1vLLM 0.14.0 대비 엔진에 추가된 환경변수
생성 2026-09-30 10:18 · 기준 커밋 v0.14.0 (b17039bcc) · 엔진 = vllm · vllm_cpu · verify · compressspec · config · 엔진 286개 (실험 하니스 194개는 맨 아래 접힘) · vLLM 원래 변수와 HF_·CUDA_·OMP_ 등 외부 변수는 제외
1.1
82
기능 · 실험에서 설정함
1.2
116
기능 · 기본값으로만 씀
1.3
27
로그·디버그 · 실험에서 설정함
1.4
61
로그·디버그 · 설정 안 함
1.5분류 기준 (펼치기)
- 로그·디버그: 이름에 LOG · PRINT · DEBUG · DUMP · TRACE · STAT · PROFILE · BREAKDOWN · TIMING · CHECK · FAKE · TEST 등이 들어간 것(이름 기준 자동 분류라 예외가 있을 수 있음)
- 기능: 그 외. 동작(선택 규칙·전송·스케줄·메모리)을 바꾸는 스위치
- 실험에서 설정함: 001 scripts · .env · test · jh_script, ysshin scripts · .env 중 한 곳이라도
NAME=로 값을 준 것 - 추가 기준: vllm · vllm_cpu 는 v0.14.0 에서 읽지 않던 이름만. verify · compressspec 등 우리 코드는 전부
정리 계획 문서 → 카테고리별: 지금 → 통합 · 이름 변경 · 삭제 ↗
1.6
1.7안 쓰는 환경변수 후보 (48)
코드는 아직 고치지 않았다. 근거 열의 커밋·파일로 확인한 뒤 지운다. A 설정만 남음 38 · C 폐기(옛 이름) 8 · C 폐기(LEGACY) 2
판정 기준 (펼치기)
- A 설정만 남음: 러너·.env 가 export·.env·NAME=값 명령 으로 주는데 지금 코드(엔진·하니스·jh_script·third-party) 어디서도 읽지 않음. 마지막 변경 = 그 이름이 코드에서 마지막으로 바뀐(지워진) 커밋
- B 죽은 코드에서만 읽음: 읽는 파일이 모두 같은 repo 어디서도 import 되지 않음(스크립트·C++ 에서 읽는 이름은 판정 보류)
- C 폐기: 새 이름으로 바뀐 옛 이름(_envcompat), 이름에 LEGACY·DEPRECATED
| # | 종류 | 이름 | 지금 읽는 곳 수 | 위치(설정한 곳·읽는 파일) | 설정 횟수 | 마지막 변경 커밋 | 근거 |
|---|---|---|---|---|---|---|---|
| 1.7.1 | A 설정만 남음 | BREAKDOWN_DUMP_DIR | 0 | .env | 1 | 001:d08c25e 2026-07-01 feat: compaction-threshold sweep harness + no-coldstart breakdown plott | 지금 엔진(001·ysshin dev)·하니스·third-party·g07 010 트리 어디서도 읽지 않음. 러너에만 남은 설정 |
| 1.7.2 | A 설정만 남음 | BREAKDOWN_TAG_COMPACT | 0 | jh_script/claude_script/27_g_curve_skipk.sh | 1 | 001:d08c25e 2026-07-01 feat: compaction-threshold sweep harness + no-coldstart breakdown plott | 지금 엔진(001·ysshin dev)·하니스·third-party·g07 010 트리 어디서도 읽지 않음. 러너에만 남은 설정 |
| 1.7.3 | A 설정만 남음 | CLIENT_PY | 0 | jh_script/claude_script/rtx_pro6000_sweep/script/nebius_arm_run.sh jh_script/claude_script/rtx_pro6000_sweep/script/nebius_arm_run2.sh | 2 | 지금 엔진(001·ysshin dev)·하니스·third-party·g07 010 트리 어디서도 읽지 않음. 러너에만 남은 설정 | |
| 1.7.4 | A 설정만 남음 | COMPACT_NO_STACK | 0 | ysshin:.env | 1 | 001:cd24516 2026-06-06 Chore: Add Benchmark and compaction debug script Add acceptance-sweep + | 지금 엔진(001·ysshin dev)·하니스·third-party·g07 010 트리 어디서도 읽지 않음. 러너에만 남은 설정 |
| 1.7.5 | A 설정만 남음 | COMPACT_RECOMPACT_ON_FULL | 0 | jh_script/claude_script/runners/run_aime2.sh | 1 | 001:ce27b98 2026-09-19 docs: hand-off note for delta transfer, and bump both submodules | 지금 엔진(001 feature·ysshin dev)은 읽지 않음. g07 010 옛 트리에서만 읽음: vllm/vllm/cpu_verify/compaction_upload/scheduler_recompact.py |
| 1.7.6 | A 설정만 남음 | COMPACT_RECOMPACT_RATIO | 0 | jh_script/claude_script/35_fullonly_vs_vanilla.sh scripts/saturate/lib/sweep_engine.sh ysshin:scripts/saturate/lib/sweep_engine.sh | 3 | 001:af64fb0 2026-09-19 saturate: forward the delta-transfer knobs | 지금 엔진(001 feature·ysshin dev)은 읽지 않음. g07 010 옛 트리에서만 읽음: vllm/vllm/cpu_verify/compaction_upload/scheduler_recompact.py vllm_cpu/vllm/compaction/runner_diff_upload.py |
| 1.7.7 | A 설정만 남음 | COMPACT_SCORE_RECOMPUTE | 0 | jh_script/claude_script/e2e_ab.sh jh_script/claude_script/e2e_ab2.sh jh_script/claude_script/e2e_node.sh jh_script/claude_script/e2e_on3.sh | 9 | 지금 엔진(001·ysshin dev)·하니스·third-party·g07 010 트리 어디서도 읽지 않음. 러너에만 남은 설정 | |
| 1.7.8 | A 설정만 남음 | DUMP_DIR | 0 | .env jh_script/claude_script/27_g_curve_skipk.sh | 2 | 001:b310946 2026-09-05 top_p/top_k long-context verify: bf16+sparse logprob relay, stale-row c | 지금 엔진(001 feature·ysshin dev)은 읽지 않음. g07 010 옛 트리에서만 읽음: vllm/vllm/cpu_verify/cspec_debug.py |
| 1.7.9 | A 설정만 남음 | EXP_DECODE_ATTN_INTERVAL | 0 | scripts/saturate/lib/sweep_engine.sh ysshin:scripts/saturate/lib/sweep_engine.sh | 2 | 001:3367962 2026-09-08 saturate: do not track the editor backup | 지금 엔진(001·ysshin dev)·하니스·third-party·g07 010 트리 어디서도 읽지 않음. 러너에만 남은 설정 |
| 1.7.10 | A 설정만 남음 | HELD_ORDER_DIAG | 0 | jh_script/claude_script/runners/run_ab.sh | 1 | 지금 엔진(001·ysshin dev)·하니스·third-party·g07 010 트리 어디서도 읽지 않음. 러너에만 남은 설정 | |
| 1.7.11 | A 설정만 남음 | HI_GATE_MODEL | 0 | scripts/longcontext/Qwen3-30B-A3B-Thinking-2507/01_gates_tiny.sh scripts/longcontext/Qwen3-30B-A3B-Thinking-2507/02_gates_real.sh ysshin:scripts/longcontext/Qwen3-30B-A3B-Thinking-2507/01_gates_tiny.sh ysshin:scripts/longcontext/Qwen3-30B-A3B-Thinking-2507/02_gates_real.sh | 4 | 001:5682a9a 2026-09-06 longcontext: Qwen3-30B-A3B (MoE) harness + cloud sm_120 5-way | 지금 엔진(001·ysshin dev)·하니스·third-party·g07 010 트리 어디서도 읽지 않음. 러너에만 남은 설정 |
| 1.7.12 | A 설정만 남음 | INDEXMAP_SHORTDRAFT_SHRINK_KEEP_HELD | 0 | jh_script/claude_script/runners/probe_util.sh jh_script/claude_script/runners/run_ab.sh jh_script/claude_script/runners/run_aime.sh jh_script/claude_script/runners/run_aime2.sh | 54 | 지금 엔진(001·ysshin dev)·하니스·third-party·g07 010 트리 어디서도 읽지 않음. 러너에만 남은 설정 | |
| 1.7.13 | A 설정만 남음 | KEEP_SPREAD_DIAG | 0 | jh_script/claude_script/runners/run_ks.sh | 1 | vllm_cpu:17132da6e 2026-09-19 feat(compaction): self-describing keep-map SHM and a correct held led | 지금 엔진(001·ysshin dev)·하니스·third-party·g07 010 트리 어디서도 읽지 않음. 러너에만 남은 설정 |
| 1.7.14 | A 설정만 남음 | KV_OFFLOAD_COALESCE | 0 | jh_script/claude_script/kv_coalesce_verify.sh jh_script/claude_script/kv_determinism_check.sh | 2 | 지금 엔진(001·ysshin dev)·하니스·third-party·g07 010 트리 어디서도 읽지 않음. 러너에만 남은 설정 | |
| 1.7.15 | A 설정만 남음 | MULTIBP_SINGLE | 0 | jh_script/claude_script/jh_q5q6.sh jh_script/claude_script/jh_q5x16.sh jh_script/claude_script/jh_q5xN.sh jh_script/claude_script/jh_revert_2tests.sh | 4 | 001:80a874b 2026-06-20 refactor: extract config/common helpers + always-on CPU-verify profilin | 지금 엔진(001·ysshin dev)·하니스·third-party·g07 010 트리 어디서도 읽지 않음. 러너에만 남은 설정 |
| 1.7.16 | A 설정만 남음 | POOLGB_256K | 0 | jh_script/claude_script/runners/run_rs256.sh jh_script/claude_script/runners/run_rs256b.sh jh_script/claude_script/runners/run_rs3.sh | 3 | 지금 엔진(001·ysshin dev)·하니스·third-party·g07 010 트리 어디서도 읽지 않음. 러너에만 남은 설정 | |
| 1.7.17 | A 설정만 남음 | PRESERVE_DRAFT_ON_ALLACCEPT | 0 | .env | 1 | 001:71771cc 2026-06-11 docs: debug knob registry (P0) + unified-levels and probe-status update | 지금 엔진(001 feature·ysshin dev)은 읽지 않음. g07 010 옛 트리에서만 읽음: vllm/tests/CPUVerify/test_engine_aftermath.py |
| 1.7.18 | A 설정만 남음 | QSTART | 0 | jh_script/claude_script/runners/probe_util.sh jh_script/claude_script/runners/run_bn.sh jh_script/claude_script/runners/run_ff.sh jh_script/claude_script/runners/run_fx2_all.sh | 12 | 001:0a6f35c 2026-09-19 feat(verify): score-server handshake, per-request arms, and the sweep d | 지금 엔진(001·ysshin dev)·하니스·third-party·g07 010 트리 어디서도 읽지 않음. 러너에만 남은 설정 |
| 1.7.19 | A 설정만 남음 | REQS_OVERRIDE | 0 | jh_script/claude_script/runners/probe_util.sh jh_script/claude_script/runners/retry_128k_full.sh jh_script/claude_script/runners/run_bn.sh jh_script/claude_script/runners/run_bprobe.sh | 43 | 지금 엔진(001·ysshin dev)·하니스·third-party·g07 010 트리 어디서도 읽지 않음. 러너에만 남은 설정 | |
| 1.7.20 | A 설정만 남음 | SSPEC_ACCEPT_CHECK | 0 | jh_script/claude_script/10_run_sparsespec_short_pod.sh jh_script/claude_script/18_dryrun_l40s.sh | 2 | SparseSpec 소스 계측용으로 보임. SparseSpec 은 클라우드 트리에 있어 이 검사로 확인 불가(노드 접속 불가). 우리 엔진·하니스·dev 트리는 읽지 않음 | |
