cleanup(tools): delete finished one-shot experiment scripts

[BlivionIaG]

Purpose

Delete finished, unreferenced one-shot experiment scripts BlivionIaG added under tools/ in the last month. Deletion only, plus the one stale test_startup_plan_cpu.py line in docs/rdna2/V620-CANDIDATE.md. No launcher, test, or committed reproduction recipe loses a script it points at.

AI assistance was used (Cursor Grok 4.7). No model-output, accuracy, or serving change, so no model eval.

Not a duplicate

Open PR #42 (cursor/dead-tooling-cleanup-6e7b) deletes the same scripts and also rewrites scripts/ and both tunableop_env.sh files, and it leaves the stale test_startup_plan_cpu.py line in docs/rdna2/V620-CANDIDATE.md. This PR does not touch scripts/, tunableop_env.sh, csrc/, or upstream vllm/ files, and it removes that one doc line. Other open PRs (#2, #10, #18, #23, #25, #31, #33, #35, #37, #39, #40, #41) do not modify these paths.

Kept (not in this diff)

• tools/rdna2_028/campaign_matrix.sh — reproduction recipe in bench_results/2026-09-30_campaign-matrix/SUMMARY.md.
• tools/rdna2_028/probe_w4a8_op.py, probe_real_weights.py, build_rocm_c_incr.sh — cited by bench_results/2026-09-29_w4a8-splitk-fix/SUMMARY.md.
• tools/rdna2_028/flashnext_w4a8_arm.sh and scripts/serve_gfx1030_27b_dense.sh — cited by docs/explore/w4a16-compile-dispatch/RESULTS-2026-09-30.md.
• tools/probe_greedy_correctness.py — loaded by tests/entrypoints/openai/test_probe_greedy_correctness.py.
• tools/rdna2_028/audit_split/split_old.txt and split_fixed.txt — not scripts and not on the delete list. Their only caller was audit_w4a16.sh.
• Out of scope, untouched: step0_*, exl3_analyze.py, open-PR files (serve_v620_baseline.sh, check_rdna_ar_native.py, qualify_tool_turns.py, probe_hetero_peer.py, pr33_*), tunableop_env.sh, scripts/, and VLLM_RDNA2_PREFILL_FORCE_* in csrc.

No candidate was kept because of a live reproduction citation. bench_results/2026-09-30_campaign-matrix/SUMMARY.md still names campaign_3a_capture.sh as the prior wrong design (line 23) and mentions a killed campaign_3a server whose intermediate rows were superseded (lines 140–145). The reproduction block there is campaign_matrix.sh. That historical note was not rewritten.

Pre-delete rg -n <basename> (whole repo)

Hits below are the script itself, a sibling on this delete list, the superseded SUMMARY note, or the one doc line removed in this PR.

basename	hits
campaign_3a_capture.sh	bench_results/2026-09-30_campaign-matrix/SUMMARY.md:23 (prior wrong design)
campaign_3bcd_arm.sh	none
campaign_3bcd_curate.sh	none
campaign_3d_freeze.sh	none
pr30_b1_probe.sh	self comment; pr30_b2_rocprof.sh:12 (sibling)
pr30_b2_rocprof.sh	self comment
pr30_b3_driver.sh	self comment
audit_w4a16.sh	self comments
audit_split_swap.py	audit_w4a16.sh:58,62 (sibling)
probe_prefill_split.py	audit_w4a16.sh:50 (sibling)
probe_force_split.py	none
probe_gptq_k.py	none
parse_kernel_sass.py	none
probe_model_inspect2.py	none
probe_inmodel_layer.py	none
flashnext_w4a8_matrix.sh	self comment
tunableop_rows_e2e.sh	self comment
probe_concurrent_greedy.py	none
tunableop_rows_probe.py	none
test_startup_plan_cpu.py	self comment; docs/rdna2/V620-CANDIDATE.md:124 (line deleted; tests/v1/worker/test_startup_plan.py does not exist)

Test Plan

for b in campaign_3a_capture.sh campaign_3bcd_arm.sh campaign_3bcd_curate.sh \
  campaign_3d_freeze.sh pr30_b1_probe.sh pr30_b2_rocprof.sh pr30_b3_driver.sh \
  audit_w4a16.sh audit_split_swap.py probe_prefill_split.py probe_force_split.py \
  probe_gptq_k.py parse_kernel_sass.py probe_model_inspect2.py probe_inmodel_layer.py \
  flashnext_w4a8_matrix.sh tunableop_rows_e2e.sh probe_concurrent_greedy.py \
  tunableop_rows_probe.py test_startup_plan_cpu.py
do
  rg -n "$b" scripts tests tools vllm || true
done

# remaining tools/rdna2_028 shell that names a deleted basename, then bash -n
.venv/bin/python -m pytest tests/tools -q

Test Result

Post-delete rg of every deleted basename under scripts/, tests/, tools/, and vllm/: no hits (20/20 clean).

Whole-repo rg after deletion: the only remaining hit is bench_results/2026-09-30_campaign-matrix/SUMMARY.md:23 for campaign_3a_capture.sh (the prior-wrong-design note). test_startup_plan_cpu.py has zero hits. Every other deleted basename has zero hits.

No remaining shell script under tools/rdna2_028 names a deleted basename, so bash -n had nothing to check.

.venv/bin/python -m pytest tests/tools -q does not collect on this CPU checkout. tests/conftest.py imports tblib (then torch and vllm), and this environment has no project install:

ImportError while loading conftest '/workspace/tests/conftest.py'.
tests/conftest.py:7: in <module>
    from tblib import pickling_support
E   ModuleNotFoundError: No module named 'tblib'

Diff is 21 files, 2055 deletions. The only non-deletion is the one-line removal in docs/rdna2/V620-CANDIDATE.md.

---

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

  

---

Closed as duplicate of #42. Same dead-tooling deletions; #42 also covers the env-export trims and tunableop_env dedupe. Prefer #42.