sh	bench_results/2026-10-01_exl3-tunableop/SUMMARY.md, bench_results/2026-10-01_exl3-27b-mtp2/SUMMARY.md, tunableop/rocm7.14-rocblas5.5/provenance.json
Flash-Next W4A8	flashnext_w4a8_arm.sh, flashnext_w4a8_extract.py, flashnext_w4a8_matrix.sh	bench_results/2026-09-29_w4a16-regression-audit/SUMMARY.md, bench_results/2026-09-29_tunableop-rows/SUMMARY.md
Matrix / A-B shells	fp32_accum_matrix.sh, fp_aroff_27b.sh, m27b_matrix.sh, w4a8_ab.sh, wt_ab.sh	bench_results/2026-09-29_tp4-fp-aroff/SUMMARY.md, bench_results/2026-09-29_w4a16-regression-audit/SUMMARY.md, bench_results/2026-09-30_fp32-accum-matrix/SUMMARY.md, bench_results/2026-10-02_sweep-gaps/SUMMARY.md
One-off probes	probe_force_split.py, probe_gptq_k.py, probe_inmodel_layer.py, probe_model_inspect2.py, probe_mtp_sampling.py, probe_prefill_split.py, probe_real_weights.py, probe_w4a8.py, probe_w4a8_moe.py, probe_w4a8_op.py	bench_results notes listed above, plus docs/rdna2/w4a8-moe-wiring.md (see follow-ups)
Misc one-shots	parse_kernel_sass.py, build_rocm_c_incr.sh, analyze_captures.py, tunableop_rows_e2e.sh	bench_results/2026-09-29_w4a8-splitk-fix/SUMMARY.md, bench_results/2026-09-30_campaign-matrix/SUMMARY.md, comment in bench_results/2026-10-02_pr34-fa-prefill-d128/remote/build_arm.sh

Doc how-to lines dropped, and nothing else in those docs:

• docs/explore/w4a16-compile-dispatch/RESULTS-2026-09-30.md: the Flash-Next bullet that only named flashnext_w4a8_arm.sh.
• docs/rdna2/w4a8-wiring.md: the sentence that only named m27b_matrix.sh as the matrix driver.

Follow-ups (not done here)

• docs/rdna2/w4a8-moe-wiring.md still names probe_w4a8_moe.py in the "what changed" table and as the caption of the L2 micro table. Those lines record a past result; they are not a how-to command, so they were left for the docs cleanup.
• tunableop/rocm7.14-rocblas5.5/provenance.json still records the historical tune_method / capture_method strings that name exl3_tunableop_campaign.sh and exl3_capture.sh. That is shipped-row provenance, not a caller.
• bench_results/ summaries (and the archived build_arm.sh comment) still name deleted drivers. The data was not edited.
• Draft cleanup(tools): drop finished-experiment drivers and unread env exports #42 also trimmed unread env exports and other tools outside tools/rdna2_028/. Those edits are out of scope for this PR.

Test Plan

No GPU. Proof is reference search plus parse of the kept pipeline, and that the launcher still points at an existing tunableop_env.sh:

rg -n 'campaign_3a_capture|campaign_3bcd_arm|campaign_3bcd_curate|campaign_3d_freeze|campaign_matrix\.sh|pr30_b1_probe|pr30_b2_rocprof|pr30_b3_driver|audit_w4a16|audit_split|step0_parse|step0_profile|step0_torchprof|step0_tp_parse|exl3_capture|exl3_validate|exl3_mtp2_validate|exl3_analyze|exl3_freezer|exl3_tunableop_campaign|flashnext_w4a8_|fp32_accum_matrix|fp_aroff_27b|m27b_matrix|w4a8_ab\.sh|wt_ab\.sh|probe_force_split|probe_gptq_k|probe_inmodel_layer|probe_model_inspect2|probe_mtp_sampling|probe_prefill_split|probe_real_weights|probe_w4a8|parse_kernel_sass|build_rocm_c_incr|analyze_captures|tunableop_rows_e2e' \
  vllm csrc tools tests benchmarks .github CMakeLists.txt tools/rdna/recipes

bash -n tools/rdna2_028/tunableop_env.sh \
  tools/rdna2_028/tunableop_rows_pipeline.sh \
  tools/rdna2_028/verify_lookup_hit.sh

python3 -m py_compile \
  tools/rdna2_028/tune_prod_shapes.py \
  tools/rdna2_028/curate_tunableop_rows.py \
  tools/rdna2_028/verify_tunableop_lookup.py \
  tools/rdna2_028/probe_record_untuned.py

rg -n 'source "\$source_dir/tools/rdna2_028/tunableop_env.sh"' tools/rdna/serve_rdna.sh
test -f tools/rdna2_028/tunableop_env.sh

Test Result

• Live-tree search (vllm/, csrc/, tools/, tests/, benchmarks/, CMakeLists.txt, .github/, tools/rdna/recipes): no matches. The seven kept files: no matches.
• bash -n on the three kept shell scripts: exit 0.
• python3 -m py_compile on the four kept Python files: exit 0. This environment has no .venv and no uv; the check is stdlib syntax only.
• tools/rdna/serve_rdna.sh:244 is source "$source_dir/tools/rdna2_028/tunableop_env.sh", and that file is still in the tree. tools/rdna/rdna_launcher_common.sh:97 is the same source line and was not edited.

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing