1엔진 환경변수 정리 계획: 카테고리별 현재 → 통합 · 이름 변경 · 삭제

생성 2026-09-30 10:41 · vLLM 0.14.0 대비 엔진에 추가된 286개 · 설계 Astra(GPU 쪽 89) + Fable(CPU 쪽 197) · 코드는 아직 고치지 않음 · 전체 목록 /envs/ ↗ · CSV ↗
한 줄 요약
로그·trace·dump 스위치는 LOG_PRINT_<카테고리>=0~3, DUMP=, CS_OUT, BREAKDOWN 으로 통합하고, 동작을 바꾸는 기능 변수는 유지하되 접두어를 카테고리 이름으로 변경(옛 이름은 한동안 경고와 함께 받음), 한 번 쓰고 남은 디버그 probe 는 삭제한다.

1b20190511 우선 정리 목록 (121개)

20190511 이 주 작업자인 변수만 카테고리별로 모았다(커밋 수 기준). 이 목록부터 정리한다. 통합 25 · 이름 변경 29 · 삭제 12 · 그대로 유지 55
카테고리20190511 주 작업카테고리 전체비율통합이름 변경삭제유지
전송(delta transfer)2828100%71812
출력 · 덤프 · 시간 분해264459%1178
CPU verify216731%318
실행 · 하니스183748%117
KV offload · pool · NUMA143737%545
선택(recompression)83622%143
GPU draft 엔진61346%312
전송(delta transfer) — 28개
처리지금앞으로 / 이유도입 날짜 · 커밋
통합COMPACT_ACK_LOG→LOG_PRINT_DELTA=22026-08-21 · vllm:27a6d3e61
통합COMPACT_DIFF_STAT→DELTA 1: existing-value summaries; DELTA 2: _diff_stat comparisons and synchronized recompaction timing. Preserve exact old feature selection in compa…2026-07-19 · top:d026cff92
통합DELTA_COMPACTION_LOG→LOG_PRINT_DELTA; translate legacy features individually, not numeric levels directly.2026-09-08 · vllm:4773d97f4
통합HELD_TRACE→DELTA 2 observation **plus**, for the old alias only, DELTA_PUBLISH_FULL_KEEPMAP=1; new log levels do not enable publication.2026-08-21 · top:366f94d23
통합LOG_DELTA_SELECTION→Value-1 summary → DELTA 1; detailed lifecycle fields → DELTA 2; value-2 full-pool/selection fingerprints → DELTA 3. Legacy alias preserves its origina…2026-09-17 · vllm:4a474f4fc
통합RECOMPACT_FP_LOG→LOG_PRINT_DELTA=3 (GPU sync)2026-09-18 · vllm:444079ef2
통합RECOMPACT_TRACE→LOG_PRINT_DELTA=2 (jsonl traces)2026-08-21 · top:366f94d23
이름 변경COMPACT_DIFF_UPLOAD→DELTA_UPLOAD; same boolean meaning, including request-level overrides.2026-07-19 · vllm_cpu:e3a54bc6c
이름 변경COMPACT_FINISHED_RESEND→rename DELTA_FINISHED_RESEND2026-09-04 · vllm:cad0868ec
이름 변경COMPACT_HOLD_FULL→DUMP_METADATA_ROUNDS; diagnostic generation-range modifier, not an algorithm setting and not an enable switch.2026-09-20 · vllm_cpu:71ec15f0b
이름 변경COMPACT_INDEX_MAP_MAX→DELTA_INDEX_MAP_MAX; preserve bounded-ledger/LRU behavior.2026-07-19 · vllm_cpu:915f8700c
이름 변경COMPACT_KEEPMAP_SLACK→DELTA_KEEPMAP_SLACK; preserve selection-versus-union behavior.2026-09-17 · vllm_cpu:e37119de9
이름 변경COMPACT_KEEPMAP_SLOT_MULT→DELTA_KEEPMAP_SLOT_MULT; preserve slot-capacity multiplier and fallback rules.2026-09-19 · vllm_cpu:627db3a05
이름 변경COMPACT_UPLOAD_PEND→DELTA_UPLOAD_PEND; preserve pending-upload branch exactly.2026-07-19 · top:b00ff022a
이름 변경CS_DRAFT_HOLD→rename DELTA_PRUNE_HOLD (alias CS_DRAFT_HOLD)2026-09-18 · vllm:665ac0687
이름 변경CS_DRAFT_HOLD_MAX→rename DELTA_PRUNE_HOLD_MAX2026-09-18 · vllm:665ac0687
이름 변경CS_LANDING_PRUNE→rename DELTA_LANDING_PRUNE2026-09-18 · vllm:a0d2add69
이름 변경CS_PARITY_ASSERT→DELTA_ASSERT_SERIAL_REQUESTS; preserve explicit assertion, default off, independent of level 3.2026-09-18 · vllm_cpu:546a74454
이름 변경DIFF_NATIVE→DELTA_NATIVE; same native-versus-fallback diff planner choice.2026-08-21 · top:366f94d23
이름 변경DIFF_PAD_SHORT_FATAL→DELTA_PAD_SHORT_FATAL; same explicit failure policy; never a verbosity effect.2026-09-08 · vllm_cpu:7b237c96c
이름 변경DIFF_PARTIAL_RELOAD→rename DELTA_PARTIAL_RELOAD (alias)2026-09-18 · vllm:a0c9f6ee1
이름 변경KEEPMAP_PAD_LAYERS→DELTA_KEEPMAP_PAD_LAYERS; same survivor-count padding behavior.2026-09-04 · vllm_cpu:832edfa60
이름 변경KEEPMAP_STRICT_SLOTS→rename DELTA_KEEPMAP_STRICT_SLOTS (alias)2026-09-04 · vllm:104462b63
이름 변경RECOMPACT_SLOT_ALIGN→DELTA_RECOMPACT_SLOT_ALIGN; preserve slot-alignment behavior.2026-09-08 · vllm:4773d97f4
이름 변경UPLOAD_SKIP_FATAL→rename DELTA_UPLOAD_SKIP_FATAL (alias)2026-09-08 · vllm:f8127c335
삭제KEEPMAP_SLOT_LOG✕-2026-09-04 · vllm:104462b63
그대로 유지COMPACT_ACT_RESERVE_GIB→unchanged (runners set it)2026-08-21 · top:366f94d23
그대로 유지COMPACT_MAX_GPUBLK_UTIL→unchanged (runners set it 14x)2026-08-21 · top:366f94d23
출력 · 덤프 · 시간 분해 — 26개
처리지금앞으로 / 이유도입 날짜 · 커밋
통합ATTN_KV_DUMP_DIR→DUMP=kv2026-05-14 · vllm:a1313a187
통합BREAKDOWN_DIR→BREAKDOWN=rows plus CS_OUT; retain old directory/timestamp behavior in legacy mode.2026-05-20 · vllm:541d1de36
통합BREAKDOWN_KERNEL_INLINE→Jointly with profile mode → BREAKDOWN_KERNEL; (1,1) → inline+gpu-profiler, (1,0) → profiler. INLINE alone does not activate profiling.2026-07-11 · top:249e00398
통합BREAKDOWN_KERNEL_PROFILE_MODE→BREAKDOWN_KERNEL; zero disables this timing family; one uses the joint translation above.2026-05-22 · top:b2ea7da4a
통합BREAKDOWN_LIGHT_LAT→BREAKDOWN=light2026-08-21 · vllm:0fbbc02c0
통합BREAKDOWN_RUN_DIR→derived: CS_OUT_RUN/breakdown (launcher no longer exports it)2026-05-20 · vllm:541d1de36
통합COMPACT_HOLD_DUMP→Fingerprints → DELTA 3, trace/held.jsonl; full records → explicit DUMP=metadata. Legacy explicit path retains its narrow enablement and exact destinat…2026-09-20 · vllm_cpu:71ec15f0b
통합DEBUG_METADATA_DUMP_DIR→DUMP=accept plus CS_OUT; preserve separately gated extensions and legacy directory semantics.2026-04-04 · vllm:97f6af266
통합LOGPROB_DUMP_DIR→DUMP=logprob2026-05-12 · top:19f58c922
통합OUT_ROOT→CS_OUT for migrated engine output routing; legacy held-output fallback remains <OUT_ROOT>/held.jsonl. Path only, no enablement.2026-07-06 · top:1afaede19
통합REJECT_DISTRIB_DUMP→DUMP=reject (reject-only) or DUMP=reject_all2026-07-11 · top:249e00398
삭제BREAKDOWN_FLUSH_EVERY✕-2026-05-20 · vllm:541d1de36
삭제BREAKDOWN_GPU_EVENT_DEPTH✕- (hard-code 4)2026-05-20 · vllm:541d1de36
삭제BREAKDOWN_PER_REQ✕-2026-05-20 · vllm:541d1de36
삭제COMPACTION_IDX_DUMP✕Remove obsolete full selected-index writer. Do not silently fold it into DUMP=accept.2026-07-19 · vllm_cpu:915f8700c
삭제COMPACTION_IDX_EVERY✕Remove sampling modifier with its writer.2026-07-19 · vllm_cpu:915f8700c
삭제COMPACTION_IDX_LAYERS✕Remove layer filter with its writer.2026-07-19 · vllm_cpu:915f8700c
삭제INIT_FILE_PATH✕-2026-02-06 · top:1dfcd3a72
그대로 유지CPU_PROFILE_CHUNK→unchanged (tool-local knob of verify/profile_cpu_forward.py)2026-06-06 · top:cd24516c7
그대로 유지CPU_PROFILE_CHUNK_ITERS→unchanged (tool-local knob of verify/profile_cpu_forward.py)2026-06-06 · top:cd24516c7
그대로 유지CPU_PROFILE_COMPACT_RATE→unchanged (tool-local knob of verify/profile_cpu_forward.py)2026-06-20 · top:80a874b3e
그대로 유지CPU_PROFILE_CTX_TRUNC→unchanged (tool-local knob of verify/profile_cpu_forward.py)2026-06-20 · top:80a874b3e