| 1.7.21 | A 설정만 남음 | SSPEC_BOUND_TRACE | 0 | jh_script/claude_script/_sspec_bound.sh jh_script/claude_script/_sspec_boundfix.sh jh_script/claude_script/_sspec_full.sh jh_script/claude_script/_sspec_valid.sh | 5 | SparseSpec 소스 계측용으로 보임. SparseSpec 은 클라우드 트리에 있어 이 검사로 확인 불가(노드 접속 불가). 우리 엔진·하니스·dev 트리는 읽지 않음 | |
| 1.7.22 | A 설정만 남음 | SSPEC_BOUND_TRACE_PATH | 0 | jh_script/claude_script/32_sspec_repro.sh | 1 | SparseSpec 소스 계측용으로 보임. SparseSpec 은 클라우드 트리에 있어 이 검사로 확인 불가(노드 접속 불가). 우리 엔진·하니스·dev 트리는 읽지 않음 | |
| 1.7.23 | A 설정만 남음 | SSPEC_ITER_TRACE | 0 | jh_script/claude_script/25_spec_cycle_probe.sh jh_script/claude_script/_sspec_trace.sh | 2 | 001:1d9e75c 2026-09-21 chore: short context tps/batch size plot monitor utility updated | SparseSpec 소스 계측용으로 보임. SparseSpec 은 클라우드 트리에 있어 이 검사로 확인 불가(노드 접속 불가). 우리 엔진·하니스·dev 트리는 읽지 않음 |
| 1.7.24 | A 설정만 남음 | SSPEC_LAYER_TRACE | 0 | jh_script/claude_script/_sspec_plan.sh | 1 | SparseSpec 소스 계측용으로 보임. SparseSpec 은 클라우드 트리에 있어 이 검사로 확인 불가(노드 접속 불가). 우리 엔진·하니스·dev 트리는 읽지 않음 | |
| 1.7.25 | A 설정만 남음 | SSPEC_LAYER_TRACE_PATH | 0 | jh_script/claude_script/32_sspec_repro.sh | 1 | SparseSpec 소스 계측용으로 보임. SparseSpec 은 클라우드 트리에 있어 이 검사로 확인 불가(노드 접속 불가). 우리 엔진·하니스·dev 트리는 읽지 않음 | |
| 1.7.26 | A 설정만 남음 | SSPEC_LOGIT_TRACE | 0 | jh_script/claude_script/_sspec_budget.sh jh_script/claude_script/_sspec_dettopk.sh jh_script/claude_script/_sspec_layer.sh jh_script/claude_script/_sspec_logit.sh | 7 | SparseSpec 소스 계측용으로 보임. SparseSpec 은 클라우드 트리에 있어 이 검사로 확인 불가(노드 접속 불가). 우리 엔진·하니스·dev 트리는 읽지 않음 | |
| 1.7.27 | A 설정만 남음 | SSPEC_LOGIT_TRACE_PATH | 0 | jh_script/claude_script/32_sspec_repro.sh | 1 | SparseSpec 소스 계측용으로 보임. SparseSpec 은 클라우드 트리에 있어 이 검사로 확인 불가(노드 접속 불가). 우리 엔진·하니스·dev 트리는 읽지 않음 | |
| 1.7.28 | A 설정만 남음 | SSPEC_PLAN_TRACE | 0 | jh_script/claude_script/_sspec_bound.sh jh_script/claude_script/_sspec_boundfix.sh jh_script/claude_script/_sspec_fp.sh jh_script/claude_script/_sspec_full.sh | 10 | SparseSpec 소스 계측용으로 보임. SparseSpec 은 클라우드 트리에 있어 이 검사로 확인 불가(노드 접속 불가). 우리 엔진·하니스·dev 트리는 읽지 않음 | |
| 1.7.29 | A 설정만 남음 | SSPEC_PLAN_TRACE_PATH | 0 | jh_script/claude_script/32_sspec_repro.sh | 1 | SparseSpec 소스 계측용으로 보임. SparseSpec 은 클라우드 트리에 있어 이 검사로 확인 불가(노드 접속 불가). 우리 엔진·하니스·dev 트리는 읽지 않음 | |
| 1.7.30 | A 설정만 남음 | SSPEC_TIE_TRACE | 0 | jh_script/claude_script/_sspec_tie.sh | 1 | SparseSpec 소스 계측용으로 보임. SparseSpec 은 클라우드 트리에 있어 이 검사로 확인 불가(노드 접속 불가). 우리 엔진·하니스·dev 트리는 읽지 않음 | |
| 1.7.31 | A 설정만 남음 | SSPEC_TIE_TRACE_PATH | 0 | jh_script/claude_script/32_sspec_repro.sh | 1 | SparseSpec 소스 계측용으로 보임. SparseSpec 은 클라우드 트리에 있어 이 검사로 확인 불가(노드 접속 불가). 우리 엔진·하니스·dev 트리는 읽지 않음 | |
| 1.7.32 | A 설정만 남음 | SSPEC_TOPK_TRACE | 0 | jh_script/claude_script/_sspec_seal.sh | 1 | SparseSpec 소스 계측용으로 보임. SparseSpec 은 클라우드 트리에 있어 이 검사로 확인 불가(노드 접속 불가). 우리 엔진·하니스·dev 트리는 읽지 않음 | |
| 1.7.33 | A 설정만 남음 | SSPEC_TOPK_TRACE_PATH | 0 | jh_script/claude_script/32_sspec_repro.sh | 1 | SparseSpec 소스 계측용으로 보임. SparseSpec 은 클라우드 트리에 있어 이 검사로 확인 불가(노드 접속 불가). 우리 엔진·하니스·dev 트리는 읽지 않음 | |
| 1.7.34 | A 설정만 남음 | THRS | 0 | jh_script/claude_script/runners/run_n82_queue.sh | 1 | 지금 엔진(001·ysshin dev)·하니스·third-party·g07 010 트리 어디서도 읽지 않음. 러너에만 남은 설정 | |
| 1.7.35 | A 설정만 남음 | TP2DIR | 0 | jh_script/claude_script/fi_dir.sh | 1 | 지금 엔진(001·ysshin dev)·하니스·third-party·g07 010 트리 어디서도 읽지 않음. 러너에만 남은 설정 | |
| 1.7.36 | A 설정만 남음 | VERIFY_EOS_OVERRULED_KEEP_RELOAD | 0 | jh_script/claude_script/runners/probe_util.sh jh_script/claude_script/runners/run_ab.sh jh_script/claude_script/runners/run_aime.sh jh_script/claude_script/runners/run_aime2.sh | 56 | 지금 엔진(001·ysshin dev)·하니스·third-party·g07 010 트리 어디서도 읽지 않음. 러너에만 남은 설정 | |
| 1.7.37 | A 설정만 남음 | VLLM_HOME | 0 | .env | 1 | 001:8a9d491 2026-02-06 feat: relative path bug fixed | 지금 엔진(001·ysshin dev)·하니스·third-party·g07 010 트리 어디서도 읽지 않음. 러너에만 남은 설정 |
| 1.7.38 | A 설정만 남음 | VLLM_TEST_HOME | 0 | .env | 1 | 001:d5e50ad 2026-06-01 feat(figure): intro attention-cost + evaluation acceptance figures | 지금 엔진(001·ysshin dev)·하니스·third-party·g07 010 트리 어디서도 읽지 않음. 러너에만 남은 설정 |
| 1.7.39 | C 폐기(LEGACY) | COMPACT_SCORE_EXPORT_LEGACY | 2 | vllm_cpu/vllm/v1/attention/backends/cpu_attn.py ysshin:vllm_cpu/vllm/v1/attention/backends/cpu_attn.py | 0 | 001:2e472cc 2026-06-11 docs(cpu-verify): step op-breakdown + GQA head-reduce result | 이름에 LEGACY/DEPRECATED |
| 1.7.40 | C 폐기(LEGACY) | LOGPROB_READ_LEGACY_ANCHOR | 2 | vllm/vllm/cpu_verify/cpu_verify_engine.py ysshin:vllm/vllm/cpu_verify/cpu_verify_engine.py | 0 | vllm:e39c49308 2026-08-21 fix(cpu-verify): read logprob rows the writer actually booked | 이름에 LEGACY/DEPRECATED |
| 1.7.41 | C 폐기(옛 이름) | COMPRESSSPEC_DRAFT_ENV | 1 | compressspec/_tag.py:28 | 0 | 001:fe3223c 2026-06-10 feat: compressspec CLI + env bootstrap & packaging | 새 이름 COMPRESS_SPEC_DRAFT_ENV 의 옛 이름. 읽으면 경고 |
| 1.7.42 | C 폐기(옛 이름) | COMPRESSSPEC_NO_AUTOBUILD | 1 | compressspec/cli.py:73 | 0 | 001:71771cc 2026-06-11 docs: debug knob registry (P0) + unified-levels and probe-status update | 새 이름 COMPRESS_SPEC_NO_AUTOBUILD 의 옛 이름. 읽으면 경고 |
| 1.7.43 | C 폐기(옛 이름) | COMPRESSSPEC_NO_CLEAN | 1 | compressspec/cli.py:629 | 0 | 001:71771cc 2026-06-11 docs: debug knob registry (P0) + unified-levels and probe-status update | 새 이름 COMPRESS_SPEC_NO_CLEAN 의 옛 이름. 읽으면 경고 |
| 1.7.44 | C 폐기(옛 이름) | COMPRESSSPEC_SKIP_CHECK | 1 | compressspec/_verify.py:93 | 0 | 001:e8f85c0 2026-06-11 feat: hybrid sliding-window support | 새 이름 COMPRESS_SPEC_SKIP_CHECK 의 옛 이름. 읽으면 경고 |
| 1.7.45 | C 폐기(옛 이름) | COMPRESSSPEC_SKIP_CHECK | 1 | compressspec/_verify.py:182 | 0 | 001:e8f85c0 2026-06-11 feat: hybrid sliding-window support | 새 이름 COMPRESS_SPEC_SKIP_CHECK 의 옛 이름. 읽으면 경고 |
| 1.7.46 | C 폐기(옛 이름) | COMPRESSSPEC_TAG | 1 | compressspec/_tag.py:27 | 0 | 001:fe3223c 2026-06-10 feat: compressspec CLI + env bootstrap & packaging | 새 이름 COMPRESS_SPEC_TAG 의 옛 이름. 읽으면 경고 |
| 1.7.47 | C 폐기(옛 이름) | COMPRESSSPEC_TARGET | 1 | compressspec/_verify.py:161 | 0 | 001:fe3223c 2026-06-10 feat: compressspec CLI + env bootstrap & packaging | 새 이름 COMPRESS_SPEC_TARGET 의 옛 이름. 읽으면 경고 |
| 1.7.48 | C 폐기(옛 이름) | COMPRESSSPEC_VERIFY_ENV | 1 | compressspec/_tag.py:29 | 0 | 001:fe3223c 2026-06-10 feat: compressspec CLI + env bootstrap & packaging | 새 이름 COMPRESS_SPEC_VERIFY_ENV 의 옛 이름. 읽으면 경고 |
1.8trace·log 경량화 감사
Fable 감사 + 핵심 두 건(1·3행) 코드 직접 확인. 실행 시간은 측정하지 않음(추정). 전체 근거 file:line 은 trace_audit.md ↗
| # | 항목 | 근거 | 조치안 |
|---|---|---|---|
| 1.8.1 | 기본으로 도는 비용 | 아무 스위치도 켜지 않아도 요청·라운드마다 도는 logger.info 6곳. vllm_cpu gpu_model_runner.py:3984 등은 로그 레벨 확인 전에 f-string 을 만들고 그 안에서 env 를 읽음. 기본 GPU sync 는 없음 | 레벨 확인 뒤로 옮기거나 CS_TRACE≥1 로 묶기 |
| 1.8.2 | hot path env 읽기 | 요청·라운드마다 os.environ.get 35~45회, 같은 _diag() 함수가 6개 파일에 복사됨 (keep_map_reader·cpu_verify_engine·gpu_recompact·scheduler_recompact·index_map·runner_diff_upload) | import 시 한 번 읽어 캐시, _diag() 하나로 |
| 1.8.3 | 켜면 GPU sync | COMPACT_DIFF_STAT=1(러너 19곳) 이 재압축 step 마다 torch.cuda.synchronize() 2회(gpu_recompact.py:21-23,79). DELTA_COMPACTION_LOG≥2 는 layer 마다 .item(). SCORE_DUMP_DIR·LOG_DELTA_SELECTION=2·KV_DUMP_* 도 step 마다 sync·파일 | sync 가 필요한 계측은 CS_TRACE=2 이상으로만 |
