[BlivionIaG]

1. Dead probe (#104)

tools/rdna2_028/probe_force_split.py set VLLM_RDNA2_PREFILL_FORCE_SPLIT_K, whose only reader was removed in #104, so it silently ran the default config. Deleted the probe. No other references outside bench_results/ (history, left as is).

2. Port notes (#102 vs #103)

docs/rdna2/v0.31.0-port-notes.md listed VLLM_BT_DEBUG, VLLM_MAMBA_COPY_DEBUG, DBG_VLLM_STEP_TIMING as still read, and said VLLM_BT_DEBUG was kept. #103 removed all three readers. Both places fixed; the list now holds only VLLM_MOE_NAN_DEBUG, VLLM_LAYER_NAN_DEBUG (vllm/models/qwen4_exp/amd/model.py), VLLM_TRACE_PREFIX_CACHE, VLLM_DEBUG_PREFIX_HIT, VLLM_NO_PARTIAL_HASH (vllm/v1/core/kv_cache_coordinator.py), each verified by grep. A note says the other three were dropped in #103.

3. TunableOp README (#101)

tunableop/README.md still named the deleted tools/rdna2/tunableop_env.sh mirror (two places). Now points only at tools/rdna2_028/tunableop_env.sh.

4. Skip message (#106)

tests/kernels/quantization/test_gdn_prefill_kkt_rdna2.py skipped unconditionally while telling readers to run it in isolation. It now skips unless VLLM_RUN_GDN_KKT_HANG_TEST=1, and the message says how to run it and why it is skipped (recorded GDN GPU hang). tests/kernels/test_gdn_prefill_rdna2.py had the same unconditional skip on its hang test; it got the same opt-in (skipif). tools/rdna/port_v031/unit.sh header now documents the env var.

Checks

• grep -rn for probe_force_split|VLLM_RDNA2_PREFILL_FORCE, the three removed knobs, and tools/rdna2/tunableop_env: only the references fixed here remain (excluding bench_results/).
• grep -rl for the five kept knobs under vllm csrc tools: all still read.
• python3 -m py_compile on both edited test files: ok.
• uvx ruff check and uvx ruff format --check on both: passed.
• Tests not run (no GPU, per task).