그대로 유지CPU_PROFILE_MAX_TOKENS→unchanged (tool-local knob of verify/profile_cpu_forward.py)2026-06-05 · top:6e0e5149e
그대로 유지CPU_PROFILE_PROMPT_LOGPROBS→unchanged (tool-local knob of verify/profile_cpu_forward.py)2026-06-06 · top:cd24516c7
그대로 유지REJECT_DISTRIB_TOPK→unchanged (dump parameter of DUMP=reject)2026-07-11 · top:249e00398
그대로 유지TOKEN_DUMP_JSON→unchanged (path-valued; see s3)2026-09-17 · top:338f77bfa
CPU verify — 21개
처리지금앞으로 / 이유도입 날짜 · 커밋
통합DRAFT_TOKEN_LOG→LOG_PRINT_VERIFY=22026-09-18 · top:f5afcb189
통합FSM_RUN_DIR→LOG_PRINT_VERIFY=2 (writer on)2026-06-17 · vllm:029702e50
통합LAST_VERIFY_LOG→LOG_PRINT_VERIFY=22026-09-18 · vllm:049dc2f31
그대로 유지ACCEPT_MODE→unchanged (VERIFY domain; judge_debugger sets it)2026-07-24 · top:d3d63e183
그대로 유지CACTUS_DELTA→unchanged2026-07-24 · top:d3d63e183
그대로 유지LOGPROB_POOL_SIZE→unchanged (shared contract GPU<->CPU)2026-06-05 · top:6e0e5149e
그대로 유지LOGPROB_SHM_NAME→unchanged2026-05-12 · top:19f58c922
그대로 유지MAXPOS_OVERRIDE→unchanged2026-07-12 · top:e39fe49ac
그대로 유지REJECT_DISTRIB_PT→unchanged (dump parameter: every N-th full-vocab .pt)2026-07-11 · top:249e00398
그대로 유지REJECT_DISTRIB_PT_MAX→unchanged (dump parameter)2026-07-11 · top:249e00398
그대로 유지REJECT_DISTRIB_SOURCE→unchanged (dump parameter: server or both)2026-07-11 · top:249e00398
그대로 유지RELAX_RESIDUAL→unchanged2026-07-24 · top:d3d63e183
그대로 유지RELAX_TAU→unchanged2026-07-24 · top:d3d63e183
그대로 유지TAS_DELTA→unchanged2026-07-24 · top:d3d63e183
그대로 유지TAS_EPS→unchanged2026-07-24 · top:d3d63e183
그대로 유지VERIFY_ENFORCE_EAGER→unchanged2026-03-12 · top:45b9be148
그대로 유지VERIFY_KV_PREFAULT→VERIFY_KV_PREFAULT; preserve optional background page-in behavior.2026-08-21 · top:366f94d23
그대로 유지VERIFY_KV_PREFAULT_ARM_IDLE→VERIFY_KV_PREFAULT_ARM_IDLE; retain seconds and default 2.0.2026-08-21 · top:366f94d23
그대로 유지VERIFY_MAX_NUM_SEQS→unchanged2026-06-03 · top:4ad643b05
그대로 유지VERIFY_MODEL→unchanged2026-03-12 · top:45b9be148
그대로 유지VERIFY_PROMPT_LOGPROBS→unchanged2026-03-12 · top:45b9be148
실행 · 하니스 — 18개
처리지금앞으로 / 이유도입 날짜 · 커밋
이름 변경ARM→CS_RUN_TAG; opaque run/artifact label. Preserve old ARM acceptance and existing harness meaning; do not confuse it with per-request policy dictionarie…2026-01-17 · vllm_cpu:9e078d058
그대로 유지BENCH_BATCH→unchanged; canonical over RULER_BATCH (category: functional)2026-07-24 · top:d3d63e183
그대로 유지BENCH_CTX→unchanged; canonical over RULER_CTX (category: functional)2026-07-24 · top:d3d63e183
그대로 유지GPU_CORES→unchanged2026-07-11 · top:249e00398
그대로 유지MAX_MODEL_LEN→unchanged (profile tool / test harness max_model_len)2026-01-21 · vllm_cpu:f0feb1cf8
그대로 유지MODEL_DIR→unchanged2026-02-05 · top:65001e5be
그대로 유지MULTIBP_COMPRESS→unchanged2026-06-05 · top:2d8c73527
그대로 유지MULTIBP_MAX_MODEL_LEN→unchanged2026-06-05 · top:2d8c73527
그대로 유지MULTIBP_MAX_NUM_SEQS→unchanged2026-06-05 · top:2d8c73527
그대로 유지MULTIBP_QUESTION→unchanged2026-06-03 · top:8fc22680a
그대로 유지MULTIBP_REPEAT→unchanged2026-06-20 · top:80a874b3e
그대로 유지RULER_OFFSET→unchanged2026-07-09 · top:7462c1b8b
그대로 유지RULER_QDIR→unchanged2026-07-11 · top:249e00398
그대로 유지RULER_SEQ→unchanged2026-06-05 · top:6e0e5149e
그대로 유지RULER_SUFFIX→unchanged2026-07-11 · top:249e00398
그대로 유지SEED→unchanged2026-01-16 · vllm_cpu:7a1030431
그대로 유지SMALL_MODEL_DIR→unchanged2026-02-27 · top:6e9c5f0fd
그대로 유지TASK_NAME→unchanged2026-07-11 · top:249e00398
KV offload · pool · NUMA — 14개
처리지금앞으로 / 이유도입 날짜 · 커밋
이름 변경COMPACT_FULL_COPY_GB→rename KV_FULL_COPY_GB (alias)2026-08-21 · vllm:0fbbc02c0
이름 변경COMPACT_TEMP_BUFFER_GB→rename KV_TEMP_BUFFER_GB (alias)2026-08-21 · vllm:0fbbc02c0
이름 변경COMPRESS_SPEC_CPU_POOL_HOLD→rename KV_CPU_POOL_HOLD (alias)2026-07-19 · vllm:9bbec5634
이름 변경COMPRESS_SPEC_PIN_FULLKV→KV_PIN_FULLKV; independent block-lifetime control, default off.2026-07-19 · vllm_cpu:915f8700c
이름 변경COMPSPEC_RELOAD_STALL_S→rename KV_RELOAD_STALL_S2026-07-19 · vllm:9bbec5634
삭제ATTN_BP_DEBUG_FILL✕- (delete the fill path from the real-run connector; keep debug_fill_slot_random for test/test_score.py)2026-04-04 · vllm:97f6af266
삭제KV_DUMP_DIR✕Remove destination for the retired parity-dump family; coordinate paired writers with the other slice. Do not alias it to all KV dumps.2026-05-14 · vllm:a1313a187
삭제KV_DUMP_NCT✕-2026-09-17 · vllm:883bd67ec
삭제KV_DUMP_SEQ✕Remove request-prefix/sequence-range parity dump and its tensor reads.2026-09-17 · vllm_cpu:e37119de9
그대로 유지ATTN_BP_STREAM_CAP→unchanged2026-07-13 · vllm:d10dda1f3
그대로 유지MAX_NUM_SEQS→MAX_NUM_SEQS; shared capacity input, including existing fallback/headroom rules.2026-06-03 · top:4ad643b05
그대로 유지NUM_GPU_BLOCKS→NUM_GPU_BLOCKS; shared GPU-capacity input, not a DELTA-only budget.2026-07-19 · top:f015dc6be
그대로 유지VLLM_SHM_NO_INTERLEAVE→unchanged2026-07-02 · top:55c6c47cb
그대로 유지VLLM_SHM_SUFFIX→VLLM_SHM_SUFFIX; preserve cross-process namespace and UID fallback.2026-06-16 · vllm:81fe55182
선택(recompression) — 8개
처리지금앞으로 / 이유도입 날짜 · 커밋
통합SCORE_LOG→LOG_PRINT_COMPACTION=22026-07-19 · top:b00ff022a
이름 변경COMPACT_BLOCK_SIZE→derive from BLOCK_SIZE; keep COMPACT_BLOCK_SIZE as alias2026-09-04 · vllm:76f0d2c5b
이름 변경COMPACT_METHOD→COMPACTION_METHOD; same selection method and dependent offload-pool behavior.2026-05-06 · vllm_cpu:68cbf974a
이름 변경COMPACT_NATIVE→COMPACTION_NATIVE; preserve requested backend, warmup, and runtime fallback.2026-06-05 · vllm_cpu:aefabbd48
이름 변경COMPACT_TOKEN_GLOBAL_RANK→COMPACTION_TOKEN_GLOBAL_RANK; changes the score used across layers.2026-05-06 · vllm_cpu:68cbf974a
그대로 유지SCORE_BLOCK_SIZE→SCORE_BLOCK_SIZE; retain SHM-layout meaning.2026-05-15 · top:6ccc92878
그대로 유지SCORE_POOL_SIZE→SCORE_POOL_SIZE; retain capacity meaning and runner values.2026-05-15 · top:6ccc92878
그대로 유지STREAMINGLLM_SINK_TOKENS→unchanged2026-06-05 · top:2d8c73527
GPU draft 엔진 — 6개
처리지금앞으로 / 이유도입 날짜 · 커밋
통합CS_STOP_LOG→LOG_PRINT_DRAFT=22026-09-18 · vllm:049dc2f31
통합DRAFT_DISPATCH_LOG→LOG_PRINT_DRAFT=22026-09-18 · vllm:fef6646aa
통합PRESERVE_LOG→LOG_PRINT_DRAFT=22026-09-18 · vllm:fdc5aca7c
이름 변경COMPSPEC_DRAFT_STALL_S→rename DRAFT_STALL_S2026-07-19 · vllm:9bbec5634
그대로 유지DRAFT_DISPATCH_FULL→unchanged (runners pass it)2026-09-18 · vllm:fef6646aa
그대로 유지SPECULATIVE_RUN_STEP→SPECULATIVE_RUN_STEP; shared algorithm/protocol setting, including ledger shrink-window use.2026-06-05 · top:2d8c73527