| 1.8.4 | breakdown 쓰기 | 요청마다 한 줄 json.dumps + 줄 단위 버퍼 + 매 줄 flush(BREAKDOWN_FLUSH_EVERY=1). batch 16 추정 step 당 약 180 write (측정 안 함) | 배치 쓰기, flush 주기 늘리기 |
| 1.8.5 | 중복 | 라운드별 verify 통계 7곳, reload 시간 4곳, 재압축 이벤트 6곳, score dump 4곳, level 방식 5가지, 출력 경로 변수 14개 | 출력 루트 CS_OUT 하나 + 고정 하위 폴더 |
| 1.8.6 | 죽은 디버그 코드 | 6월 hybrid-sw 버그(PR4_*, gemma3 DUMP_*, *_DBG 14개), 6월 pr4-item8(*_DBG 8개), 9월 parity 추적(KV_DUMP_*, *_LOG 9개) 등 — 러너가 켜지 않음 | 삭제 후보(도입 커밋 기록 있음) |
통일안 (펼치기)
- 출력 루트 하나:
CS_OUT(기본 ./.realmode-run), 실행마다CS_OUT/<시각>. 하위 폴더 고정: logs/ · breakdown/{cpu_kernels,torch_profile} · trace/ · dump/{accept,score,kv,logprob,debug}. 옛 BREAKDOWN_DIR·SCORE_DUMP_DIR·KV_DUMP_DIR 등 14개는 이 아래 경로로 대응 - 스위치 최소화:
CS_TRACE=0/1/2/3(0 조용함, 1 카운터·jsonl·GPU sync 없음, 2 cycle trace·FSM·시간 측정 sync, 3 parity fingerprint),CS_DUMP=accept,score,kv,logprob,debug,BREAKDOWN=0|rows|light,BREAKDOWN_KERNEL=0|inline|profiler - 순서: ① 죽은 디버그 코드 삭제 → ② env 캐시·_diag 하나로 → ③ 기본 로그 guard → ④ CS_OUT·CS_TRACE 도입(옛 이름은 _envcompat 경고로 한동안 유지) → ⑤ GPU 에서 breakdown 전후 비교 측정
2기능 · 실험에서 설정함 (82)
실제로 동작을 바꾸고, 러너 스크립트나 .env 가 값을 준다. 논문 실험 재현에 필요한 것
| # | 이름 | 읽는 곳 | 읽는 횟수 | 기본값 | 러너가 준 값 | 설정한 곳 | 첫 위치 |
|---|---|---|---|---|---|---|---|
| 2.1 | ARM | vllm+vllm_cpu | 6 | $1 | $2 | $ARM | $TAG | jh_script | vllm/vllm/cpu_verify/channels/temp_buffer.py:575 | |
| 2.2 | ATTN_KV_RETAIN_DIV | vllm_cpu | 2 | 8 | {rd} | ysshin:scripts | vllm_cpu/csrc/cpu/cpu_attn_impl.hpp:1631 | |
| 2.3 | ATTN_TASK_BALANCE | vllm_cpu | 2 | scripts ysshin:scripts | vllm_cpu/csrc/cpu/cpu_attn_impl.hpp:484 | ||
| 2.4 | BLOCK_SIZE | compressspec+config+scripts+test+verify | 13 | "16" | "32" | $BS | $ST_BLOCK | 16 | 32 | .env scripts test ysshin:scripts | verify/profile_cpu_forward.py:33 |
| 2.5 | COMPACT_ACT_RESERVE_GIB | vllm | 1 | "1.0" | scripts ysshin:scripts | vllm/vllm/cpu_verify/compaction_upload/gpu_worker_sizing.py:19 | |
| 2.6 | COMPACT_DIFF_UPLOAD | vllm_cpu | 2 | "1" | $DU | ${DU:-1} | 0 | 0) | jh_script scripts ysshin:scripts | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:5056 |
| 2.7 | COMPACT_DISABLE | vllm_cpu | 1 | "0" | scripts ysshin:scripts | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:4743 | |
| 2.8 | COMPACT_GATE_SCORE | compressspec+vllm_cpu | 2 | "1" | $GATE | 0 | 0, | 0`` | .env jh_script scripts test | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:4751 |
| 2.9 | COMPACT_GATE_THRESHOLD | vllm_cpu | 1 | "0.8" | scripts | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:4892 | |
| 2.10 | COMPACT_INDEX_MAP_MAX | vllm_cpu | 1 | "512" | scripts ysshin:scripts | vllm_cpu/vllm/compaction/index_map.py:53 | |
| 2.11 | COMPACT_KEEPMAP_SLACK | vllm_cpu | 1 | "0" | 0 | jh_script | vllm_cpu/vllm/compaction/runner_diff_upload.py:317 |
| 2.12 | COMPACT_MAX_GPUBLK_UTIL | vllm | 1 | "0.9" | $CAP | 0 | 0, | 0.9 | jh_script scripts ysshin:scripts | vllm/vllm/cpu_verify/compaction_upload/scheduler_recompact.py:273 |
| 2.13 | COMPACT_METHOD | verify+vllm+vllm_cpu | 6 | "?" | "block" | "token" | $METHOD | block | jh_script test | vllm/vllm/v1/kv_offload/cpu.py:78 |
| 2.14 | COMPACT_NATIVE | verify+vllm_cpu | 3 | "0" | "1" | 1 | .env ysshin:.env | vllm_cpu/vllm/v1/worker/cpu_model_runner.py:68 |
| 2.15 | COMPACT_SCORE_FORCE_OFF | vllm_cpu | 1 | "0" | 1 | jh_script | vllm_cpu/vllm/v1/attention/backends/cpu_attn.py:51 |
| 2.16 | COMPACT_THRESHOLD | compressspec+vllm_cpu | 3 | "0.5" | "0.95" | $THR | 0.5 | 0.5, | 0.6 | .env jh_script scripts test | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:5226 |
| 2.17 | COMPACT_TOKEN_NUM_RECENT | vllm_cpu | 1 | "32" | 32 | test | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:5244 |
| 2.18 | COMPACT_TOKEN_NUM_SINK | vllm_cpu | 1 | "32" | 32 | test | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:5243 |
| 2.19 | COMPACT_UPLOAD_PEND | verify+vllm_cpu | 5 | "1" | ${LC_UPLOAD_PEND:-1} | scripts ysshin:scripts | vllm_cpu/vllm/compaction/runner_diff_upload.py:76 |
| 2.20 | COMPRESS_SPEC_CYCLE_ANNOTATE | test+vllm | 3 | "0" | 1): | test | vllm/vllm/cpu_verify/cpu_verify_engine.py:3705 |
| 2.21 | COMPRESS_SPEC_DRAFT_ENV | compressspec | 1 | "COMPRESSSPEC_DRAFT_ENV" | vllm | jh_script | compressspec/_tag.py:28 |
| 2.22 | COMPRESS_SPEC_VERIFY_ENV | compressspec | 1 | "COMPRESSSPEC_VERIFY_ENV" | verify | jh_script | compressspec/_tag.py:29 |
| 2.23 | COMPRESSION_SPEC_DECODING_INTERVAL | config+test+vllm | 13 | "0" | "1024" | "2048" | "1024" | "2048" | "64" | <int> | test | vllm/vllm/v1/attention/backends/flashinfer.py:600 |
| 2.24 | COMPRESSION_SPEC_MAX_TOTAL_CONTEXT | vllm | 2 | scripts ysshin:scripts | vllm/vllm/v1/core/sched/scheduler.py:323 | ||
| 2.25 | COMPRESSION_SPEC_SELECTED_TOKENS | config+scripts+test+vllm | 18 | "128" | "2048" | "512" | "64" | "8192" | <int> | os.environ.get( | scripts test ysshin:scripts | vllm/vllm/config/compression_spec_config.py:81 |
| 2.26 | COMPSPEC_DECODE_ATTN_INTERVAL | scripts+test+vllm | 4 | os.environ.get( | scripts test ysshin:scripts | vllm/vllm/v1/attention/backends/flashinfer.py:599 | |
| 2.27 | COMPSPEC_PREEMPT_OFFLOAD | vllm | 3 | "0" | scripts ysshin:scripts | vllm/vllm/cpu_verify/compaction_upload/manager.py:502 | |
| 2.28 | COMPSPEC_RESUME_NEED_GATE | vllm | 1 | "0" | scripts ysshin:scripts | vllm/vllm/cpu_verify/cpu_verify_engine.py:399 | |
| 2.29 | CPU_OFFLOAD_POOL_BLOCKS | vllm+vllm_cpu | 4 | "" | $_kpool | 35000) | K | scripts test ysshin:scripts | vllm/vllm/cpu_verify/cpu_verify_engine.py:2851 |
| 2.30 | CPU_VERIFY_URL | test+vllm | 4 | "http://localhost:8000" | http://localhost: | http://localhost:$ST_PORT | http://localhost:18000 | jh_script scripts ysshin:scripts | vllm/vllm/cpu_verify/_compaction_upload_dispatcher.py:96 |
| 2.31 | DRAFT_DISPATCH_FULL | vllm | 1 | "0" | ${DRAFT_DISPATCH_FULL:-0} | jh_script | vllm/vllm/cpu_verify/channels/temp_buffer.py:700 |
| 2.32 | DYN_FORCE_OUT | vllm | 1 | "0" | $LC_DYN_FORCE_OUT | scripts ysshin:scripts | vllm/vllm/cpu_verify/dyn_router.py:61 |
| 2.33 | DYN_ROUTER | scripts+test+vllm | 3 | "0" | 1 | 1) | scripts ysshin:scripts | vllm/vllm/cpu_verify/dyn_router.py:51 |
| 2.34 | DYN_ROUTER_HEADROOM_PCT | test+vllm | 2 | "0.10" | $LC_DYN_HEADROOM_PCT | scripts ysshin:scripts | vllm/vllm/cpu_verify/dyn_router.py:60 |
| 2.35 | EXP_DTYPE | compressspec+config+scripts+test+verify | 14 | "bf16" | "fp16" | "bf16" | $DT | $ST_DTYPE | bf16 | .env jh_script scripts test | verify/profile_cpu_forward.py:34 |
| 2.36 | EXP_SELECTED_TOKENS | compressspec+scripts+test+vllm_cpu | 8 | "128" | "2048" | 128`` | scripts test ysshin:scripts | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:4080 |
| 2.37 | LAYER_SERIAL_CHUNKED | vllm | 4 | "0" | False | 1 | 1) | jh_script scripts ysshin:scripts | vllm/vllm/config/scheduler.py:307 |
| 2.38 | LOGPROB_POOL_SIZE | config+dump_debugger+verify | 6 | "480" | "8320" | "32" | 4160 | 4160) | 8320 | .env jh_script | verify/test_logprob_stamp.py:31 |
| 2.39 | LOGPROB_VOCAB_SIZE | verify | 4 | 151936 | jh_script scripts ysshin:scripts | verify/test_logprob_stamp.py:30 | |
