[BlivionIaG]

Purpose

Subtraction-only cleanup of leftover owner forensics on rdna_extras. Env-unset behavior is unchanged: no new flags, helpers, state, or replacement logging, and vllm/envs.py is untouched.

Not a duplicate. Open PRs on opengfx1030/vllm-rdna (including #41 W4A8 trim, #40 hetero MoE, #37 expert residency, #33 FA decode, #31 W4A16 explore, #25 all-reduce) do not delete these probes. Hands-off files owned by those branches were not edited: vllm/envs.py, vllm/v1/worker/gpu/model_runner.py, vllm/v1/attention/backends/rdna_attn.py, vllm/models/qwen4_exp/amd/hyperconnection.py, vllm/model_executor/layers/quantization/exl3.py, vllm/distributed/device_communicators/rdna_all_reduce.py, csrc/, docs/, bench_results/.

AI assistance was used (Cursor cloud agent). No model-output, accuracy, or serving change: the deleted branches run only when an env var is set, and unset is the default.

Kept / dropped

Item	Result
1. vllm/utils/rocm_poolwatch.py, vllm/utils/rocm_graph_keepalive.py	Deleted. rg found no Python imports. Docs mention the keepalive filename only. C++ rdna2_immortal_zeros in csrc/ left in place.
2. GDN probes in qwen_gdn_linear_attn.py	Deleted VLLM_LOG_GDN_PTRS, VLLM_GDN_OUT_DEBUG, VLLM_GDN_NAN_DEBUG, VLLM_GDN_STATE_DEBUG / VLLM_GDN_SAVE (torch.save to /tmp/gdn_l6.pt), VLLM_GDN_DBG. Kept VLLM_GDN_HIP_PREFILL and VLLM_GDN_DECODE_RDNA2.
3. VLLM_MAMBA_COPY_DEBUG and the debug local in mamba_utils.py	Deleted. No test, script, or launcher reads the knob.
4. QSA _DEBUG / _qsa_dump_bad_rows	Deleted the Python clone/print/dump. The HIP call qsa_compress_groups_rdna2 takes no debug argument, so the call is unchanged. Not dropped. csrc/rocm/qsa_rdna2.cu still getenvs VLLM_QSA_RDNA2_DEBUG for a printf; that file is hands-off.
5. CUDA graph probes	Deleted VLLM_PIECE_IN_DEBUG, VLLM_CG_NAN_INPUT_CHECK, VLLM_CG_REPLAY_LOG. Kept VLLM_CG_SKIP_REPLAY and capture-time _nan_to_num_tree.
6. backends.py VLLM_CG_REPLAY_LOG and VLLM_PIECE_DUMP / VLLM_PIECE_TP	Deleted.
7. VLLM_W4A16_PTR_DEBUG in _tp_all_reduce	Deleted, including the function-local import os. Restored the pre-probe return get_tp_group().all_reduce(input_).
8. W4A16 dead AWQ prefill arm	Deleted the VLLM_W4A16_PTR_DEBUG logs, the unreachable kernel_name == "awq_prefill" arm (selector returns only w4a8_prefill / prefill / exllama / rdna2_decode), the else-arm ops.awq_gemm_rdna2_prefill call, and unused _awq_prefill_available. Kept the gptq/exllama fallback. awq_gemm_rdna2_prefill is not in csrc/.
9. VLLM_BT_DEBUG in block_table.py and worker/utils.py	Deleted those two blocks. Not removed from hands-off model_runner.py and rdna_attn.py.
10. VLLM_DEBUG_PREFIX_HIT	Deleted. Kept VLLM_NO_PARTIAL_HASH and the existing VLLM_TRACE_PREFIX_CACHE logs (not in the delete list; they are a separate trace switch).
11. VLLM_MOE_NAN_DEBUG, VLLM_LAYER_NAN_DEBUG, VLLM_PLE_NAN_DEBUG	Deleted.

Nothing in this list was dropped as a live path. Remaining references are only in hands-off trees.

Proof

rg of the deleted knobs/modules/functions. Hits left outside docs/ and bench_results/:

• VLLM_BT_DEBUG: vllm/v1/worker/gpu/model_runner.py, vllm/v1/attention/backends/rdna_attn.py (hands off).
• VLLM_QSA_RDNA2_DEBUG: csrc/rocm/qsa_rdna2.cu printf (hands off).
• awq_gemm_rdna2_prefill: string inside the kernel-name regex in benchmarks/kernels/w4a16_compile_dispatch/probe.py (not a call; not edited).
• Historical mentions of VLLM_LOG_GDN_PTRS, _awq_prefill_available, and rocm_graph_keepalive.py under docs/ only.