2카테고리별 개수

#카테고리무엇통합이름 변경그대로 유지삭제합
2.1선택(recompression)verify attention 으로 GPU 에 남길 블록을 고르는 규칙과 score 전달1241136
2.2전송(delta transfer)없는 블록만 보내기, index map · keep map · ack · landing prune7182128
2.3GPU draft 엔진draft 실행, 스케줄러, run-ahead, dispatch435113
2.4CPU verifyverify 엔진·서버, rejection sampling, logprob, 수락 규칙7746767
2.5KV offload · pool · NUMAGPU↔CPU KV 이동, CPU pool, 공유 메모리, NUMA 배치111101537
2.6prefill (layer-serial · HeadInfer)긴 프롬프트의 chunk prefill, streaming, HeadInfer decode211316
2.7모드 전환(router)resident / speculative 모드 전환134
2.8출력 · 덤프 · 시간 분해로그 파일, 덤프, breakdown, 출력 경로1891744
2.9실행 · 하니스모델 경로, 벤치 입력, CLI, 환경 이름134237
2.10기타위에 넣기 어려운 것44
전체406613347286

2b카테고리별 주 작업자

주 작업자 = 그 변수 이름을 추가·변경한 커밋(git log -S, v0.14.0 이후) 수가 가장 많은 사람. 같으면 처음 넣은 사람. 표의 작업자 이름에 마우스를 올리면 사람별 커밋 수와 도입 날짜가 보인다. 기록 없음 = 커밋 기록에서 찾지 못함(미커밋 코드 등)
카테고리20190511ysshinwonhyukother기록 없음합
선택(recompression)823536
전송(delta transfer)2828
GPU draft 엔진62513
CPU verify213213167
KV offload · pool · NUMA14101337
prefill (layer-serial · HeadInfer)1616
모드 전환(router)44
출력 · 덤프 · 시간 분해261844
실행 · 하니스18414137
기타44
전체121917220286

3새 공용 knob

knob값뜻
LOG_PRINT0–3모든 카테고리의 기본 로그 레벨
LOG_PRINT_<카테고리>0–3COMPACTION · DELTA · DRAFT · VERIFY · KV · ROUTER · PREFILL 별 덮어쓰기. 0 = 지금과 같음(조용) · 1 = 라운드당 요약, GPU sync 없음 · 2 = 요청·layer 상세, jsonl · 3 = 무거운 검사(sync, fingerprint)
LOG_PRINT_REQreq_id특정 요청만 기록
DUMPaccept,score,kv,logprob,memcheck,reject,metadata,tokens고른 덤프만 CS_OUT/<run>/dump/<kind>/ 에
CS_OUT경로출력 루트 하나. 실행마다 CS_OUT/<시각>/{logs,breakdown,trace,dump}. 경로만 정하고 아무것도 켜지 않음
BREAKDOWN · BREAKDOWN_KERNEL0/rows/light · 0/inline/profiler/inline+gpu-profiler시간 분해 기록

4선택(recompression) (36)

verify attention 으로 GPU 에 남길 블록을 고르는 규칙과 score 전달 · 통합 1 · 이름 변경 24 · 그대로 유지 11

4.1통합: 지금 변수 → 새 knob · 레벨 (1)

#주 작업자지금앞으로근거
4.1.120190511SCORE_LOG→LOG_PRINT_COMPACTION=2config/connector_config.py:185 -> attn_bp_connector.py:109,369-371 ([score-xfer] per entry); cpu_verify_engine.py:668,46…

4.2이름 변경: 지금 이름 → 새 이름 (옛 이름 호환) (24)

#주 작업자지금앞으로근거
4.2.1ysshinCOMPACT_BLOCK_KEEP_RATIO→COMPACTION_BLOCK_KEEP_RATIO; retained fraction for ratio policy, including existing rounding/overflow behavior.C-runner:4014, :4075.
4.2.2ysshinCOMPACT_BLOCK_POLICY→COMPACTION_BLOCK_POLICY; preserve mass, budget, ratio.C-runner:4011, :4024.
4.2.320190511COMPACT_BLOCK_SIZE→derive from BLOCK_SIZE; keep COMPACT_BLOCK_SIZE as aliaskeep_map_reader.py:31-32 · slot alignment for the keep map = KV block size; UNCLEAR whether any run sets it != BLOCK_SIZ…
4.2.4ysshinCOMPACT_DISABLE→COMPACTION_DISABLE; preserve disable polarity and full-KV operating assumption.C-runner:4743.
4.2.5wonhyukCOMPACT_GATE_AGG→COMPACTION_GATE_AGG; same LSE sufficiency aggregation.C-runner:4884.
4.2.6ysshinCOMPACT_GATE_SCORE→COMPACTION_GATE_SCORE; preserve code default and runner overrides separately.C-runner:4751; supplied .env sets zero.
4.2.7ysshinCOMPACT_GATE_THRESHOLD→COMPACTION_GATE_THRESHOLD; do not merge with selection threshold.C-runner:4892.
4.2.820190511COMPACT_METHOD→COMPACTION_METHOD; same selection method and dependent offload-pool behavior.vllm/vllm/v1/kv_offload/cpu.py:78; C-runner:5190.
4.2.920190511COMPACT_NATIVE→COMPACTION_NATIVE; preserve requested backend, warmup, and runtime fallback.vllm_cpu/vllm/v1/worker/cpu_model_runner.py:68.
4.2.10ysshinCOMPACT_SCORE_EXPORT_LEGACY→COMPACTION_SCORE_EXPORT_LEGACY; retain live comparison path; correct this row’s removal marker.C-attn:47, :586, :587, :611.
4.2.11ysshinCOMPACT_SCORE_FORCE_OFF→COMPACTION_SCORE_FORCE_OFF; explicit export-cost experiment setting, never activated by a log level.C-attn:51, :586; inventory records one enabled setter.
4.2.12ysshinCOMPACT_SCORE_FP16→COMPACTION_SCORE_FP16; preserve precision and reduction-dependent applicability.C-attn:65, :596.
4.2.13ysshinCOMPACT_SCORE_NO_REDUCE→COMPACTION_SCORE_NO_REDUCE; preserve negative polarity and interaction with head aggregation.C-attn:59, :587.
4.2.14ysshinCOMPACT_THRESHOLD→COMPACTION_THRESHOLD; preserve each caller’s current fallback and explicit runner values.C-runner:5185; inventory records differing consumer defaults and 28 setters.
4.2.1520190511COMPACT_TOKEN_GLOBAL_RANK→COMPACTION_TOKEN_GLOBAL_RANK; changes the score used across layers.C-runner:3737.
4.2.16ysshinCOMPACT_TOKEN_HEAD_AGG→COMPACTION_TOKEN_HEAD_AGG; preserve aggregation and score-export coupling.C-attn:60; C-runner:3725.
4.2.17ysshinCOMPACT_TOKEN_NUM_RECENT→COMPACTION_TOKEN_NUM_RECENT; same protected recent-token count.C-runner:5203.
4.2.18wonhyukCOMPACT_TOKEN_NUM_SINK→COMPACTION_TOKEN_NUM_SINK; same protected sink-token count.C-runner:5202.
4.2.19ysshinCOMPACT_TOKEN_POOL_K→COMPACTION_TOKEN_POOL_K; same score-pooling width and disable behavior.C-runner:3726.
4.2.20ysshinCOMPACT_TOKEN_POOL_TYPE→COMPACTION_TOKEN_POOL_TYPE; same pooling operator.C-runner:3727.
4.2.21ysshinCOMPACT_TOKEN_SCORE_ROWS→COMPACTION_TOKEN_SCORE_ROWS; same exported/consumed row policy.C-attn:69, :628.
4.2.22ysshinCP_BATCH_HEAD_REDUCE→COMPACTION_BATCH_HEAD_REDUCE; same batched score-reduction path.C-runner:3865.
4.2.23ysshinDENSE_NO_COMPACT→COMPACTION_DENSE_NO_COMPACT; preserve its no-draft baseline condition.C-runner:4756.
4.2.24ysshinEXP_SELECTED_TOKENS→COMPACTION_SELECTED_TOKENS; preserve all current consumers and budget fallback precedence.C-runner:4080; inventory records eight reads across engine/harness/tests.
4.3그대로 유지 (11)
#주 작업자지금앞으로근거
4.3.1wonhyukCOMPRESSION_SPEC_SELECTED_TOKENS→unchangedcompression_spec_config.py:81; cpu_verify_engine.py:389; cli.py:245; connector_config.py:58,76
4.3.2wonhyukLSE_SHM_NAME→LSE_SHM_NAME; retain exact SHM naming override.C-score:24.
4.3.320190511SCORE_BLOCK_SIZE→SCORE_BLOCK_SIZE; retain SHM-layout meaning.C-score:28.
4.3.4ysshinSCORE_NUM_LAYERS→SCORE_NUM_LAYERS; retain layout dimension and existing model-derived overrides.C-score:25; inventory lists other consumers’ model-derived defaults.
4.3.5ysshinSCORE_NUM_Q_HEADS→SCORE_NUM_Q_HEADS; retain layout dimension and model-derived overrides.C-score:27.
4.3.620190511SCORE_POOL_SIZE→SCORE_POOL_SIZE; retain capacity meaning and runner values.C-score:26.
4.3.7wonhyukSCORE_SHM_NAME→SCORE_SHM_NAME; retain exact SHM naming override.C-score:23.
4.3.8ysshinSNAPKV_FUSE_RELOAD→unchangedflashinfer.py:96
4.3.9ysshinSNAPKV_KEY_CHUNK→unchangedflashinfer.py:94-95
4.3.10ysshinSNAPKV_WIN_CHUNK→unchangedflashinfer.py:2477
4.3.1120190511STREAMINGLLM_SINK_TOKENS→unchangedconnector_config.py:56,204