| 2.40 | LS_CHUNK_TILE | vllm | 1 | "8192" | 2048 | 8192 | jh_script scripts ysshin:scripts | vllm/vllm/v1/worker/gpu_model_runner.py:6486 |
| 2.41 | LS_PIPELINE_RELOAD | vllm | 1 | "1" | 1 | scripts ysshin:scripts | vllm/vllm/v1/attention/backends/flashinfer.py:2041 |
| 2.42 | LS_RELOAD_MEMCPY | vllm | 1 | "1" | 1 | scripts ysshin:scripts | vllm/vllm/v1/attention/backends/flashinfer.py:2093 |
| 2.43 | LS_STREAM_STAGE | vllm | 1 | m | 1 | scripts ysshin:scripts | vllm/vllm/v1/attention/backends/flashinfer.py:91 |
| 2.44 | MAX_NUM_SEQS | test+vllm_cpu | 4 | "0" | "2" | ysshin:scripts | vllm_cpu/vllm/compaction/keep_map_shm.py:88 | |
| 2.45 | MODEL_DIR | breakdown_debugger+compressspec+config+dump_debugger+scripts+test | 28 | "Qwen/Qwen3-8B" | "meta-llama/Llama-3.1-8B-Instruct" | default_model | $R/models/Qwen/Qwen3-4B-Thinki | ${VLLM_TEST_HOME}/models/AmanP | ${VLLM_TEST_HOME}/models/googl | ${VLLM_TEST_HOME}/models/meta- | .env jh_script scripts test | compressspec/cli.py:274 |
| 2.46 | MULTIBP_MAX_MODEL_LEN | config | 1 | "25300" | "2048" | $MML | 4096 | 45056 | jh_script | config/connector_config.py:53 |
| 2.47 | MULTIBP_MAX_NUM_SEQS | config+dump_debugger | 2 | "10" | "2" | 100 | 16 | 2 | .env jh_script | config/connector_config.py:52 |
| 2.48 | MULTIBP_QUESTION | config | 1 | 5 | jh_script | config/common_util.py:27 | |
| 2.49 | MULTIBP_REPEAT | config | 1 | "1" | $REP | 1 | 16 | jh_script | config/common_util.py:29 |
| 2.50 | NUM_GPU_BLOCKS | dump_debugger+test+vllm_cpu | 9 | "0" | $LC_NUM_GPU_BLOCKS | ${LC_NUM_GPU_BLOCKS:-5500} | 5500 | <n> | scripts test ysshin:scripts | vllm_cpu/vllm/compaction/runner_diff_upload.py:172 |
| 2.51 | NUMA_AWARE_VERIFY_NODES | verify+vllm+vllm_cpu | 7 | "4" | $nn | scripts ysshin:scripts | vllm/vllm/cpu_verify/channels/naive_copy.py:419 |
| 2.52 | NUMA_FIRST_TOUCH | vllm+vllm_cpu | 3 | "0" | scripts ysshin:scripts | vllm/vllm/v1/kv_offload/worker/cpu_gpu.py:1074 | |
| 2.53 | NUMA_FT_NODES | compressspec+vllm_cpu | 2 | "" | scripts ysshin:scripts | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:7510 | |
| 2.54 | OUT_ROOT | vllm_cpu | 1 | '/tmp' | $OUT | $R/jh_script/claude_script/res | $REPO/dump_debugger/dump/sweep | jh_script | vllm_cpu/vllm/compaction/index_map.py:92 |
| 2.55 | PORT | compressspec+verify | 3 | "8000" | $ST_PORT | 18000 | 8123 | 8155 | jh_script scripts ysshin:scripts | verify/cpu_verify_server_score.py:602 |
| 2.56 | RS_TRUNCATE | verify | 1 | "1" | scripts ysshin:scripts | verify/utils.py:583 | |
| 2.57 | RULER_BATCH | config+dump_debugger+test | 12 | "10" | "16" | $B | 1 | 16 | 16, | jh_script scripts test ysshin:scripts | config/common_util.py:38 |
| 2.58 | RULER_CTX | config+test | 5 | "16384" | "4096" | $CTX | 16384 | 2048 | 4096 | jh_script scripts test ysshin:scripts | config/common_util.py:49 |
| 2.59 | RULER_OFFSET | config+dump_debugger+test | 6 | "0" | '0' | str(_do | $((DOC-1)) | str(_it * _bsz) | jh_script test | config/common_util.py:68 |
| 2.60 | RULER_QDIR | config+dump_debugger | 4 | _dq | _pd | os.path.join(REPO, "test", "RulerQuestions" | $BAB | $CLOUD_ROOT/prompts/lb_v2_qwen | $LB | $NIAH | jh_script | config/common_util.py:67 |
| 2.61 | RULER_SEQ | config+verify | 4 | <token_len> | test | verify/profile_cpu_forward.py:242 | |
| 2.62 | RULER_SUFFIX | config | 1 | _ds | _babilong | _longbench | _niah | jh_script | config/common_util.py:82 |
| 2.63 | RULER_TASK | config+dump_debugger+verify | 4 | "cwe,vt" | "niah_single_1" | all | cwe | cwe,vt | .env jh_script | verify/profile_cpu_forward.py:249 |
| 2.64 | SCORE_BLOCK_SIZE | config+test+verify+vllm_cpu | 5 | "128" | 128 | .env | vllm_cpu/vllm/v1/worker/score_shm_reader.py:28 |
| 2.65 | SCORE_NUM_LAYERS | verify+vllm_cpu | 5 | "32" | str(_hf.num_hidden_layers | 32 | .env scripts ysshin:scripts | vllm_cpu/vllm/v1/worker/score_shm_reader.py:25 |
| 2.66 | SCORE_NUM_Q_HEADS | verify+vllm_cpu | 5 | "32" | str(_hf.num_attention_heads | 32 | .env scripts ysshin:scripts | vllm_cpu/vllm/v1/worker/score_shm_reader.py:27 |
| 2.67 | SCORE_POOL_SIZE | config+dump_debugger+test+verify+vllm_cpu | 6 | "100" | "8" | 4160 | 4160, | 8320 | .env jh_script | vllm_cpu/vllm/v1/worker/score_shm_reader.py:26 |
| 2.68 | SEED | config | 1 | "0" | 0 | jh_script | config/connector_config.py:51 |
| 2.69 | SHARED_CPU_POOL_GB | compressspec+test | 5 | "2" | "8" | "2" | 140 | 20 | 200 | .env jh_script | compressspec/cli.py:219 |
| 2.70 | SMALL_MODEL_DIR | config+test | 4 | $M | $MODEL | M | jh_script | config/connector_config.py:31 | |
| 2.71 | SPECULATIVE_RUN_STEP | config+vllm_cpu | 2 | "16" | "4" | 16 | .env jh_script | vllm_cpu/vllm/compaction/index_map.py:59 |
| 2.72 | TASK_NAME | config | 1 | RULER | babilong | longbench_v2 | .env jh_script | config/common_util.py:53 | |
| 2.73 | VERIFY_DP_SIZE | verify+vllm+vllm_cpu | 7 | "1" | scripts ysshin:scripts | vllm/vllm/cpu_verify/channels/naive_copy.py:417 | |
| 2.74 | VERIFY_ENFORCE_EAGER | verify | 2 | "1" | 1 | jh_script | verify/cpu_verify_server_score.py:111 |
| 2.75 | VERIFY_KV_SHM_RANK | vllm_cpu | 1 | self.parallel_config.rank | 0 | 0) | scripts ysshin:scripts | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:7388 |
| 2.76 | VERIFY_MAX_MODEL_LEN | compressspec+verify | 3 | "50000" | $MML | 1044480 | 20000 | 30000 | jh_script scripts ysshin:scripts | verify/cpu_verify_server_score.py:139 |
| 2.77 | VERIFY_MAX_NUM_SEQS | dump_debugger+verify | 4 | "1" | "32" | "4" | ${VSEQS:-$B} | 100 | 64 | .env jh_script test | verify/profile_cpu_forward.py:278 |
| 2.78 | VERIFY_MODEL | breakdown_debugger+compressspec+dump_debugger+verify | 7 | "meta-llama/Llama-3.1-8B-Instruct" | $M | $MODEL | $R/models/Qwen/Qwen3-4B-Thinki | $R/models/meta-llama/Llama-3.1 | .env jh_script scripts ysshin:scripts | verify/profile_cpu_forward.py:32 |
| 2.79 | VERIFY_NUMA_INTERLEAVE | verify | 2 | "1" | 0 | 0: | scripts ysshin:scripts | verify/cpu_verify_server.py:35 |
| 2.80 | VERIFY_TOKENIZER | dump_debugger+verify | 4 | MODEL_NAME | model | $R/models/meta-llama/Llama-3.1 | jh_script scripts ysshin:scripts | verify/cpu_verify_server_score.py:46 |
| 2.81 | VLLM_ATTENTION_BACKEND | config+scripts+test | 16 | "FLASHINFER" | "FLASHINFER" | 'FLASHINFER' | BACKEND | FLASHINFER | jh_script scripts test ysshin:scripts | config/connector_config.py:71 |
| 2.82 | VLLM_SHM_NO_INTERLEAVE | config | 1 | 1 | 1. | jh_script scripts ysshin:scripts | config/common_util.py:188 |
3기능 · 기본값으로만 씀 (116)
동작을 바꾸지만 어떤 러너도 값을 주지 않는다. 기본값이 곧 운영값
| # | 이름 | 읽는 곳 | 읽는 횟수 | 기본값 | 러너가 준 값 | 설정한 곳 | 첫 위치 |
|---|---|---|---|---|---|---|---|
| 3.1 | ACCEPT_MODE | judge_debugger+verify | 2 | "standard" | verify/utils.py:1306 | ||
| 3.2 | ATTN_BP_STREAM_CAP | vllm | 1 | "64" | vllm/vllm/distributed/kv_transfer/kv_connector/v1/attn_byproduct_handler/handler.py:291 | ||
| 3.3 | CACTUS_DELTA | judge_debugger+verify | 2 | "1.0" | verify/utils.py:1309 | ||
| 3.4 | COMPACT_BLOCK_KEEP_RATIO | vllm_cpu | 1 | "0.1" | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:4075 | ||
| 3.5 | COMPACT_BLOCK_POLICY | vllm_cpu | 1 | "mass" | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:4024 | ||
