[BlivionIaG]

Purpose

Live serve path only. Four items, no new flags, functions, or files.

1. Stop exporting env knobs nothing in vllm/, csrc/, or tests/ reads. rg -n NAME vllm csrc tests is empty for each name below, so none of the findings were dropped.
   • VLLM_RDNA_FORCE_FP16: removed from rdna_init and from serve_rdna.sh's print list.
   • VLLM_ROCM_SKIP_LIVE_TAIL_HASH: removed from scripts/recipes/full.env (assignment and the stale comment) and from the print list.
   • VLLM_ROCM_MOE_PREFILL: removed from flashnext-mtp0.env, flashnext-mtp2.env, and the print list. VLLM_GDN_HIP_PREFILL stays; it is read in vllm/model_executor/layers/mamba/gdn/qwen_gdn_linear_attn.py.
2. tools/rdna2/tunableop_env.sh was byte-identical to tools/rdna2_028/tunableop_env.sh below a 3-line header. Its body is now one source of the rdna2_028 file, path resolved from this script's directory. SPDX header kept. Callers (serve_v620_baseline.sh included; that file is not edited) still get configure_tunableop, configure_v620_tunableop, and _tunableop_profiles.
3. Deleted the dead alias configure_mtp_tunableop and its "kept for existing callers" comment. rg of the name is empty after the edit.
4. rdna_kill_stale slept 8s after TERM and 3s after the loop even when pgrep matched nothing. It now sleeps only after a pid was actually signaled. The pgrep/kill loops are unchanged.

Not a duplicate of #42. That PR deletes finished-experiment drivers under tools/ and leaves VLLM_ROCM_MOE_PREFILL in the serve_rdna.sh print list. It does not change rdna_kill_stale. This PR is only the live serve path (scripts/serve_rdna.sh, scripts/rdna_launcher_common.sh, scripts/recipes/*.env, the two TunableOp helpers). Untouched: tools/rdna2/serve_v620_baseline.sh (#25), docs/, bench_results/, every vllm/ and csrc/ file, and the other tools/ scripts.

AI assistance was used (Cursor Grok 4.7). No model-output, accuracy, or serving change, so no model eval. Launch behavior changes only by not exporting unread names and by skipping the 11s wait when there is nothing to kill.

Test Plan

rg -n VLLM_RDNA_FORCE_FP16 vllm csrc tests
rg -n VLLM_ROCM_SKIP_LIVE_TAIL_HASH vllm csrc tests
rg -n VLLM_ROCM_MOE_PREFILL vllm csrc tests
rg -n configure_mtp_tunableop

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

shellcheck -s bash --format=gcc \
  scripts/serve_rdna.sh scripts/rdna_launcher_common.sh \
  scripts/recipes/full.env scripts/recipes/flashnext-mtp0.env \
  scripts/recipes/flashnext-mtp2.env \
  tools/rdna2/tunableop_env.sh tools/rdna2_028/tunableop_env.sh

# PRINT=1 dry-run, before vs after, for each recipe
# full, flashnext, flashnext-mtp0, flashnext-mtp2, 27b-awq, 27b-exl3
PRINT=1 bash scripts/serve_rdna.sh RECIPE=<name> MODEL=/tmp/model VENV=/tmp/venv

bash -c 'source tools/rdna2/tunableop_env.sh; declare -F'
diff <(bash -c 'source tools/rdna2/tunableop_env.sh; declare -f') \
     <(bash -c 'source tools/rdna2_028/tunableop_env.sh; declare -f')

python -m pytest tests/tools/test_v620_tuning_coverage.py -q --noconftest

rdna_kill_stale under set -euo pipefail: time a call with no matching pid; then a pid whose command line matches but whose VLLM_CACHE_ROOT does not; then a pid that matches both (sleep overridden so the positive path does not actually wait 11s).

Test Result

• rg of the three env names under vllm/, csrc/, and tests/: no matches. configure_mtp_tunableop: no matches.

• bash -n on the seven touched shell files: ok.

• shellcheck 0.9.0 (same invocation, .shellcheckrc applied): 126 findings before, 123 after, 0 new. The three that disappeared are SC2034 on the deleted recipe assignments.

• PRINT=1 env diff, only these lines removed (commands unchanged):

  • every recipe: export VLLM_RDNA_FORCE_FP16='1'
  • full only: export VLLM_ROCM_SKIP_LIVE_TAIL_HASH='1'
  • flashnext-mtp0 and flashnext-mtp2 only: export VLLM_ROCM_MOE_PREFILL='0'
  • VLLM_GDN_HIP_PREFILL='0' still printed for the two flashnext-mtp recipes.

• declare -F after source tools/rdna2/tunableop_env.sh:

  declare -f _tunableop_profiles
  declare -f configure_tunableop
  declare -f configure_v620_tunableop
  

  Before, the same list also had configure_mtp_tunableop. declare -f bodies are identical to sourcing tools/rdna2_028/tunableop_env.sh directly.

• rdna_kill_stale:

  • no matching pid: returned 0 in 8 ms (set -e did not abort). Before this change the same call took 11015 ms.
  • pgrep hit with a different VLLM_CACHE_ROOT: no sleep, process left alive.
  • pgrep hit with the launcher's VLLM_CACHE_ROOT: kill succeeded, recorded SLEEP 8 then SLEEP 3, process was gone, function returned 0.

• python -m pytest tests/tools/test_v620_tuning_coverage.py -q cannot load tests/conftest.py here (it imports torch and vllm; no project venv). The same file with --noconftest on CPU: 4 passed, 2 subtests passed in 0.20s. python3 -m unittest tests.tools.test_v620_tuning_coverage -q: Ran 4 tests ... OK. That test sources tools/rdna2/tunableop_env.sh and calls configure_v620_tunableop.

---

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

  

[BlivionIaG]

Closed as overlapping #42. Same no-op env knobs + tunableop_env.sh dedupe. The unique rdna_kill_stale sleep-only-after-signal tweak is being folded into #42 instead.