5전송(delta transfer) (28)

없는 블록만 보내기, index map · keep map · ack · landing prune · 통합 7 · 이름 변경 18 · 그대로 유지 2 · 삭제 1

5.1통합: 지금 변수 → 새 knob · 레벨 (7)

#주 작업자지금앞으로근거
5.1.120190511COMPACT_ACK_LOG→LOG_PRINT_DELTA=2cpu_verify_engine.py:2068,2092,2801,3748 · 1-3 lines per ack; runners pass ${COMPACT_ACK_LOG:-0} (2)
5.1.220190511COMPACT_DIFF_STAT→DELTA 1: existing-value summaries; DELTA 2: _diff_stat comparisons and synchronized recompaction timing. Preserve exact old feature selection in compatibility.C-index:351, :367; G-move:21, :78; Audit:53, :102.
5.1.320190511DELTA_COMPACTION_LOG→LOG_PRINT_DELTA; translate legacy features individually, not numeric levels directly.vllm/vllm/cpu_verify/compaction_trace.py:10; G-move:21; Audit:54.
5.1.420190511HELD_TRACE→DELTA 2 observation **plus**, for the old alias only, DELTA_PUBLISH_FULL_KEEPMAP=1; new log levels do not enable publication.G-engine:1857; C-upload:229; Audit:52.
5.1.520190511LOG_DELTA_SELECTION→Value-1 summary → DELTA 1; detailed lifecycle fields → DELTA 2; value-2 full-pool/selection fingerprints → DELTA 3. Legacy alias preserves its original bundle.G-temp:325; C-runner:5406; Audit:55, :56.
5.1.620190511RECOMPACT_FP_LOG→LOG_PRINT_DELTA=3 (GPU sync)gpu_recompact.py:26-37 (_slot_sums(...).numpy() = full pool D2H per plan)
5.1.720190511RECOMPACT_TRACE→LOG_PRINT_DELTA=2 (jsonl traces) · -> CS_OUT/<run>/trace/{reload_trace,verify_trace,recompact_trace,draft_trace}.jsonlcompaction_trace.py:10-40 (writes to BREAKDOWN_DIR root, mode a); cpu_tracker.py:212; breakdown_logger.py:296-306 (draft…

5.2이름 변경: 지금 이름 → 새 이름 (옛 이름 호환) (18)

#주 작업자지금앞으로근거
5.2.120190511COMPACT_DIFF_UPLOAD→DELTA_UPLOAD; same boolean meaning, including request-level overrides.C-runner:5187; C-upload:74; inventory records 33 setters.
5.2.220190511COMPACT_FINISHED_RESEND→rename DELTA_FINISHED_RESENDcpu_verify_engine.py:4341-4342
5.2.320190511COMPACT_HOLD_FULL→DUMP_METADATA_ROUNDS; diagnostic generation-range modifier, not an algorithm setting and not an enable switch.C-index:96, :117.
5.2.420190511COMPACT_INDEX_MAP_MAX→DELTA_INDEX_MAP_MAX; preserve bounded-ledger/LRU behavior.C-index:53.
5.2.520190511COMPACT_KEEPMAP_SLACK→DELTA_KEEPMAP_SLACK; preserve selection-versus-union behavior.C-upload:295; inventory records 76 zero-valued setters.
5.2.620190511COMPACT_KEEPMAP_SLOT_MULT→DELTA_KEEPMAP_SLOT_MULT; preserve slot-capacity multiplier and fallback rules.C-keep:106.
5.2.720190511COMPACT_UPLOAD_PEND→DELTA_UPLOAD_PEND; preserve pending-upload branch exactly.C-upload:77.
5.2.820190511CS_DRAFT_HOLD→rename DELTA_PRUNE_HOLD (alias CS_DRAFT_HOLD)cpu_verify_engine.py:2261 · holds drafting while a landing prune is pending
5.2.920190511CS_DRAFT_HOLD_MAX→rename DELTA_PRUNE_HOLD_MAXcpu_verify_engine.py:2264,2268-2273
5.2.1020190511CS_LANDING_PRUNE→rename DELTA_LANDING_PRUNEcpu_verify_engine.py:2082,2085
5.2.1120190511CS_PARITY_ASSERT→DELTA_ASSERT_SERIAL_REQUESTS; preserve explicit assertion, default off, independent of level 3.C-upload:110.
5.2.1220190511DIFF_NATIVE→DELTA_NATIVE; same native-versus-fallback diff planner choice.C-diff:97.
5.2.1320190511DIFF_PAD_SHORT_FATAL→DELTA_PAD_SHORT_FATAL; same explicit failure policy; never a verbosity effect.C-diff:148.
5.2.1420190511DIFF_PARTIAL_RELOAD→rename DELTA_PARTIAL_RELOAD (alias)temp_buffer.py:504-508
5.2.1520190511KEEPMAP_PAD_LAYERS→DELTA_KEEPMAP_PAD_LAYERS; same survivor-count padding behavior.C-keep:47.
5.2.1620190511KEEPMAP_STRICT_SLOTS→rename DELTA_KEEPMAP_STRICT_SLOTS (alias)keep_map_reader.py:28-30,164-165
5.2.1720190511RECOMPACT_SLOT_ALIGN→DELTA_RECOMPACT_SLOT_ALIGN; preserve slot-alignment behavior.C-index:58.
5.2.1820190511UPLOAD_SKIP_FATAL→rename DELTA_UPLOAD_SKIP_FATAL (alias)manager.py:174-180,594-599

5.3삭제: 이유 (1)

#주 작업자지금이유근거
5.3.120190511KEEPMAP_SLOT_LOG✕-keep_map_reader.py:157-163 · warn-once gate only inside if not _same:; keep the warning unconditional (audit s1 #15)
5.4그대로 유지 (2)
#주 작업자지금앞으로근거
5.4.120190511COMPACT_ACT_RESERVE_GIB→unchanged (runners set it)gpu_worker_sizing.py:19-21
5.4.220190511COMPACT_MAX_GPUBLK_UTIL→unchanged (runners set it 14x)scheduler_recompact.py:269-274; runner_diff_upload.py:173 (per-arm)

6GPU draft 엔진 (13)

draft 실행, 스케줄러, run-ahead, dispatch · 통합 4 · 이름 변경 3 · 그대로 유지 5 · 삭제 1

6.1통합: 지금 변수 → 새 knob · 레벨 (4)

#주 작업자지금앞으로근거
6.1.1wonhyukCOMPRESSION_SPEC_DEBUG→LOG_PRINT_DRAFT=3 (GPU sync), then delete the namegpu_model_runner.py:329,1266,2034-2038,3887; engine/core.py:91,444,505; single_type_kv_cache_manager.py:81,297,308,313,3…
6.1.220190511CS_STOP_LOG→LOG_PRINT_DRAFT=2sched/utils.py:93-100 · per request per step when near a cap
6.1.320190511DRAFT_DISPATCH_LOG→LOG_PRINT_DRAFT=2temp_buffer.py:701-712
6.1.420190511PRESERVE_LOG→LOG_PRINT_DRAFT=2temp_buffer.py:516-526 (RELOAD_LIST); cpu_verify_engine.py:1774-1784 (PRESERVE_RANGE) · one line per reload / per all-ac…

6.2이름 변경: 지금 이름 → 새 이름 (옛 이름 호환) (3)

#주 작업자지금앞으로근거
6.2.1wonhyukCOMPRESS_SPEC_SYNC_VERIFY→rename DRAFT_SYNC_VERIFY (alias)scheduler.py:1725-1736 · debug-only but behaviour-altering (serializes draft/verify); never a level
6.2.220190511COMPSPEC_DRAFT_STALL_S→rename DRAFT_STALL_Scpu_verify_engine.py:513-514
6.2.3ysshinCOMPSPEC_RUNAHEAD→rename DRAFT_RUNAHEAD (alias COMPSPEC_RUNAHEAD)cpu_verify_engine.py:414-428,1738-1747,2249

6.3삭제: 이유 (1)

#주 작업자지금이유근거
6.3.1wonhyukDUMMY_BATCH_DBG✕Remove capped dummy-batch log probe; retain the independent functional skip control.C-core:1373; Audit:140.
6.4그대로 유지 (5)
#주 작업자지금앞으로근거
6.4.1wonhyukCOMPRESSION_SPEC_DECODING_INTERVAL→unchanged; document as alias of COMPSPEC_DECODE_ATTN_INTERVALflashinfer.py:594-600 (fallback only); connector_config.py:57,75 · only reader is the flashinfer fallback -> the two nam…
6.4.2wonhyukCOMPRESSION_SPEC_MAX_TOTAL_CONTEXT→unchangedscheduler.py:323-325; scripts/longcontext/run_compressspec.sh:146
6.4.3ysshinCOMPSPEC_DECODE_ATTN_INTERVAL→unchanged (runners set it); canonical of the pairflashinfer.py:599
6.4.420190511DRAFT_DISPATCH_FULL→unchanged (runners pass it)temp_buffer.py:692-700,713 · fixed-window dispatch for parity runs; behaviour
6.4.520190511SPECULATIVE_RUN_STEP→SPECULATIVE_RUN_STEP; shared algorithm/protocol setting, including ledger shrink-window use.C-index:59, :122.