| 3.6 | COMPACT_BLOCK_SIZE | vllm | 1 | "32" | vllm/vllm/cpu_verify/keep_map_reader.py:32 | ||
| 3.7 | COMPACT_FINISHED_RESEND | vllm | 1 | "4" | vllm/vllm/cpu_verify/cpu_verify_engine.py:4342 | ||
| 3.8 | COMPACT_FULL_COPY_GB | vllm | 1 | "8" | vllm/vllm/config/compression_spec_config.py:64 | ||
| 3.9 | COMPACT_GATE_AGG | vllm_cpu | 1 | "min" | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:4884 | ||
| 3.10 | COMPACT_HOLD_FULL | vllm_cpu | 1 | "" | vllm_cpu/vllm/compaction/index_map.py:96 | ||
| 3.11 | COMPACT_KEEPMAP_SLOT_MULT | vllm_cpu | 1 | "2" | vllm_cpu/vllm/compaction/keep_map_shm.py:106 | ||
| 3.12 | COMPACT_SCORE_EXPORT_LEGACY | vllm_cpu | 1 | "0" | vllm_cpu/vllm/v1/attention/backends/cpu_attn.py:47 | ||
| 3.13 | COMPACT_SCORE_FP16 | vllm_cpu | 1 | "1" | vllm_cpu/vllm/v1/attention/backends/cpu_attn.py:65 | ||
| 3.14 | COMPACT_SCORE_NO_REDUCE | vllm_cpu | 1 | "0" | vllm_cpu/vllm/v1/attention/backends/cpu_attn.py:59 | ||
| 3.15 | COMPACT_TEMP_BUFFER_GB | vllm | 1 | "8" | vllm/vllm/config/compression_spec_config.py:62 | ||
| 3.16 | COMPACT_TOKEN_GLOBAL_RANK | vllm_cpu | 2 | "0" | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:3737 | ||
| 3.17 | COMPACT_TOKEN_HEAD_AGG | verify+vllm_cpu | 6 | "gqa_max" | 'gqa_max' | vllm_cpu/vllm/v1/attention/backends/cpu_attn.py:60 | ||
| 3.18 | COMPACT_TOKEN_POOL_K | vllm_cpu | 2 | "7" | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:3726 | ||
| 3.19 | COMPACT_TOKEN_POOL_TYPE | vllm_cpu | 2 | "max" | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:3727 | ||
| 3.20 | COMPACT_TOKEN_SCORE_ROWS | verify+vllm_cpu | 4 | "draft_first_last" | vllm_cpu/vllm/v1/attention/backends/cpu_attn.py:69 | ||
| 3.21 | COMPACTION_IDX_EVERY | vllm_cpu | 1 | "1" | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:4204 | ||
| 3.22 | COMPACTION_IDX_LAYERS | vllm_cpu | 1 | "all" | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:4207 | ||
| 3.23 | COMPRESS_SPEC_CPU_POOL_HOLD | vllm | 2 | "1" | vllm/vllm/cpu_verify/cpu_verify_engine.py:1182 | ||
| 3.24 | COMPRESS_SPEC_NO_AUTOBUILD | compressspec | 1 | "COMPRESSSPEC_NO_AUTOBUILD" | compressspec/cli.py:73 | ||
| 3.25 | COMPRESS_SPEC_NO_CLEAN | compressspec | 1 | "COMPRESSSPEC_NO_CLEAN" | compressspec/cli.py:629 | ||
| 3.26 | COMPRESS_SPEC_NO_FREE_SW | dump_debugger+vllm_cpu | 4 | "0" | vllm_cpu/vllm/v1/core/single_type_kv_cache_manager.py:608 | ||
| 3.27 | COMPRESS_SPEC_PIN_FULLKV | dump_debugger+vllm_cpu | 4 | "0" | vllm_cpu/vllm/v1/core/kv_cache_manager.py:344 | ||
| 3.28 | COMPRESS_SPEC_SYNC_VERIFY | vllm | 1 | "0" | vllm/vllm/v1/core/sched/scheduler.py:1730 | ||
| 3.29 | COMPRESS_SPEC_TARGET | compressspec | 1 | "COMPRESSSPEC_TARGET" | compressspec/_verify.py:161 | ||
| 3.30 | COMPSPEC_BONUS_TOKEN | verify | 1 | "0" | verify/cpu_verify_server.py:290 | ||
| 3.31 | COMPSPEC_DRAFT_STALL_S | vllm | 1 | "60" | vllm/vllm/cpu_verify/cpu_verify_engine.py:514 | ||
| 3.32 | COMPSPEC_OFFLOAD_COVER | vllm | 1 | "0" | vllm/vllm/cpu_verify/cpu_verify_engine.py:2773 | ||
| 3.33 | COMPSPEC_RELOAD_STALL_S | vllm | 1 | "120" | vllm/vllm/cpu_verify/cpu_verify_engine.py:519 | ||
| 3.34 | COMPSPEC_RESUME_NEED_MARGIN | vllm | 1 | "0" | vllm/vllm/cpu_verify/cpu_verify_engine.py:401 | ||
| 3.35 | COMPSPEC_RESUME_NEED_TIMEOUT_S | vllm | 1 | "600" | vllm/vllm/cpu_verify/cpu_verify_engine.py:403 | ||
| 3.36 | COMPSPEC_RUNAHEAD | vllm | 1 | "" | vllm/vllm/cpu_verify/cpu_verify_engine.py:424 | ||
| 3.37 | COMPSPEC_SKIP_WATCH_SECS | vllm | 1 | "120" | vllm/vllm/cpu_verify/cpu_verify_engine.py:546 | ||
| 3.38 | CP_BATCH_HEAD_REDUCE | verify+vllm_cpu | 3 | "1" | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:3865 | ||
| 3.39 | CPU_FWD_READ_HASH | vllm_cpu | 1 | "0" | vllm_cpu/vllm/v1/attention/backends/cpu_attn.py:442 | ||
| 3.40 | CPU_VERIFY_TIMEOUT_S | vllm | 2 | "500" | vllm/vllm/cpu_verify/_compaction_upload_dispatcher.py:99 | ||
| 3.41 | CS_DRAFT_HOLD | vllm | 1 | "1" | vllm/vllm/cpu_verify/cpu_verify_engine.py:2261 | ||
| 3.42 | CS_DRAFT_HOLD_MAX | vllm | 1 | "64" | vllm/vllm/cpu_verify/cpu_verify_engine.py:2264 | ||
| 3.43 | CS_LANDING_PRUNE | vllm | 2 | "1" | vllm/vllm/cpu_verify/cpu_verify_engine.py:2082 | ||
| 3.44 | DENSE_NO_COMPACT | vllm_cpu | 1 | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:4756 | |||
| 3.45 | DIFF_NATIVE | vllm_cpu | 2 | "1" | vllm_cpu/vllm/compaction/diff_plan.py:97 | ||
| 3.46 | DIFF_PAD_SHORT_FATAL | vllm_cpu | 1 | "0" | vllm_cpu/vllm/compaction/diff_plan.py:148 | ||
| 3.47 | DIFF_PARTIAL_RELOAD | vllm | 1 | "1" | vllm/vllm/cpu_verify/channels/temp_buffer.py:504 | ||
| 3.48 | DYN_ROUTER_HOLD_STEPS | vllm | 1 | "500" | vllm/vllm/cpu_verify/dyn_router.py:69 | ||
| 3.49 | FORCE_ATTN_ISA | vllm+vllm_cpu | 2 | "" | vllm/vllm/v1/kv_offload/worker/cpu_gpu.py:248 | ||
| 3.50 | FSM_RUN_DIR | breakdown_debugger+vllm | 2 | "" | vllm/vllm/cpu_verify/_fsm_log.py:16 | ||
| 3.51 | GPU_CORES | config | 1 | "8" | config/common_util.py:313 | ||
| 3.52 | HEADINFER_DECODE | vllm | 1 | False | vllm/vllm/v1/attention/backends/headinfer_decode.py:39 | ||
| 3.53 | HEADINFER_DECODE_REAL | vllm | 1 | False | vllm/vllm/v1/attention/backends/headinfer_decode.py:41 | ||
| 3.54 | HI_NTB | vllm | 1 | "32" | vllm/vllm/v1/attention/backends/headinfer_decode.py:148 | ||
| 3.55 | INIT_FILE_PATH | vllm | 2 | vllm/vllm/cpu_verify/utils.py:664 | |||
| 3.56 | KEEPMAP_PAD_LAYERS | vllm_cpu | 1 | "1" | vllm_cpu/vllm/compaction/keep_map_shm.py:47 | ||
| 3.57 | KEEPMAP_STRICT_SLOTS | vllm | 1 | "1" | vllm/vllm/cpu_verify/keep_map_reader.py:30 | ||
| 3.58 | LAYER_SERIAL_BARRIER | vllm | 1 | "stream" | vllm/vllm/cpu_verify/cpu_verify_engine.py:4281 | ||
| 3.59 | LAYER_SERIAL_COVERAGE_BLOCKS | vllm | 1 | vllm/vllm/v1/core/kv_cache_utils.py:871 | |||
| 3.60 | LAYER_SERIAL_K | vllm | 9 | "0" | vllm/vllm/cpu_verify/channels/naive_copy.py:371 | ||
| 3.61 | LAYER_SERIAL_TAIL_BLOCKS | vllm | 3 | "16" | vllm/vllm/cpu_verify/cpu_verify_engine.py:1365 | ||
| 3.62 | LOGPROB_READ_LEGACY_ANCHOR | vllm | 1 | "0" | vllm/vllm/cpu_verify/cpu_verify_engine.py:3478 | ||
| 3.63 | LOGPROB_READ_POS_OFFSET | vllm | 1 | "0" | vllm/vllm/cpu_verify/cpu_verify_engine.py:3481 | ||
| 3.64 | LOGPROB_SHM_NAME | verify | 3 | f"vllm_logprobs_shm{shm_suffix( | verify/test_logprob_stamp.py:29 | ||
| 3.65 | LOGPROB_SPARSE | verify+vllm | 5 | "0" | vllm/vllm/cpu_verify/logprob_handler.py:25 | ||
| 3.66 | LOGPROB_SPARSE_CAP | verify+vllm | 3 | "128" | vllm/vllm/cpu_verify/logprob_handler.py:27 | ||
| 3.67 | LS_CHUNKED_VERIFY | vllm | 1 | "0" | vllm/vllm/v1/attention/backends/flashinfer.py:1878 | ||
| 3.68 | LS_STREAM_FREE | vllm | 1 | m | vllm/vllm/v1/attention/backends/flashinfer.py:93 | ||
| 3.69 | LS_STREAM_RELOAD | vllm | 1 | m | vllm/vllm/v1/attention/backends/flashinfer.py:92 | ||
| 3.70 | LSE_SHM_NAME | vllm_cpu | 1 | f"vllm_rank_0_attn_lse{shm_suffix( | vllm_cpu/vllm/v1/worker/score_shm_reader.py:24 | ||
| 3.71 | MAX_MODEL_LEN | test+verify | 7 | "50000" | "8192" | verify/profile_cpu_forward.py:36 | ||
| 3.72 | MAXPOS_OVERRIDE | test+verify | 7 | None | verify/cpu_verify_server_score.py:142 | ||
