[BlivionIaG]

Purpose

Remove finished-experiment drivers under tools/ and three env exports that nothing in vllm/ or csrc/ reads. Live scripts/serve_rdna.sh and scripts/rdna_launcher_common.sh keep the same serve behavior aside from those unread exports and one stale-process wait.

Not a duplicate: open PRs #25, #2, #33, #37, #40, and #41 do not touch these files. Closed drafts #43 and #46 overlapped this change; this PR (#42) is the one to keep. csrc/, vllm/, W4A8 kernels, and tools/rdna2/serve_v620_baseline.sh are untouched.

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

Deleted (rg of each basename is empty outside historical docs / bench_results/)

• TunableOp step-3 drivers: campaign_3a_capture.sh, campaign_3bcd_arm.sh, campaign_3bcd_curate.sh, campaign_3d_freeze.sh
• Closed PR [explore] RDNA2 W4A16: does torch.compile freeze the M-dispatch? #30 harness: pr30_b1_probe.sh, pr30_b2_rocprof.sh, pr30_b3_driver.sh
• Split-K source-swap audit: audit_w4a16.sh, audit_split/audit_split_swap.py, probe_prefill_split.py, probe_force_split.py, probe_gptq_k.py
• Unreferenced one-shots: parse_kernel_sass.py, probe_model_inspect2.py, probe_inmodel_layer.py, tools/probe_concurrent_greedy.py, tools/rdna2/tunableop_rows_probe.py, flashnext_w4a8_matrix.sh, tunableop_rows_e2e.sh, tools/rdna2/test_startup_plan_cpu.py

No listed file was still called by a live path, so none were skipped.

Follow-up: the repro line .venv/bin/python tools/rdna2/test_startup_plan_cpu.py is removed from docs/rdna2/V620-CANDIDATE.md. The rest of that doc is unchanged. rg test_startup_plan_cpu is now empty.

tools/rdna2_028/audit_split/ is not empty after the .py deletion. split_old.txt and split_fixed.txt were not on the kill list (they were only inputs to the deleted audit_w4a16.sh), so the directory stays.

One stale mention left on purpose: bench_results/2026-09-30_campaign-matrix/SUMMARY.md still names campaign_3a_capture.sh.

Trimmed

• configure_mtp_tunableop had no callers. Removed once from tools/rdna2_028/tunableop_env.sh. tools/rdna2/tunableop_env.sh is now a thin source of that file. configure_tunableop and configure_v620_tunableop are unchanged.
• VLLM_RDNA_FORCE_FP16 dropped from scripts/rdna_launcher_common.sh and the plan-key list in scripts/serve_rdna.sh.
• VLLM_ROCM_SKIP_LIVE_TAIL_HASH dropped from scripts/recipes/full.env and the plan-key list in scripts/serve_rdna.sh.
• VLLM_ROCM_MOE_PREFILL=0 dropped from scripts/recipes/flashnext-mtp0.env and flashnext-mtp2.env. The name stays in the serve_rdna.sh plan-key list because other launchers still export it (serve_v620_*.sh, step0_*.sh); those files were out of scope.

rdna_kill_stale waits

rdna_kill_stale still sends TERM, then KILL, only to pids whose VLLM_CACHE_ROOT matches. The 8s sleep runs only after a TERM actually landed, and the 3s sleep runs only if any pid was signaled. A launch where pgrep finds nobody no longer waits 11s. The function still returns success in that case (serve_rdna.sh is set -e).

Test Plan

No GPU. Proof is reference search plus shell parse, and a source of the TunableOp helper through the thin mirror:

rg -n 'campaign_3a_capture|campaign_3bcd_arm|campaign_3bcd_curate|campaign_3d_freeze|pr30_b1_probe|pr30_b2_rocprof|pr30_b3_driver|audit_w4a16|audit_split_swap|probe_prefill_split|probe_force_split|probe_gptq_k|parse_kernel_sass|probe_model_inspect2|probe_inmodel_layer|probe_concurrent_greedy|tunableop_rows_probe|flashnext_w4a8_matrix|tunableop_rows_e2e|test_startup_plan_cpu|configure_mtp_tunableop'

rg -n 'VLLM_RDNA_FORCE_FP16|VLLM_ROCM_SKIP_LIVE_TAIL_HASH|VLLM_ROCM_MOE_PREFILL' vllm csrc

bash -n scripts/serve_rdna.sh scripts/rdna_launcher_common.sh \
  tools/rdna2/tunableop_env.sh tools/rdna2_028/tunableop_env.sh \
  scripts/recipes/full.env scripts/recipes/flashnext-mtp0.env \
  scripts/recipes/flashnext-mtp2.env

Also source tools/rdna2/tunableop_env.sh and call configure_v620_tunableop on an unknown library (same contract as tests/tools/test_v620_tuning_coverage.py::test_unknown_library_disables_tuning_without_stopping_service).

rdna_kill_stale was checked under set -euo pipefail with sleep and kill logged: no matching pid does not sleep; a pid with a different VLLM_CACHE_ROOT is left alone; a matching pid gets kill -TERM, sleep 8, kill -KILL, sleep 3.

Test Result

• Basename / configure_mtp_tunableop search: only bench_results/2026-09-30_campaign-matrix/SUMMARY.md (campaign_3a_capture.sh). test_startup_plan_cpu is gone from the tree, including docs/rdna2/V620-CANDIDATE.md.
• vllm/ and csrc/ search for the three env names: empty.
• bash -n on the edited shell files: all ok.
• Thin source defines configure_tunableop and configure_v620_tunableop, and does not define configure_mtp_tunableop.
• Unknown-library call via the mirror: stdout 0:unset, stderr contains WARNING.
• rdna_kill_stale: empty and cache-root mismatch skip both sleeps; a matching pid is signaled TERM then KILL with the 8s and 3s waits.

---

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