7CPU verify (67)

verify 엔진·서버, rejection sampling, logprob, 수락 규칙 · 통합 7 · 이름 변경 7 · 그대로 유지 46 · 삭제 7

7.1통합: 지금 변수 → 새 knob · 레벨 (7)

#주 작업자지금앞으로근거
7.1.1wonhyukCOMPRESS_SPEC_TRACE→LOG_PRINT_VERIFY=2 (optional category filter LOG_PRINT_VERIFY_CATS)_trace.py:41-42 (categories),58-59,141 · one multi-line logger.info per verify cycle; no sync
7.1.220190511DRAFT_TOKEN_LOG→LOG_PRINT_VERIFY=2cpu_verify_server_score.py:455-461 · hard-codes Qwen stop ids 151645/151643 in the line; fix when folding
7.1.320190511FSM_RUN_DIR→LOG_PRINT_VERIFY=2 (writer on) · -> CS_OUT/<run>/trace/fsm/<req>.jsonl_fsm_log.py:15-26; breakdown_util.py:62-68 derives it from FSM_DEBUG_DIR · open/append/close per transition; cache the h…
7.1.420190511LAST_VERIFY_LOG→LOG_PRINT_VERIFY=2_tracker_registry.py:141-147,170-176 · runners pass ${LAST_VERIFY_LOG:-0}
7.1.5ysshinVERIFY_BATCH_LOG→LOG_PRINT_VERIFY=1; new interface emits once per logical forward; legacy adapter preserves old emission behavior until consumers migrate.C-attn:228, :231; Audit:71, :107.
7.1.6wonhyukVERIFY_DEBUG→LOG_PRINT_VERIFY=2cpu_verify_server.py:131,357-360,523-525; cpu_verify_server_score.py:114 (read, never used) · prints decoded verified ta…
7.1.7wonhyukVERIFY_DEBUG_PAYLOAD→LOG_PRINT_VERIFY=2cpu_verify_server.py:409-420 · first 2 payloads only; print() not logger

7.2이름 변경: 지금 이름 → 새 이름 (옛 이름 호환) (7)

#주 작업자지금앞으로근거
7.2.1ysshinATTN_KV_RETAIN_DIV→VERIFY_ATTN_KV_RETAIN_DIV; preserve divisor parsing, clamping, and default 8.C-native:1631.
7.2.2ysshinATTN_TASK_BALANCE→VERIFY_ATTN_TASK_BALANCE; preserve kernel work-partition behavior.C-native:484.
7.2.3ysshinCOMPSPEC_BONUS_TOKEN→rename VERIFY_BONUS_TOKEN (alias COMPSPEC_BONUS_TOKEN)cpu_verify_server.py:290,287-289; verify/dataclass.py:67 · changes what the GPU appends; behaviour knob
7.2.4ysshinCOMPSPEC_SKIP_WATCH_SECS→rename VERIFY_SKIP_WATCH_Scpu_verify_engine.py:542-546 · log-only watchdog threshold (comment :542); still a timing parameter, not a level
7.2.5ysshinFORCE_ATTN_ISA→VERIFY_ATTN_ISA; preserve forced ISA and matching offload layout across both engines.G-offload:243, :247; supplied debug registry line 85.
7.2.6ysshinNUMA_AWARE_VERIFY_NODES→VERIFY_NUMA_NODES; preserve request-to-node hashing and node-count meaning.G-copy:419; G-offload:1072.
7.2.7wonhyukSKIP_DUMMY_BATCH→VERIFY_SKIP_DUMMY_BATCH; preserve explicit execution override, default off.C-core:1381.

7.3삭제: 이유 (7)

#주 작업자지금이유근거
7.3.1wonhyukCPU_ATTN_BUILD_DBG✕Remove capped PR4 attention-build probe.C-attn:195; Audit:140.
7.3.2wonhyukCPU_FWD_READ_HASH✕Remove capped forward-read hash probe.C-attn:442; Audit:140.
7.3.3wonhyukFWD_TENSOR_STATS✕Remove capped pre/post-cache-write tensor-stat probes.C-attn:484; Audit:140.
7.3.4ysshinLOGPROB_READ_DEBUG✕-cpu_verify_engine.py:3517-3524 · measurement aid for the 08-21 anchor fix; 0 setters (audit s3.2)
7.3.5ysshinLOGPROB_READ_LEGACY_ANCHOR✕-cpu_verify_engine.py:3478,3595 · legacy A/B path (dead_kind C)
7.3.6ysshinLOGPROB_READ_POS_OFFSET✕-cpu_verify_engine.py:3481 · only read inside the LEGACY_ANCHOR branch
7.3.7wonhyukPAYLOAD_BUILD_DBG✕env removed; emitter stays under LOG_PRINT_VERIFY=2cpu_verify_engine.py:2899 (or _cdbg.TRACE) · already gated by cspec_debug.TRACE; first-N counter
7.4그대로 유지 (46)
#주 작업자지금앞으로근거
7.4.120190511ACCEPT_MODE→unchanged (VERIFY domain; judge_debugger sets it)verify/utils.py:1306,1235-1244 · standard / dual / cactus / tas acceptance rule
7.4.220190511CACTUS_DELTA→unchangedverify/utils.py:1309,1239
7.4.3wonhyukCPU_VERIFY_TIMEOUT_S→unchanged_compaction_upload_dispatcher.py:99; _verify_dispatcher.py:125
7.4.4wonhyukCPU_VERIFY_URL→unchanged_compaction_upload_dispatcher.py:96; _verify_dispatcher.py:124; cpu_verify_engine.py:4178
7.4.520190511LOGPROB_POOL_SIZE→unchanged (shared contract GPU<->CPU)verify/utils.py:53; config/connector_config.py:48
7.4.620190511LOGPROB_SHM_NAME→unchangedverify/utils.py:51
7.4.7ysshinLOGPROB_SPARSE→unchangedlogprob_handler.py:25; logprob_offload_connector.py:88; verify/utils.py:59
7.4.8ysshinLOGPROB_SPARSE_CAP→unchangedlogprob_handler.py:27; verify/utils.py:61,509
7.4.9ysshinLOGPROB_VOCAB_SIZE→unchanged (runners set 151936)verify/utils.py:55; server :240 overrides from model
7.4.1020190511MAXPOS_OVERRIDE→unchangedcpu_verify_server.py:216; _score.py:142
7.4.11otherPORT→unchanged (optional alias VERIFY_PORT)cli.py:445,458; cpu_verify_server.py:756; _score.py:602
7.4.1220190511REJECT_DISTRIB_PT→unchanged (dump parameter: every N-th full-vocab .pt) · -> dump/accept/reject_pt/verify/utils.py:1301,1386-1394
7.4.1320190511REJECT_DISTRIB_PT_MAX→unchanged (dump parameter)verify/utils.py:1302,1394-1411
7.4.1420190511REJECT_DISTRIB_SOURCE→unchanged (dump parameter: server or both)verify/utils.py:1298,1337
7.4.1520190511RELAX_RESIDUAL→unchangedverify/utils.py:1308 · read but effect not in the read lines; UNCLEAR: read verify/utils.py around :1240-1260
7.4.1620190511RELAX_TAU→unchangedverify/utils.py:1307,1236
7.4.17ysshinRS_ALLOW_UNSUPPORTED_SAMPLING→unchangedchannels/temp_buffer.py:83 (GPU side: refuse vs warn)
7.4.18ysshinRS_FAST_TRUNC→unchangedverify/utils.py:620,689
7.4.19ysshinRS_FAST_TRUNC_MAX_K→unchangedverify/utils.py:623,689
7.4.20ysshinRS_KTH_CHUNKS→unchangedverify/utils.py:702,709
7.4.21ysshinRS_KTH_MIN_VOCAB→unchangedverify/utils.py:703
7.4.22ysshinRS_Q_FLOOR→unchangedverify/utils.py:874
7.4.23ysshinRS_ROW_MISMATCH→unchangedverify/utils.py:875,867-871 (fatal or reject)
7.4.24ysshinRS_STALE_MAX_RATE→unchangedverify/utils.py:608,1113
7.4.25ysshinRS_STALE_MIN_POS→unchangedverify/utils.py:609,1112
7.4.26ysshinRS_TIE_EPS→unchangedverify/utils.py:599,1181
7.4.27ysshinRS_TOPP_EPS→unchangedverify/utils.py:589,1144
7.4.28ysshinRS_TRUNCATE→unchangedverify/utils.py:583,987 (runners set it)
7.4.2920190511TAS_DELTA→unchangedverify/utils.py:1311,1244
7.4.3020190511TAS_EPS→unchangedverify/utils.py:1310,1244
7.4.31ysshinVERIFY_BATCH_WINDOW_MS→unchangedcpu_verify_server.py:139,168,353
7.4.32wonhyukVERIFY_DISABLE_HYBRID_KVCM→unchanged (behaviour: engine config)cpu_verify_server.py:229-233 · 0 setters; behaviour knob, must not fold into a level
7.4.33ysshinVERIFY_DISPATCH_WORKERS→unchangedcpu_verify_engine.py:459-464
7.4.34ysshinVERIFY_DP_SIZE→VERIFY_DP_SIZE; preserve verifier replica and NUMA-routing behavior.G-copy:417.
7.4.3520190511VERIFY_ENFORCE_EAGER→unchangedcpu_verify_server.py:119; _score.py:111
7.4.3620190511VERIFY_KV_PREFAULT→VERIFY_KV_PREFAULT; preserve optional background page-in behavior.C-runner:2977, :2984.
7.4.3720190511VERIFY_KV_PREFAULT_ARM_IDLE→VERIFY_KV_PREFAULT_ARM_IDLE; retain seconds and default 2.0.C-runner:3001.
7.4.38ysshinVERIFY_KV_SHM_RANK→VERIFY_KV_SHM_RANK; preserve override and parallel-rank fallback.C-runner:7347.
7.4.39wonhyukVERIFY_MAX_MODEL_LEN→unchangedcpu_verify_server.py:212; _score.py:139; cli.py:281-283
7.4.4020190511VERIFY_MAX_NUM_SEQS→unchangedcpu_verify_server.py:126; _score.py:113; profile_cpu_forward.py:278
7.4.4120190511VERIFY_MODEL→unchangedverify/cpu_verify_server.py:116; cpu_verify_server_score.py:109; cli.py:274
7.4.42ysshinVERIFY_NUMA_INTERLEAVE→unchanged (runners set 0)cpu_verify_server.py:35,55
7.4.43ysshinVERIFY_NUMA_LOCAL→unchanged (diagnostic A/B, but changes NUMA policy)cpu_verify_server.py:42-45
7.4.44wonhyukVERIFY_NUM_SPEC_BLOCKS→VERIFY_NUM_SPEC_BLOCKS; preserve speculative sliding-window allocation agreement.vllm_cpu/vllm/attention/layer.py:458.
7.4.4520190511VERIFY_PROMPT_LOGPROBS→unchangedcpu_verify_server.py:118,362
7.4.46ysshinVERIFY_TOKENIZER→unchangedcpu_verify_server.py:117; cpu_verify_server_score.py:110,46