| 3.73 | MINIFORGE_DIR | compressspec | 1 | str(Path.home( | compressspec/cli.py:49 | ||
| 3.74 | MULTIBP_COMPRESS | config | 1 | "streamingllm" | config/connector_config.py:55 | ||
| 3.75 | OFFLOAD_PREFIX_CACHE | vllm | 1 | "0" | vllm/vllm/cpu_verify/channels/naive_copy.py:408 | ||
| 3.76 | PR4_DISABLE_COMPRESS_CALL | vllm | 1 | "0" | vllm/vllm/v1/attention/backends/flashinfer.py:2874 | ||
| 3.77 | READY_TIMEOUT | compressspec | 1 | "600" | compressspec/cli.py:446 | ||
| 3.78 | RECOMPACT_SLOT_ALIGN | vllm_cpu | 1 | "1" | vllm_cpu/vllm/compaction/index_map.py:58 | ||
| 3.79 | REJECT_DISTRIB_PT | verify | 1 | "0" | verify/utils.py:1301 | ||
| 3.80 | REJECT_DISTRIB_PT_MAX | judge_debugger+verify | 2 | "2000" | verify/utils.py:1302 | ||
| 3.81 | REJECT_DISTRIB_SOURCE | verify | 1 | "server" | verify/utils.py:1298 | ||
| 3.82 | REJECT_DISTRIB_TOPK | verify | 1 | "8" | verify/utils.py:1296 | ||
| 3.83 | RELAX_RESIDUAL | judge_debugger+verify | 2 | "v1" | verify/utils.py:1308 | ||
| 3.84 | RELAX_TAU | judge_debugger+verify | 2 | "0.5" | verify/utils.py:1307 | ||
| 3.85 | RESERVE_SKIPPED_SW_BLOCKS | vllm_cpu | 1 | "0" | vllm_cpu/vllm/v1/core/single_type_kv_cache_manager.py:198 | ||
| 3.86 | REUSE_ENGINE_RID | verify | 1 | "0" | verify/cpu_verify_server.py:505 | ||
| 3.87 | RS_ALLOW_UNSUPPORTED_SAMPLING | vllm | 1 | "0" | vllm/vllm/cpu_verify/channels/temp_buffer.py:83 | ||
| 3.88 | RS_FAST_TRUNC | verify | 1 | "1" | verify/utils.py:620 | ||
| 3.89 | RS_FAST_TRUNC_MAX_K | verify | 1 | "4096" | verify/utils.py:623 | ||
| 3.90 | RS_KTH_CHUNKS | verify | 1 | "32" | verify/utils.py:702 | ||
| 3.91 | RS_KTH_MIN_VOCAB | verify | 1 | "16384" | verify/utils.py:703 | ||
| 3.92 | RS_Q_FLOOR | verify | 1 | "1e-6" | verify/utils.py:874 | ||
| 3.93 | RS_ROW_MISMATCH | verify | 2 | "fatal" | "reject" | verify/utils.py:875 | ||
| 3.94 | RS_STALE_MAX_RATE | verify | 1 | "0.02" | verify/utils.py:608 | ||
| 3.95 | RS_STALE_MIN_POS | verify | 1 | "200" | verify/utils.py:609 | ||
| 3.96 | RS_TIE_EPS | verify | 1 | "1e-3" | verify/utils.py:599 | ||
| 3.97 | RS_TOPP_EPS | verify | 1 | "1e-5" | verify/utils.py:589 | ||
| 3.98 | RULER_DYNAMIC | config | 1 | config/common_util.py:54 | |||
| 3.99 | SCORE_SHM_NAME | vllm_cpu | 1 | f"vllm_rank_0_attn_score{shm_suffix( | vllm_cpu/vllm/v1/worker/score_shm_reader.py:23 | ||
| 3.100 | SKIP_DUMMY_BATCH | vllm_cpu | 1 | "0" | vllm_cpu/vllm/v1/engine/core.py:1381 | ||
| 3.101 | SNAPKV_FUSE_RELOAD | vllm | 1 | m | vllm/vllm/v1/attention/backends/flashinfer.py:96 | ||
| 3.102 | SNAPKV_KEY_CHUNK | vllm | 1 | vllm/vllm/v1/attention/backends/flashinfer.py:94 | |||
| 3.103 | SNAPKV_WIN_CHUNK | vllm | 1 | "8" | vllm/vllm/v1/attention/backends/flashinfer.py:2477 | ||
| 3.104 | STREAMINGLLM_SINK_TOKENS | config | 1 | "64" | config/connector_config.py:56 | ||
| 3.105 | TAS_DELTA | verify | 1 | "0.3" | verify/utils.py:1311 | ||
| 3.106 | TAS_EPS | verify | 1 | "0.09" | verify/utils.py:1310 | ||
| 3.107 | UPLOAD_SKIP_FATAL | vllm | 2 | "0" | vllm/vllm/cpu_verify/compaction_upload/manager.py:175 | ||
| 3.108 | VERIFY_BATCH_WINDOW_MS | verify | 1 | "0" | verify/cpu_verify_server.py:139 | ||
| 3.109 | VERIFY_DISABLE_HYBRID_KVCM | verify | 1 | "0" | verify/cpu_verify_server.py:232 | ||
| 3.110 | VERIFY_DISPATCH_WORKERS | vllm | 1 | "" | vllm/vllm/cpu_verify/cpu_verify_engine.py:459 | ||
| 3.111 | VERIFY_KV_PREFAULT | vllm_cpu | 1 | "0" | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:2984 | ||
| 3.112 | VERIFY_KV_PREFAULT_ARM_IDLE | vllm_cpu | 1 | "2.0" | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:3001 | ||
| 3.113 | VERIFY_NUMA_LOCAL | verify | 2 | verify/cpu_verify_server.py:42 | |||
| 3.114 | VERIFY_NUM_SPEC_BLOCKS | verify+vllm_cpu | 3 | "0" | vllm_cpu/vllm/attention/layer.py:458 | ||
| 3.115 | VERIFY_PROMPT_LOGPROBS | verify | 1 | "-1" | verify/cpu_verify_server.py:118 | ||
| 3.116 | VLLM_SHM_SUFFIX | vllm+vllm_cpu | 2 | vllm/vllm/shm_ns.py:13 |
4로그·디버그 · 실험에서 설정함 (27)
출력·계측용인데 러너가 켠다(breakdown 등)
| # | 이름 | 읽는 곳 | 읽는 횟수 | 기본값 | 러너가 준 값 | 설정한 곳 | 첫 위치 |
|---|---|---|---|---|---|---|---|
| 4.1 | ATTN_BP_DEBUG_FILL | config+test+vllm | 3 | "" | "1" | 0 | .env | vllm/vllm/distributed/kv_transfer/kv_connector/v1/offloading_attn_bp_connector.py:336 |
| 4.2 | ATTN_BP_DUMP_SCORE_CHECK | vllm | 1 | "" | 0 | .env | vllm/vllm/v1/worker/gpu_model_runner.py:6836 |
| 4.3 | ATTN_KV_DUMP_DIR | test+vllm | 4 | "" | ${DUMP_DIR}/kvcache | .env | vllm/vllm/cpu_verify/debug_utils.py:98 |
| 4.4 | BENCH_BATCH | config | 3 | 1 | .env jh_script | config/common_util.py:38 | |
| 4.5 | BENCH_CTX | config+test | 4 | 262144 | 32768 | _saved_ctx | _wctx | .env jh_script test | config/common_util.py:49 | |
| 4.6 | BREAKDOWN_DIR | breakdown_debugger+test+verify+vllm+vllm_cpu | 13 | "" | $BD | $R/breakdown_debugger/dump/bre | $ROOT/breakdown_debugger/dump/ | ${BREAKDOWN_DUMP_DIR}/breakdow | .env jh_script | vllm/vllm/breakdown_logger.py:149 |
| 4.7 | BREAKDOWN_KERNEL_INLINE | breakdown_debugger+vllm_cpu | 3 | "0" | "1" | 0 | 1 | jh_script | vllm_cpu/vllm/model_executor/models/_cpu_kernel_timer.py:18 |
| 4.8 | BREAKDOWN_KERNEL_PROFILE_MODE | breakdown_debugger+config+test+verify+vllm_cpu | 6 | "0" | "1" | 0 | 1 | .env jh_script test | vllm_cpu/vllm/model_executor/models/_cpu_kernel_timer.py:17 |
| 4.9 | COMPACT_ACK_LOG | vllm | 4 | "0" | ${COMPACT_ACK_LOG:-0} | jh_script | vllm/vllm/cpu_verify/cpu_verify_engine.py:2068 |
| 4.10 | COMPACT_DIFF_STAT | verify+vllm_cpu | 3 | "0" | 0 | 1 | jh_script | vllm_cpu/vllm/v1/core/sched/scheduler.py:1321 |
| 4.11 | COMPRESS_SPEC_SKIP_CHECK | compressspec | 2 | "COMPRESSSPEC_SKIP_CHECK" | 1 | jh_script | compressspec/_verify.py:93 |
| 4.12 | COMPRESSION_SPEC_DEBUG | test+vllm | 14 | "true" | true | test | vllm/vllm/v1/core/single_type_kv_cache_manager.py:81 | |
| 4.13 | CPU_KV_HASH_DBG | vllm_cpu | 1 | "0" | 1 | jh_script | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:5106 |
| 4.14 | DEBUG_METADATA_DUMP_DIR | dump_debugger+verify+vllm_cpu | 3 | "" | $OUT/accept_meta | $R512/accept_meta | $ROOT/results/gptoss20b_e2e | $ROOT/results/gptoss_e2e | .env jh_script | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:4195 |
| 4.15 | DELTA_COMPACTION_LOG | vllm+vllm_cpu | 10 | "0" | 1. | jh_script scripts ysshin:scripts | vllm/vllm/cpu_verify/compaction_trace.py:16 |
| 4.16 | DRAFT_DISPATCH_LOG | vllm | 1 | "0" | ${DRAFT_DISPATCH_LOG:-0} | jh_script | vllm/vllm/cpu_verify/channels/temp_buffer.py:701 |
| 4.17 | DUMP_EXTEND_BEFORE_VERIFY | vllm | 1 | "0" | 1 | jh_script | vllm/vllm/v1/core/sched/scheduler.py:2303 |
| 4.18 | EXAMPLE_DIAG_QID | compressspec+test | 2 | "1" | "DIAG_QID", "5" | 1 | test | compressspec/cli.py:254 |
| 4.19 | GPU_LOG | compressspec | 1 | str(log_dir / "gpu_client.log" | test | compressspec/cli.py:450 | |
| 4.20 | HELD_TRACE | vllm+vllm_cpu | 2 | "0" | 0 | jh_script | vllm/vllm/cpu_verify/cpu_verify_engine.py:1857 |
| 4.21 | LAST_VERIFY_LOG | vllm | 2 | "0" | ${LAST_VERIFY_LOG:-0} | jh_script | vllm/vllm/cpu_verify/_tracker_registry.py:141 |
| 4.22 | LOG_DELTA_SELECTION | vllm+vllm_cpu | 9 | "0" | ${LOG_DELTA_SELECTION:-1} | jh_script | vllm/vllm/cpu_verify/channels/temp_buffer.py:325 |