Zero hits in vllm/**/*.py for: VLLM_LOG_GDN_PTRS, VLLM_GDN_OUT_DEBUG, VLLM_GDN_NAN_DEBUG, VLLM_GDN_STATE_DEBUG, VLLM_GDN_SAVE, VLLM_GDN_DBG, VLLM_MAMBA_COPY_DEBUG, VLLM_QSA_RDNA2_DEBUG, _qsa_dump_bad_rows, VLLM_PIECE_IN_DEBUG, VLLM_CG_NAN_INPUT_CHECK, VLLM_CG_REPLAY_LOG, VLLM_PIECE_DUMP, VLLM_PIECE_TP, VLLM_W4A16_PTR_DEBUG, _awq_prefill_available, VLLM_DEBUG_PREFIX_HIT, VLLM_MOE_NAN_DEBUG, VLLM_LAYER_NAN_DEBUG, VLLM_PLE_NAN_DEBUG, rocm_poolwatch, rocm_graph_keepalive.

.venv/bin/python -m py_compile <12 touched .py files>
PY_COMPILE_OK

Ruff 0.14.0 (repo pre-commit rev) on the touched files. No new diagnostics versus rdna_extras HEAD. Remaining ruff check hits are pre-existing on untouched lines (HEAD had more: the deleted probes were the extra E501/UP031/SIM105/B018):

• backends.py E501 line 714 (same line on HEAD)
• cuda_graph.py SIM102 in kept _nan_to_num_tree
• rdna2_w4a16.py B018 on the kept torch.ops._rocm_C.w4a8_gemm_rdna2 probe
• qwen_gdn_linear_attn.py I001 plus two pre-existing E501s
• qsa_rdna2.py I001 (pre-existing import gap)

ruff format --check: 7 touched files already formatted (communication_op.py, ple_offload_layer.py, model.py, kv_cache_coordinator.py, block_table.py, mamba_utils.py, utils.py). The other 5 also fail format --check on HEAD; they were not reformatted, so surviving lines stay as they were.

Diff: 14 files, 4 insertions, 985 deletions. The four insertions restore the pre-probe return / gptq fallback (elif became if after the AWQ arm was removed).

Test Plan

• python -m py_compile on every touched file
• ruff check and ruff format --check with ruff 0.14.0 and pyproject.toml
• rg for each deleted knob
• CPU pytest, if the modules import without a ROCm build:
  • tests/compile/test_cudagraph_replay_inputs.py
  • tests/kernels/quantization/test_rdna2_w4a16_selection.py
  • tests/v1/core/prefix_cache/test_partial_prefix_cache_hits.py

Test Result

py_compile: pass.

Ruff: see Proof. No new findings versus rdna_extras.

Pytest (CPU torch 2.13.0+cpu, PYTHONPATH=/workspace, no ROCm extension build):

6 failed, 67 passed, 11 skipped, 73 errors in 11.51s

Imports succeeded without a compiled ROCm extension. import_kernels only warns when vllm._C is missing.

• 11 skipped are the GPU/ROCm gates (CUDA/HIP required, ROCm only), including the three VLLM_CG_SKIP_REPLAY capture/replay cases.
• 73 errors are teardown only: cleanup_dist_env_and_memory calls torch.accelerator.empty_cache() because platform detection does not see this tree as CPU (vLLM package was not found when importing from PYTHONPATH). The test bodies of the W4A16 selector cases (test_rdna2_w4a16_inner_dispatch, registry order, backend map, compile dispatch) ran and were not assertion failures.
• 6 failures are not from this diff:
  • test_should_copy_and_wrap_eager_piecewise_graphmodules and test_mrope_get_positions_contiguous_per_capture_size raise Failed to infer device type / pin_memory=True with no accelerator.
  • four test_mamba_align_split_* cases raise AttributeError: 'types.SimpleNamespace' object has no attribute '_effective_lpt' in vllm/v1/core/sched/scheduler.py, which this PR does not touch.

Model eval: not applicable. No kernel math, sampler, or default serving path changed.

  

[BlivionIaG]

Closed as duplicate of #44. Same Python bring-up /tmp probe deletions. Prefer #44.