8KV offload · pool · NUMA (37)

GPU↔CPU KV 이동, CPU pool, 공유 메모리, NUMA 배치 · 통합 1 · 이름 변경 11 · 그대로 유지 10 · 삭제 15

8.1통합: 지금 변수 → 새 knob · 레벨 (1)

#주 작업자지금앞으로근거
8.1.1ysshinCOMPSPEC_OFFLOAD_COVER→LOG_PRINT_KV=2cpu_verify_engine.py:2767-2792 · log-only invariant check per dispatch (comment :2767); no sync

8.2이름 변경: 지금 이름 → 새 이름 (옛 이름 호환) (11)

#주 작업자지금앞으로근거
8.2.120190511COMPACT_FULL_COPY_GB→rename KV_FULL_COPY_GB (alias)compression_spec_config.py:63-64
8.2.220190511COMPACT_TEMP_BUFFER_GB→rename KV_TEMP_BUFFER_GB (alias)compression_spec_config.py:61-62 · connector memory size, not a compaction knob
8.2.320190511COMPRESS_SPEC_CPU_POOL_HOLD→rename KV_CPU_POOL_HOLD (alias)cpu_verify_engine.py:1182,1248 · admission hold on CPU pool capacity; default on
8.2.4wonhyukCOMPRESS_SPEC_NO_FREE_SW→KV_NO_FREE_SW; explicit behavior-changing evidence hook, default off.C-single:602, :608; supplied debug registry line 86.
8.2.520190511COMPRESS_SPEC_PIN_FULLKV→KV_PIN_FULLKV; independent block-lifetime control, default off.C-cache:344.
8.2.620190511COMPSPEC_RELOAD_STALL_S→rename KV_RELOAD_STALL_Scpu_verify_engine.py:518-519
8.2.7ysshinCOMPSPEC_RESUME_NEED_MARGIN→rename KV_RESUME_NEED_MARGINcpu_verify_engine.py:400-401
8.2.8ysshinCOMPSPEC_RESUME_NEED_TIMEOUT_S→rename KV_RESUME_NEED_TIMEOUT_Scpu_verify_engine.py:402-403
8.2.9ysshinCPU_OFFLOAD_POOL_BLOCKS→KV_OFFLOAD_POOL_BLOCKS; preserve integer/all/unset semantics and GPU/CPU agreement.G-engine:2840; vllm/vllm/v1/kv_offload/cpu.py:78.
8.2.10ysshinOFFLOAD_PREFIX_CACHE→rename KV_OFFLOAD_PREFIX_CACHE (alias); removal candidatenaive_copy.py:400-410 · restores a documented-broken 2x-pool behaviour for A/B; 0 setters
8.2.11wonhyukRESERVE_SKIPPED_SW_BLOCKS→KV_RESERVE_SKIPPED_SW_BLOCKS; preserve explicit reservation behavior.C-single:198.

8.3삭제: 이유 (15)

#주 작업자지금이유근거
8.3.1wonhyukALLOC_NEW_COMPUTED_DBG✕Remove capped allocation probe and its counter.C-single:244; Audit:140.
8.3.220190511ATTN_BP_DEBUG_FILL✕- (delete the fill path from the real-run connector; keep debug_fill_slot_random for test/test_score.py)offloading_attn_bp_connector.py:336,348-356; handler.py:415-427 (real random fill, NOT a no-op); config/connector_config…
8.3.3wonhyukATTN_BP_DUMP_SCORE_CHECK✕-gpu_model_runner.py:6834-6849 (allocates L x T x H x S fp32 persist tensor) · .env=0 only; test-only cross-check
8.3.4wonhyukFREE_BLOCKS_DBG✕Remove block-ID-specific free probe.C-pool:514; Audit:140.
8.3.5wonhyukINPUT_PROCESSOR_DBG✕Remove capped input block-table probe.C-input:612; Audit:140.
8.3.6wonhyukKV_DST_PTR_DBG✕-cpu_gpu.py:262,779-790
8.3.720190511KV_DUMP_DIR✕Remove destination for the retired parity-dump family; coordinate paired writers with the other slice. Do not alias it to all KV dumps.G-temp:572; G-move:69; Audit:145.
8.3.820190511KV_DUMP_NCT✕-gpu_model_runner.py:866-911 (torch.save per step, .cpu()), :3934; temp_buffer.py:566-580; gpu_recompact.py:41-44 · 09-19…
8.3.920190511KV_DUMP_SEQ✕Remove request-prefix/sequence-range parity dump and its tensor reads.C-runner:5322, :5415; Audit:145.
8.3.10wonhyukKV_TRANSFER_DBG✕-offloading_naivecopy_connector.py:335 · PR4 item 8 evidence, first-10
8.3.11wonhyukNAIVE_STORE_DBG✕-offloading_naivecopy_connector.py:471
8.3.12ysshinNUMA_VERIFY_DEBUG✕-temp_cpu.py:80-88 · print() every 10000 allocs; 0 setters
8.3.13wonhyukRESERVE_DBG✕Remove block-ID-specific reserve probe.C-pool:386; Audit:140.
8.3.14wonhyukSAVE_KV_LAYER_DBG✕-offloading_naivecopy_connector.py:2,431-445
8.3.15wonhyukTEMPBUF_DBG✕-temp_buffer.py:666-673
8.4그대로 유지 (10)
#주 작업자지금앞으로근거
8.4.120190511ATTN_BP_STREAM_CAP→unchangedattn_byproduct_handler/handler.py:286-298 · stream/event pool cap; resource limit
8.4.2ysshinCOMPSPEC_PREEMPT_OFFLOAD→unchanged (runners set it); alias KV_PREEMPT_OFFLOADcpu_verify_engine.py:2985; manager.py:502; scheduler.py:1319
8.4.3ysshinCOMPSPEC_RESUME_NEED_GATE→unchanged (runners set it); alias KV_RESUME_NEED_GATEcpu_verify_engine.py:398-399
8.4.420190511MAX_NUM_SEQS→MAX_NUM_SEQS; shared capacity input, including existing fallback/headroom rules.C-keep:88.
8.4.5ysshinNUMA_FIRST_TOUCH→NUMA_FIRST_TOUCH; existing subsystem prefix is adequate; preserve placement behavior.G-offload:1073.
8.4.6ysshinNUMA_FT_NODES→NUMA_FT_NODES; preserve node-list parsing and unset behavior.C-runner:7469.
8.4.720190511NUM_GPU_BLOCKS→NUM_GPU_BLOCKS; shared GPU-capacity input, not a DELTA-only budget.C-upload:150; inventory records nine reads.
8.4.8wonhyukSHARED_CPU_POOL_GB→unchangedcli.py:218-236; connector_config.py:45-46
8.4.920190511VLLM_SHM_NO_INTERLEAVE→unchangedconfig/common_util.py:185-189 (both processes)
8.4.1020190511VLLM_SHM_SUFFIX→VLLM_SHM_SUFFIX; preserve cross-process namespace and UID fallback.vllm/vllm/shm_ns.py:13.