| 4.23 | LOG_DIR | compressspec | 1 | str(ROOT / ".realmode-run" | test | compressspec/cli.py:447 | |
| 4.24 | LOGPROB_DUMP_DIR | dump_debugger+verify | 2 | ${DUMP_DIR}/logprobs | .env | verify/utils.py:85 | |
| 4.25 | RECOMPACT_TRACE | vllm | 1 | "0" | 1 | jh_script | vllm/vllm/breakdown_logger.py:296 |
| 4.26 | SCORE_DUMP_DIR | dump_debugger+verify+vllm | 7 | "" | ${DUMP_DIR}/score | .env | vllm/vllm/cpu_verify/cpu_verify_engine.py:2480 |
| 4.27 | TOKEN_DUMP_JSON | config | 1 | jh_script | config/common_util.py:129 |
5로그·디버그 · 설정 안 함 (61)
출력·덤프·검증용. 필요할 때 손으로 켜는 것
| # | 이름 | 읽는 곳 | 읽는 횟수 | 기본값 | 러너가 준 값 | 설정한 곳 | 첫 위치 |
|---|---|---|---|---|---|---|---|
| 5.1 | ALLOC_NEW_COMPUTED_DBG | vllm_cpu | 1 | "0" | vllm_cpu/vllm/v1/core/single_type_kv_cache_manager.py:244 | ||
| 5.2 | BREAKDOWN_FLUSH_EVERY | vllm | 1 | "1" | vllm/vllm/breakdown_logger.py:88 | ||
| 5.3 | BREAKDOWN_GPU_EVENT_DEPTH | vllm | 1 | "4" | vllm/vllm/breakdown_logger.py:91 | ||
| 5.4 | BREAKDOWN_LIGHT_LAT | vllm | 1 | "0" | vllm/vllm/breakdown_logger.py:105 | ||
| 5.5 | BREAKDOWN_PER_REQ | vllm | 1 | "0" | vllm/vllm/breakdown_logger.py:89 | ||
| 5.6 | BREAKDOWN_RUN_DIR | breakdown_debugger+vllm | 2 | "" | vllm/vllm/breakdown_logger.py:147 | ||
| 5.7 | BREAKDOWN_SKIP_WARMUP | breakdown_debugger+vllm | 2 | "0" | "1" | vllm/vllm/breakdown_logger.py:159 | ||
| 5.8 | COMPACT_DUMP_INDICES | vllm_cpu | 1 | "0" | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:4298 | ||
| 5.9 | COMPACT_DUMP_KV | vllm_cpu | 1 | "0" | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:4464 | ||
| 5.10 | COMPACT_HOLD_DUMP | vllm_cpu | 1 | vllm_cpu/vllm/compaction/index_map.py:91 | |||
| 5.11 | COMPACTION_IDX_DUMP | vllm_cpu | 1 | "0" | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:4193 | ||
| 5.12 | COMPRESS_SPEC_DEBUG | verify+vllm | 4 | "0" | vllm/vllm/cpu_verify/cspec_debug.py:33 | ||
| 5.13 | COMPRESS_SPEC_TAG | compressspec | 1 | "COMPRESSSPEC_TAG" | compressspec/_tag.py:27 | ||
| 5.14 | COMPRESS_SPEC_TRACE | vllm | 1 | "" | vllm/vllm/cpu_verify/_trace.py:58 | ||
| 5.15 | COMPRESS_SPEC_TRACE_REQ | vllm | 1 | "" | vllm/vllm/cpu_verify/_trace.py:60 | ||
| 5.16 | CPU_ATTN_BUILD_DBG | vllm_cpu | 2 | "0" | vllm_cpu/vllm/v1/attention/backends/cpu_attn.py:195 | ||
| 5.17 | CPU_LOG | compressspec | 1 | str(log_dir / "cpu_server.log" | compressspec/cli.py:449 | ||
| 5.18 | CPU_PROFILE_CHUNK | verify | 1 | "0" | verify/profile_cpu_forward.py:47 | ||
| 5.19 | CPU_PROFILE_CHUNK_ITERS | verify | 1 | "12" | verify/profile_cpu_forward.py:48 | ||
| 5.20 | CPU_PROFILE_COMPACT_RATE | verify | 3 | "" | "0" | verify/profile_cpu_forward.py:65 | ||
| 5.21 | CPU_PROFILE_CTX_TRUNC | verify | 1 | "0" | verify/profile_cpu_forward.py:292 | ||
| 5.22 | CPU_PROFILE_MAX_TOKENS | verify | 1 | "32" | verify/profile_cpu_forward.py:35 | ||
| 5.23 | CPU_PROFILE_PROMPT_LOGPROBS | verify | 1 | verify/profile_cpu_forward.py:52 | |||
| 5.24 | CPU_VERIFY_DUMP_DIR | vllm | 2 | vllm/vllm/cpu_verify/utils.py:650 | |||
| 5.25 | CS_PARITY_ASSERT | vllm_cpu | 1 | "0" | vllm_cpu/vllm/compaction/runner_diff_upload.py:109 | ||
| 5.26 | CS_STOP_LOG | vllm | 1 | "0" | vllm/vllm/v1/core/sched/utils.py:93 | ||
| 5.27 | DRAFT_TOKEN_LOG | verify | 1 | "0" | verify/cpu_verify_server_score.py:455 | ||
| 5.28 | DUMMY_BATCH_DBG | vllm_cpu | 1 | "0" | vllm_cpu/vllm/v1/engine/core.py:1373 | ||
| 5.29 | DUMP_FORWARD_OUT | vllm | 1 | "0" | vllm/vllm/model_executor/models/gemma3.py:373 | ||
| 5.30 | DUMP_LAST_HIDDEN | vllm | 1 | "0" | vllm/vllm/model_executor/models/gemma3.py:532 | ||
| 5.31 | DUMP_LOGITS_OUT | vllm | 1 | "0" | vllm/vllm/model_executor/models/gemma3.py:550 | ||
| 5.32 | DUMP_SAMPLED_IDS | vllm | 1 | "0" | vllm/vllm/v1/worker/gpu_model_runner.py:4213 | ||
| 5.33 | DUMP_TRACKER_DISPATCH | vllm | 1 | "0" | vllm/vllm/cpu_verify/channels/temp_buffer.py:907 | ||
| 5.34 | FAKE_COMPACTION_VERIFY | vllm_cpu | 1 | "0.0" | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:5525 | ||
| 5.35 | FAKE_NEEDS_KV_RELOAD | vllm | 1 | "0" | vllm/vllm/cpu_verify/_fake_verify.py:247 | ||
| 5.36 | FREE_BLOCKS_DBG | vllm_cpu | 1 | "0" | vllm_cpu/vllm/v1/core/block_pool.py:514 | ||
| 5.37 | FWD_TENSOR_STATS | vllm_cpu | 2 | "0" | vllm_cpu/vllm/v1/attention/backends/cpu_attn.py:484 | ||
| 5.38 | INPUT_PROCESSOR_DBG | vllm_cpu | 1 | "0" | vllm_cpu/vllm/v1/engine/input_processor.py:612 | ||
| 5.39 | KEEPMAP_SLOT_LOG | vllm | 1 | "1" | vllm/vllm/cpu_verify/keep_map_reader.py:157 | ||
| 5.40 | KV_DST_PTR_DBG | vllm | 1 | "0" | vllm/vllm/v1/kv_offload/worker/cpu_gpu.py:262 | ||
| 5.41 | KV_DUMP_DIR | vllm+vllm_cpu | 4 | "/tmp/qwg_kvdump" | vllm/vllm/cpu_verify/channels/temp_buffer.py:572 | ||
| 5.42 | KV_DUMP_NCT | vllm | 5 | vllm/vllm/cpu_verify/channels/temp_buffer.py:566 | |||
| 5.43 | KV_DUMP_SEQ | vllm_cpu | 2 | vllm_cpu/vllm/v1/worker/gpu_model_runner.py:5363 | |||
| 5.44 | KV_TRANSFER_DBG | vllm | 1 | "0" | vllm/vllm/distributed/kv_transfer/kv_connector/v1/offloading_naivecopy_connector.py:335 | ||
| 5.45 | LOGPROB_READ_DEBUG | vllm | 1 | "0" | vllm/vllm/cpu_verify/cpu_verify_engine.py:3517 | ||
| 5.46 | LS_PEAK_PROBE | vllm | 1 | "0" | vllm/vllm/v1/core/sched/scheduler.py:1914 | ||
| 5.47 | NAIVE_STORE_DBG | vllm | 1 | "0" | vllm/vllm/distributed/kv_transfer/kv_connector/v1/offloading_naivecopy_connector.py:471 | ||
| 5.48 | NUMA_VERIFY_DEBUG | vllm | 1 | vllm/vllm/v1/kv_offload/backends/temp_cpu.py:80 | |||
| 5.49 | PAYLOAD_BUILD_DBG | vllm | 1 | "0" | vllm/vllm/cpu_verify/cpu_verify_engine.py:2899 | ||
| 5.50 | PR4_8_DBG_EXTENDED | vllm | 1 | "0" | vllm/vllm/cpu_verify/channels/naive_copy.py:327 | ||
| 5.51 | PR4_SW_GUARD_DEBUG | vllm | 1 | "0" | vllm/vllm/v1/attention/backends/flashinfer.py:1841 | ||
| 5.52 | PRESERVE_LOG | vllm | 2 | "0" | vllm/vllm/cpu_verify/channels/temp_buffer.py:516 | ||
| 5.53 | RECOMPACT_FP_LOG | vllm | 1 | "0" | vllm/vllm/cpu_verify/compaction_upload/gpu_recompact.py:26 | ||
| 5.54 | REJECT_DISTRIB_DUMP | verify | 1 | "0" | verify/utils.py:1295 | ||
| 5.55 | RESERVE_DBG | vllm_cpu | 1 | "0" | vllm_cpu/vllm/v1/core/block_pool.py:386 | ||
| 5.56 | SAVE_KV_LAYER_DBG | vllm | 1 | "0" | vllm/vllm/distributed/kv_transfer/kv_connector/v1/offloading_naivecopy_connector.py:2 | ||
| 5.57 | SCORE_LOG | config | 1 | "0" | config/connector_config.py:185 | ||
| 5.58 | TEMPBUF_DBG | vllm | 1 | "0" | vllm/vllm/cpu_verify/channels/temp_buffer.py:666 | ||
| 5.59 | VERIFY_BATCH_LOG | vllm_cpu | 1 | "0" | vllm_cpu/vllm/v1/attention/backends/cpu_attn.py:231 | ||
| 5.60 | VERIFY_DEBUG | verify | 2 | "0" | verify/cpu_verify_server_score.py:114 | ||
| 5.61 | VERIFY_DEBUG_PAYLOAD | verify | 1 | "0" | verify/cpu_verify_server.py:409 |
6실험 하니스 변수 (194)
펼치기: test · scripts · debugger 가 읽는 변수(RULER_ · SAT_ · SWEEP_ 등). 엔진 동작이 아니라 실험 구성용
| # | 이름 | 분류 | 읽는 곳 | 기본값 | 러너가 준 값 | 첫 위치 |
|---|---|---|---|---|---|---|
| 6.1 | AIME_DATA_PATH | functional | scripts | scripts/saturate/lib/aime_data.py:33 | ||
| 6.2 | AIME_N | functional | scripts | "960" | scripts/saturate/lib/aime_data.py:179 | |
