[BlivionIaG]

Purpose

Subtraction-only removal of leftover runtime debug that only logged when an env var was set. With those vars unset, the serving path is unchanged. phase_timer (VLLM_RDNA_PHASE_TIMING) stays. VLLM_GDN_NAN_DEBUG and VLLM_QSA_RDNA2_DEBUG stay.

Not a duplicate. Open PRs were checked before editing:

• cleanup(rdna): drop leftover bring-up debug probes #44 / cleanup(rocm): drop W4A16 prefill dispatcher diary and experiment env hooks #48 are old-base drafts and were left alone. cleanup(rdna): drop leftover bring-up debug probes #44 also deletes VLLM_GDN_NAN_DEBUG and VLLM_QSA_RDNA2_DEBUG, which this PR keeps.
• perf(rdna2): opt-in MoE skinny v2 decode, M=16 decode kernels, small-op pass (port of leapdragon) #57 adds envs.py knobs and _custom_ops.py. This PR only deletes the unread DBG_VLLM_STEP_TIMING entry at the old lines 1445-1447. Those perf(rdna2): opt-in MoE skinny v2 decode, M=16 decode kernels, small-op pass (port of leapdragon) #57 additions are not in this diff.
• feat(rocm): Glm5Next on gfx1030 through the upstream model #51 does not touch these files.
• perf(rdna2): QSA/GDN HIP prefill paths vs Triton/FLA + TunableOp decode/MTP grid (implements #64) #99's net diff on mamba_utils.py, worker/utils.py, gpu/block_table.py, and envs.py is empty. Its tuning.md insert is the hybrid-model rows near line 102. This edit is the diagnostics paragraph near line 217.
• fix(tools): drop unread launcher exports and chassis GPU pins #100 deletes the unread-export note in tuning.md (old lines 207-209). This PR does not touch that note.
• tools(rdna): drop the TunableOp mirror and unused probes #101 does not touch these files.
• docs(rdna): drop stale serve instructions #102 owns docs/rdna2/v0.31.0-port-notes.md. That file still names VLLM_BT_DEBUG, VLLM_MAMBA_COPY_DEBUG, and DBG_VLLM_STEP_TIMING in the "still read" list. It was not edited here.

AI assistance was used (Cursor Grok 4.7). A human should review every line. No model-output change is expected, so no model eval was run.

Removed

• vllm/utils/rocm_poolwatch.py. Nothing imported it. rg before delete matched only that file.
• Diagnostics row for VLLM_RDNA_POOLWATCH (_HISTORY, _DUMP_ON_REPLAY) in docs/rdna2/tuning.md. The same paragraph no longer lists VLLM_BT_DEBUG, VLLM_MAMBA_COPY_DEBUG, or DBG_VLLM_STEP_TIMING.
• VLLM_MAMBA_COPY_DEBUG log branches in vllm/v1/worker/mamba_utils.py (create, collect, preprocess). Copy indexing is unchanged.
• VLLM_BT_DEBUG logs in vllm/v1/worker/utils.py, vllm/v1/worker/gpu/model_runner.py, and vllm/v1/worker/gpu/block_table.py. Zeroing and copy-on-write still run. gather_block_tables still returns the same tuple.
• DBG_VLLM_STEP_TIMING: _DbgPhase, _dbg_flush_step, the preprocess wrapper, the flush calls, the banner, and the unread envs.py entry. phase_timer calls are untouched.

Lines removed are from the fork delta (BlivionIaG / kev29lt@gmail.com, including the block-table probe in the port commit he committed). Upstream zeroing, copy-on-write, and preprocess_state stay.

Test Plan

rg -n 'VLLM_RDNA_POOLWATCH|VLLM_MAMBA_COPY_DEBUG|VLLM_BT_DEBUG|DBG_VLLM_STEP_TIMING|_DbgPhase|_dbg_flush_step|rocm_poolwatch' vllm docs/rdna2/tuning.md
uv venv --python 3.12
.venv/bin/python -m py_compile \
  vllm/envs.py \
  vllm/v1/worker/mamba_utils.py \
  vllm/v1/worker/utils.py \
  vllm/v1/worker/gpu/block_table.py \
  vllm/v1/worker/gpu/model_runner.py
uvx ruff@0.14.0 check <edited python files>
uvx ruff@0.14.0 format --check <edited python files>
.venv/bin/python -m pytest tests/test_envs.py -q --tb=line --noconftest

tests/conftest.py imports torch, transformers, and vllm.LLM, so the env suite was run with --noconftest. Mamba worker tests import Triton and the GPU model-state stack and were not run.

Test Result

rg over vllm/ and docs/rdna2/tuning.md: no matches.

Remaining name hits, left on purpose:

• docs/rdna2/v0.31.0-port-notes.md lines 79 and 256 (docs(rdna): drop stale serve instructions #102 owns that file; line 79 is its hunk).
• Two historical copies under bench_results/2026-10-02_pr34-fa-prefill-d128/remote/src/{patched,baseline}/rdna_attn.py.

py_compile on the five edited modules: exit 0.

ruff 0.14.0 check and format --check on those modules: all passed, already formatted.

pytest tests/test_envs.py --noconftest (CPU torch, uv venv, no system pip): 54 passed, 3 failed. The pass set includes test_getattr_with_cache, which evaluates every remaining environment-variable getter. The 3 failures are ModuleNotFoundError: No module named 'zmq' inside compile_factors() / SamplingParams (vllm/utils/network_utils.py), not the deleted debug gates.

GPU check needed

On gfx1030, greedy hybrid smoke (Flash-Next or another GDN/mamba recipe) with every removed env unset, plus one serve log line from VLLM_RDNA_PHASE_TIMING to confirm phase_timer still prints. No model eval was run here.