9prefill (layer-serial · HeadInfer) (16)

긴 프롬프트의 chunk prefill, streaming, HeadInfer decode · 통합 2 · 이름 변경 1 · 그대로 유지 13

9.1통합: 지금 변수 → 새 knob · 레벨 (2)

#주 작업자지금앞으로근거
9.1.1ysshinLS_CHUNKED_VERIFY→LOG_PRINT_PREFILL=3 (GPU sync)flashinfer.py:1876-1878,2826-2840 (.item(); runs the single causal wrapper again)
9.1.2ysshinLS_PEAK_PROBE→LOG_PRINT_PREFILL=1scheduler.py:1912-1921 · one line per free event; no sync

9.2이름 변경: 지금 이름 → 새 이름 (옛 이름 호환) (1)

#주 작업자지금앞으로근거
9.2.1ysshinHI_NTB→rename HEADINFER_NTB (alias HI_NTB)headinfer_decode.py:148
9.3그대로 유지 (13)
#주 작업자지금앞으로근거
9.3.1ysshinHEADINFER_DECODE→unchangedheadinfer_decode.py:39
9.3.2ysshinHEADINFER_DECODE_REAL→unchangedheadinfer_decode.py:41
9.3.3ysshinLAYER_SERIAL_BARRIER→unchangedcpu_verify_engine.py:4279-4291 (stream or host)
9.3.4ysshinLAYER_SERIAL_CHUNKED→unchangedconfig/scheduler.py:307; flashinfer.py:90; scheduler.py:203; gpu_model_runner.py:6483
9.3.5ysshinLAYER_SERIAL_COVERAGE_BLOCKS→unchangedkv_cache_utils.py:866-873
9.3.6ysshinLAYER_SERIAL_K→unchangednaive_copy.py:371; cpu_verify_engine.py:1364,4277; kv_cache_utils.py:870,1550; single_type_kv_cache_manager.py:34; flash…
9.3.7ysshinLAYER_SERIAL_TAIL_BLOCKS→unchangedcpu_verify_engine.py:1365; single_type_kv_cache_manager.py:36; flashinfer.py:1055
9.3.8ysshinLS_CHUNK_TILE→unchangedgpu_model_runner.py:6486
9.3.9ysshinLS_PIPELINE_RELOAD→unchangedflashinfer.py:2040-2045
9.3.10ysshinLS_RELOAD_MEMCPY→unchangedflashinfer.py:2093-2094
9.3.11ysshinLS_STREAM_FREE→unchangedflashinfer.py:93; scheduler.py:206
9.3.12ysshinLS_STREAM_RELOAD→unchangedflashinfer.py:92; scheduler.py:205
9.3.13ysshinLS_STREAM_STAGE→unchangedflashinfer.py:91; scheduler.py:204

10모드 전환(router) (4)

resident / speculative 모드 전환 · 이름 변경 1 · 그대로 유지 3

10.1이름 변경: 지금 이름 → 새 이름 (옛 이름 호환) (1)

#주 작업자지금앞으로근거
10.1.1ysshinDYN_FORCE_OUT→rename DYN_ROUTER_FORCE_OUT (alias DYN_FORCE_OUT)dyn_router.py:24-25,61 · dev knob; runners pass $LC_DYN_FORCE_OUT
10.2그대로 유지 (3)
#주 작업자지금앞으로근거
10.2.1ysshinDYN_ROUTER→unchangeddyn_router.py:51; cpu_verify_engine.py:536-541
10.2.2ysshinDYN_ROUTER_HEADROOM_PCT→unchangeddyn_router.py:60
10.2.3ysshinDYN_ROUTER_HOLD_STEPS→unchangeddyn_router.py:69

11출력 · 덤프 · 시간 분해 (44)

로그 파일, 덤프, breakdown, 출력 경로 · 통합 18 · 그대로 유지 9 · 삭제 17

11.1통합: 지금 변수 → 새 knob · 레벨 (18)

#주 작업자지금앞으로근거
11.1.120190511ATTN_KV_DUMP_DIR→DUMP=kv · -> CS_OUT/<run>/dump/kv/<req>_j<job>_<postfix>.jsondebug_utils.py:98-104 (called unconditionally scheduler.py:1760,1819); cpu_verify/utils.py:586-597 · metadata/tracker JS…
11.1.220190511BREAKDOWN_DIR→BREAKDOWN=rows plus CS_OUT; retain old directory/timestamp behavior in legacy mode.vllm/vllm/breakdown_logger.py:149; Audit:98, :165.
11.1.320190511BREAKDOWN_KERNEL_INLINE→Jointly with profile mode → BREAKDOWN_KERNEL; (1,1) → inline+gpu-profiler, (1,0) → profiler. INLINE alone does not activate profiling.C-timer:17; root breakdown_debugger/breakdown_util.py:97, :149.
11.1.420190511BREAKDOWN_KERNEL_PROFILE_MODE→BREAKDOWN_KERNEL; zero disables this timing family; one uses the joint translation above.C-timer:17; root breakdown_debugger/breakdown_util.py:97, :147.
11.1.520190511BREAKDOWN_LIGHT_LAT→BREAKDOWN=light · -> breakdown/lat_<role>.jsonbreakdown_logger.py:105,208-239 · 0 in-tree setters (audit s2)
11.1.620190511BREAKDOWN_RUN_DIR→derived: CS_OUT_RUN/breakdown (launcher no longer exports it) · -> CS_OUT/<run>/breakdown/{sched,step_<role>,cpu_verify}.jsonlbreakdown_logger.py:147-154; breakdown_util.py:94
11.1.720190511COMPACT_HOLD_DUMP→Fingerprints → DELTA 3, trace/held.jsonl; full records → explicit DUMP=metadata. Legacy explicit path retains its narrow enablement and exact destination.C-index:91, :109, :117, :125.
11.1.8wonhyukCOMPRESS_SPEC_DEBUG→LOG_PRINT=<level> (global default for all LOG_PRINT_<DOMAIN>)cspec_debug.py:21-36; readers: handler.py:327 (==1), logprob_worker.py:226 (==1), verify/utils.py:429 (==1), cpu_gpu.py:…
11.1.9wonhyukCOMPRESS_SPEC_TRACE_REQ→LOG_PRINT_REQ=<req_id> (scheme addition, see s3)_trace.py:60,120
11.1.10wonhyukCPU_LOG→derived; override removed · -> CS_OUT/<run>/logs/cpu_server.logcli.py:449 · 0 setters
11.1.11wonhyukCPU_VERIFY_DUMP_DIR→DUMP=memcheck · -> CS_OUT/<run>/dump/memcheck/cpu_verify/utils.py:649-669; cpu_verify_engine.py:382-386,2707-2715,3848 · enabled today by COMPRESS_SPEC_DEBUG>=3 or co…
11.1.1220190511DEBUG_METADATA_DUMP_DIR→DUMP=accept plus CS_OUT; preserve separately gated extensions and legacy directory semantics.Root verify/utils.py:1288, :1592, :1616; Audit:64, :100.
11.1.13wonhyukGPU_LOG→derived; override removed · -> CS_OUT/<run>/logs/gpu_client.logcli.py:450 · 1 setter (test)
11.1.1420190511LOGPROB_DUMP_DIR→DUMP=logprob · -> CS_OUT/<run>/dump/logprob/step{N}_{ts}.ptverify/utils.py:85,206,227; reader dump_debugger/logprobs_debugger.py:29 · audit s3.2 lists it under REMOVE; kept becaus…
11.1.15wonhyukLOG_DIR→CS_OUT (LOG_DIR accepted as deprecated alias) · -> CS_OUT/<run>/logs/cli.py:447; test/ruler/run_commit_bench.sh:86 sets per cell
11.1.1620190511OUT_ROOT→CS_OUT for migrated engine output routing; legacy held-output fallback remains <OUT_ROOT>/held.jsonl. Path only, no enablement.C-index:92.
11.1.1720190511REJECT_DISTRIB_DUMP→DUMP=reject (reject-only) or DUMP=reject_all · -> CS_OUT/<run>/dump/accept/reject_distrib.jsonlverify/utils.py:1295,1332-1337; cpu_verify_server_score.py:239-246 · today needs DEBUG_METADATA_DUMP_DIR too (:1333 ACCE…
11.1.18wonhyukSCORE_DUMP_DIR→DUMP=score · -> CS_OUT/<run>/dump/score/gpu_model_runner.py:3478-3500; offloading_attn_bp_connector.py:323-327 (torch.cuda.synchronize per step),401-409; cpu_ve…

11.2삭제: 이유 (17)

#주 작업자지금이유근거
11.2.120190511BREAKDOWN_FLUSH_EVERY✕-breakdown_logger.py:88,254-255 vs :102 buffering=1 · redundant with line buffering
11.2.220190511BREAKDOWN_GPU_EVENT_DEPTH✕- (hard-code 4)breakdown_logger.py:90-91 · 0 setters
11.2.320190511BREAKDOWN_PER_REQ✕-breakdown_logger.py:89 · 0 setters
11.2.420190511COMPACTION_IDX_DUMP✕Remove obsolete full selected-index writer. Do not silently fold it into DUMP=accept.C-runner:4193; Audit:150 identifies its limited historical consumer.
11.2.520190511COMPACTION_IDX_EVERY✕Remove sampling modifier with its writer.C-runner:4204; Audit:150.
11.2.620190511COMPACTION_IDX_LAYERS✕Remove layer filter with its writer.C-runner:4207; Audit:150.
11.2.7wonhyukCOMPRESS_SPEC_CYCLE_ANNOTATE✕-cpu_verify_engine.py:3704-3719 · writes hard-coded .realmode-run/cycle_map.jsonl; 'uncommitted' comment; DUMP=accept alr…
11.2.8wonhyukDUMP_EXTEND_BEFORE_VERIFY✕-scheduler.py:2303-2312 · PR4 first-8; 1 setter (jh_script)
11.2.9wonhyukDUMP_FORWARD_OUT✕-gemma3.py:373-380
11.2.10wonhyukDUMP_LAST_HIDDEN✕-gemma3.py:532-547 (.item()/.cpu() sync)
11.2.11wonhyukDUMP_LOGITS_OUT✕-gemma3.py:550-556
11.2.12wonhyukDUMP_SAMPLED_IDS✕-gpu_model_runner.py:4213-4222 (.tolist() sync, first 32)
11.2.13wonhyukDUMP_TRACKER_DISPATCH✕-temp_buffer.py:907-915
11.2.1420190511INIT_FILE_PATH✕-cpu_verify/utils.py:663-665 · base for a relative CPU_VERIFY_DUMP_DIR; CS_OUT is absolute
11.2.15wonhyukPR4_8_DBG_EXTENDED✕-naive_copy.py:327 · NOTE the [PR4-8-DBG] emitter at :321-335 is unconditional for the first 2 calls; delete it with the …
11.2.16wonhyukPR4_DISABLE_COMPRESS_CALL✕-flashinfer.py:2874 · behaviour-altering probe; 0 setters
11.2.17wonhyukPR4_SW_GUARD_DEBUG✕-flashinfer.py:1841-1847
11.3그대로 유지 (9)
#주 작업자지금앞으로근거
11.3.1wonhyukBREAKDOWN_SKIP_WARMUP→unchangedbreakdown_logger.py:158-160; breakdown_util.py:93
11.3.220190511CPU_PROFILE_CHUNK→unchanged (tool-local knob of verify/profile_cpu_forward.py) · run dir -> CS_OUT/<run>/breakdown/ (today $BREAKDOWN_DIR/B{n}_{DTYPE}_CPU/<ts>)profile_cpu_forward.py:47,286
11.3.320190511CPU_PROFILE_CHUNK_ITERS→unchanged (tool-local knob of verify/profile_cpu_forward.py) · run dir -> CS_OUT/<run>/breakdown/ (today $BREAKDOWN_DIR/B{n}_{DTYPE}_CPU/<ts>)profile_cpu_forward.py:48
11.3.420190511CPU_PROFILE_COMPACT_RATE→unchanged (tool-local knob of verify/profile_cpu_forward.py) · run dir -> CS_OUT/<run>/breakdown/ (today $BREAKDOWN_DIR/B{n}_{DTYPE}_CPU/<ts>)profile_cpu_forward.py:64-66,82,314
11.3.520190511CPU_PROFILE_CTX_TRUNC→unchanged (tool-local knob of verify/profile_cpu_forward.py) · run dir -> CS_OUT/<run>/breakdown/ (today $BREAKDOWN_DIR/B{n}_{DTYPE}_CPU/<ts>)profile_cpu_forward.py:292-294
11.3.620190511CPU_PROFILE_MAX_TOKENS→unchanged (tool-local knob of verify/profile_cpu_forward.py) · run dir -> CS_OUT/<run>/breakdown/ (today $BREAKDOWN_DIR/B{n}_{DTYPE}_CPU/<ts>)profile_cpu_forward.py:35
11.3.720190511CPU_PROFILE_PROMPT_LOGPROBS→unchanged (tool-local knob of verify/profile_cpu_forward.py) · run dir -> CS_OUT/<run>/breakdown/ (today $BREAKDOWN_DIR/B{n}_{DTYPE}_CPU/<ts>)profile_cpu_forward.py:52-53
11.3.820190511REJECT_DISTRIB_TOPK→unchanged (dump parameter of DUMP=reject)verify/utils.py:1296,1353
11.3.920190511TOKEN_DUMP_JSON→unchanged (path-valued; see s3) · optional default CS_OUT/<run>/dump/tokens/<force_name>.json when DUMP=tokenscommon_util.py:127-140; parity scripts 31_parity_greedy.sh:82, 35_fullonly_vs_vanilla.sh:81 name A/B files explicitly

12실행 · 하니스 (37)

모델 경로, 벤치 입력, CLI, 환경 이름 · 이름 변경 1 · 그대로 유지 34 · 삭제 2

12.1이름 변경: 지금 이름 → 새 이름 (옛 이름 호환) (1)

#주 작업자지금앞으로근거
12.1.120190511ARM→CS_RUN_TAG; opaque run/artifact label. Preserve old ARM acceptance and existing harness meaning; do not confuse it with per-request policy dictionaries.G-temp:575; C-runner:5436; inventory records 21 setters.

12.2삭제: 이유 (2)

#주 작업자지금이유근거
12.2.1wonhyukFAKE_COMPACTION_VERIFY✕Remove obsolete synthetic compaction mutation path.C-runner:5478, :5484; Audit:143.
12.2.2wonhyukFAKE_NEEDS_KV_RELOAD✕-_fake_verify.py:242-253 · gpu_only fake-verify toggle; 0 setters
12.3그대로 유지 (34)
#주 작업자지금앞으로근거
12.3.120190511BENCH_BATCH→unchanged; canonical over RULER_BATCH (category: functional)common_util.py:38,46-48 · CSV called it log/debug; it selects the prompt batch
12.3.220190511BENCH_CTX→unchanged; canonical over RULER_CTX (category: functional)common_util.py:49
12.3.3otherBLOCK_SIZE→unchangedconnector_config.py:38; cpu_verify_server.py:120; _score.py:112; profile_cpu_forward.py:33
12.3.4wonhyukCOMPRESS_SPEC_DRAFT_ENV→unchanged_tag.py:28
12.3.5wonhyukCOMPRESS_SPEC_NO_AUTOBUILD→unchangedcli.py:73
12.3.6wonhyukCOMPRESS_SPEC_NO_CLEAN→unchangedcli.py:629
12.3.7wonhyukCOMPRESS_SPEC_SKIP_CHECK→unchanged (category: functional, not log/debug)compressspec/_verify.py:93,182 · bypasses the build-stamp guard
12.3.8wonhyukCOMPRESS_SPEC_TAG→unchanged (COMPRESSSPEC_TAG already a deprecated alias)compressspec/_tag.py:27
12.3.9wonhyukCOMPRESS_SPEC_TARGET→unchanged_verify.py:161
12.3.10wonhyukCOMPRESS_SPEC_VERIFY_ENV→unchanged_tag.py:29
12.3.11wonhyukEXAMPLE_DIAG_QID→unchanged (DIAG_QID is its alias in test/example.py:218)cli.py:254
12.3.12wonhyukEXP_DTYPE→unchangedcli.py:241,264; connector_config.py:40; servers
12.3.1320190511GPU_CORES→unchangedconfig/common_util.py:309-313
12.3.1420190511MAX_MODEL_LEN→unchanged (profile tool / test harness max_model_len)profile_cpu_forward.py:36,277
12.3.15wonhyukMINIFORGE_DIR→unchangedcli.py:49
12.3.1620190511MODEL_DIR→unchangedcli.py:274; connector_config.py:31
12.3.1720190511MULTIBP_COMPRESS→unchangedconnector_config.py:55
12.3.1820190511MULTIBP_MAX_MODEL_LEN→unchangedconnector_config.py:53
12.3.1920190511MULTIBP_MAX_NUM_SEQS→unchangedconnector_config.py:52
12.3.2020190511MULTIBP_QUESTION→unchangedcommon_util.py:27
12.3.2120190511MULTIBP_REPEAT→unchangedcommon_util.py:29
12.3.22wonhyukREADY_TIMEOUT→unchangedcli.py:446
12.3.23ysshinRULER_BATCH→unchanged (fallback alias of BENCH_BATCH)common_util.py:38,48
12.3.24ysshinRULER_CTX→unchanged (fallback alias of BENCH_CTX)common_util.py:49
12.3.25ysshinRULER_DYNAMIC→unchangedcommon_util.py:54
12.3.2620190511RULER_OFFSET→unchangedcommon_util.py:68,84
12.3.2720190511RULER_QDIR→unchangedcommon_util.py:67,81
12.3.2820190511RULER_SEQ→unchangedcommon_util.py:39,101; profile_cpu_forward.py:242
12.3.2920190511RULER_SUFFIX→unchangedcommon_util.py:82
12.3.30ysshinRULER_TASK→unchangedcommon_util.py:57,102; profile_cpu_forward.py:249
12.3.3120190511SEED→unchangedconnector_config.py:51
12.3.3220190511SMALL_MODEL_DIR→unchangedconnector_config.py:29-32
12.3.3320190511TASK_NAME→unchangedcommon_util.py:53
12.3.34wonhyukVLLM_ATTENTION_BACKEND→unchanged -- upstream vLLM variable, out of scopevllm/vllm/config/attention.py:93 (upstream); connector_config.py:71,74

13기타 (4)

위에 넣기 어려운 것 · 삭제 4

13.1삭제: 이유 (4)

#주 작업자지금이유근거
13.1.1wonhyukCOMPACT_DUMP_INDICES✕Remove one-shot index-layout probe.C-runner:4298; Audit:142.
13.1.2wonhyukCOMPACT_DUMP_KV✕Remove one-shot KV-layout probe.C-runner:4464; Audit:142.
13.1.3wonhyukCPU_KV_HASH_DBG✕Remove issue-specific per-block hash probe; update its one recorded diagnostic setter.C-runner:5071; Audit:67, :141.
13.1.4wonhyukREUSE_ENGINE_RID✕-cpu_verify_server.py:499-508 · PR4 item-8 hypothesis toggle (comment :499); 0 setters