| 6.3 | AIME_ONLY_PIDS | functional | scripts | "" | scripts/saturate/lib/aime_data.py:92 | |
| 6.4 | AIME_REPEAT | functional | scripts+test | "32" | 32 | scripts/saturate/lib/aime_data.py:180 | |
| 6.5 | AIME_THINKING | functional | scripts+test | "1" | scripts/saturate/lib/draft_saturate_async.py:421 | |
| 6.6 | AIME_UNIQ_PREFIX | functional | scripts+test | "0" | scripts/saturate/lib/draft_saturate_async.py:424 | |
| 6.7 | ASYNC_BIN_S | functional | dump_debugger+scripts+test | "2.0" | scripts/saturate/lib/draft_saturate_async.py:304 | |
| 6.8 | ATTN_BP_DUMP_DIR | log/debug | test | "" | test/test_score.py:241 | |
| 6.9 | ATTN_BP_DUMP_TS | log/debug | test | "" | test/test_score.py:242 | |
| 6.10 | BATCH | functional | scripts | "1" | $2 | $B | 1 | 4 | scripts/longcontext/lib/hi_lb.py:18 |
| 6.11 | BD_CAP_COLD | functional | breakdown_debugger | "0" | breakdown_debugger/breakdown_utils/timeline_html.py:367 | |
| 6.12 | BLOCK_DEBUG_DUMP_DIR | log/debug | test | "" | test/test_multi_bp.py:85 | |
| 6.13 | BREAKDOWN_CPU_NAME | log/debug | breakdown_debugger | "" | breakdown_debugger/breakdown_utils/latency_plot.py:79 | |
| 6.14 | BREAKDOWN_GPU_NAME | log/debug | breakdown_debugger | "" | breakdown_debugger/breakdown_utils/latency_plot.py:78 | |
| 6.15 | BREAKDOWN_HW_NAME | log/debug | breakdown_debugger | "" | breakdown_debugger/breakdown_util.py:27 | |
| 6.16 | BREAKDOWN_KERNEL_FAMILY | log/debug | breakdown_debugger | "llama" | breakdown_debugger/breakdown_debugger.py:116 | |
| 6.17 | BREAKDOWN_KERNEL_LAYERWISE | log/debug | breakdown_debugger | "0" | breakdown_debugger/breakdown_utils/cpu_inline_kernels.py:174 | |
| 6.18 | BREAKDOWN_PROFILE_STACK | log/debug | breakdown_debugger | "0" | 1 | breakdown_debugger/breakdown_util.py:105 |
| 6.19 | BREAKDOWN_TRACE_WAIT_S | log/debug | breakdown_debugger | "20" | breakdown_debugger/breakdown_util.py:252 | |
| 6.20 | CHAT_MAX_TOKENS | functional | test | "256" | test/chat.py:156 | |
| 6.21 | CHAT_VERBOSE | log/debug | test | "0" | test/chat.py:62 | |
| 6.22 | CHUNK | functional | scripts | "16384" | $CHUNK | scripts/longcontext/lib/hi_lb.py:19 |
| 6.23 | COLLAPSE_S | functional | dump_debugger | "10" | 15 | 999 | dump_debugger/sweep_test.py:46 |
| 6.24 | COMPRESS_METHOD | functional | scripts+test | "snapkv" | "streamingllm" | snapkv | scripts/saturate/lib/draft_saturate_async.py:244 |
| 6.25 | COMPRESSSPEC_BENCH_CACHE | log/debug | test | test/make_datasets.py:68 | ||
| 6.26 | CONDA_PREFIX | functional | dump_debugger | "" | /home2/qwg724/miniconda3/envs/ | dump_debugger/sweep_test.py:36 |
| 6.27 | CONDA_SH | functional | dump_debugger | $LC_CONDA | /home2/qwg724/miniconda3/etc/p | dump_debugger/sweep_test.py:31 | |
| 6.28 | CPU_POOL_GIB | functional | scripts+test | "16" | $kvspace | scripts/saturate/lib/draft_saturate_async.py:199 |
| 6.29 | CSV | functional | scripts | "" | $(find | $CSV | $OUT/hop_probe.csv | $OUT/microbench.csv | scripts/longcontext/lib/hi_lb.py:23 |
| 6.30 | CTX | functional | dump_debugger+scripts | "16384" | $((GEN | $1 | $CTX | $RULER_CTX | scripts/longcontext/lib/hi_lb.py:17 |
| 6.31 | DEBUG_DUMP_TRACKER | log/debug | test | "0" | 1 | 1) | 1. | test/gpu_only_debug.py:96 |
| 6.32 | DEBUG_PATH | log/debug | dump_debugger | dump_debugger/dump_debugger.py:175 | ||
| 6.33 | DIAG_QID | log/debug | test | "1" | test/example_hybrid.py:158 | |
| 6.34 | DISABLE_HYBRID_KVCM | functional | test | "0" | test/gpu_only_debug.py:272 | |
| 6.35 | DISABLE_SLIDING_WINDOW | functional | test | "0" | test/gpu_only_debug.py:271 | |
| 6.36 | DRAFT_BLOCK_SIZE | functional | scripts+test | os.environ.get("BLOCK_SIZE", "32" | scripts/saturate/lib/draft_saturate_async.py:155 | |
| 6.37 | DRAIN_STALL_S | functional | dump_debugger | "240" | 1800 | dump_debugger/ruler_saturate_async_cs.py:165 |
| 6.38 | DRV_CPUS | functional | dump_debugger | "8-15,24-31,40-47,56-63" | ${DRV_CPUS:-16-19} | ${DRV_CPUS:-24-29,56-57} | 24-29,56-57 | nobind | dump_debugger/saturate_run.py:17 |
| 6.39 | DRV_SANITIZER | functional | dump_debugger | dump_debugger/sweep_test.py:245 | ||
| 6.40 | DYN_ROUTER_ADMIT_RESERVE | functional | scripts+test | "2048" | $LC_DYN_ADMIT_RESERVE | scripts/saturate/lib/draft_saturate_async.py:463 |
| 6.41 | ENFORCE_EAGER | functional | dump_debugger+scripts+test | "0" | 1 | scripts/saturate/lib/draft_saturate_async.py:172 |
| 6.42 | EXAMPLE_DIAG_BATCH | log/debug | test | "DIAG_BATCH", "1" | test/example.py:229 | |
| 6.43 | EXP_DECODE_INTERVAL | functional | scripts+test | "2048" | scripts/saturate/lib/draft_saturate_async.py:195 | |
| 6.44 | EXP_SCORE_INFLIGHT | functional | scripts+test | spec_steps * 6 + 32 | scripts/saturate/lib/draft_saturate_async.py:234 | |
| 6.45 | EXP_SCORE_POOL_SIZE | functional | scripts+test | score_pool_size | scripts/saturate/lib/draft_saturate_async.py:237 | |
| 6.46 | EXP_SEED | functional | dump_debugger+scripts+test | "0" | $ST_SEED | ${EXP_SEED:-0} | 0 | 0. | scripts/saturate/lib/draft_saturate_async.py:399 |
| 6.47 | EXP_SEED_PER_REQ | functional | scripts+test | "0" | scripts/saturate/lib/draft_saturate_async.py:407 | |
| 6.48 | EXP_SPEC_STEPS | functional | scripts+test | "16" | scripts/saturate/lib/draft_saturate_async.py:153 | |
| 6.49 | EXP_TEMP | functional | dump_debugger+scripts+test | "0.0" | "0.5" | $ST_TEMP | ${EXP_TEMP:-0.0} | 0.0 | 0.5 | scripts/saturate/lib/draft_saturate_async.py:392 |
| 6.50 | EXP_TEMPERATURE | functional | test | "0.0" | test/test_multi_bp.py:139 | |
| 6.51 | EXP_TOP_K | functional | dump_debugger+scripts+test | "-1" | "0" | "20" | $ST_TOP_K | 20 | scripts/saturate/lib/draft_saturate_async.py:398 |
| 6.52 | EXP_TOP_P | functional | dump_debugger+scripts+test | "0.95" | "1.0" | $ST_TOP_P | 0.95 | scripts/saturate/lib/draft_saturate_async.py:397 |
| 6.53 | FAKE_ALL_ACCEPT_PROB | log/debug | test | "0.0" | <float> | test/gpu_only_hybrid_debug.py:93 |
| 6.54 | FAKE_SELECTED_BLOCK_RATE | log/debug | test | "0.4" | test/gpu_only_debug.py:254 | |
| 6.55 | FSM_DEBUG_DIR | log/debug | breakdown_debugger | "" | $ROOT/results/fsm_A_${TS} | $ROOT/results/fsm_cwe_vt_b4 | ${VLLM_TEST_HOME}/breakdown | breakdown_debugger/breakdown_util.py:55 |
| 6.56 | GPU_MEMORY_UTILIZATION | functional | test | "0.5" | test/chat.py:101 | |
| 6.57 | GPU_MEM_UTIL | functional | dump_debugger+test | "0.8" | "0.9" | ${GPU_MEM_UTIL:-0.9} | 0.70 | 0.8 | 0.8. | test/test_multi_bp.py:123 |
| 6.58 | HI_MODEL | functional | scripts | os.path.expanduser("~/models/Gradient-1048k" | $R/models/ibm-granite/granite- | $R/models/mistralai/Magistral- | $R/models/skt/A.X-3.1 | scripts/longcontext/lib/hi_lb.py:16 |
| 6.59 | HI_SRC | functional | scripts | os.path.join(_REPO, "third-party", "headinfer" | /home2/qwg724/tmp/headinfer_mi | scripts/longcontext/lib/hi_lb.py:7 |
| 6.60 | HYBRID_COMPRESS_MODE | functional | test | "streamingllm" | test/chat.py:89 | |
| 6.61 | HYBRID_MODEL_DIR | functional | test | "google/gemma-3-1b-it" | test/gpu_only_hybrid_debug.py:110 | |
| 6.62 | IGNORE_EOS | functional | dump_debugger+test | "0" | "1" | $EOS | ${IGNORE_EOS:-1} | 0 | 1 | test/test_multi_bp.py:144 |
| 6.63 | ITERS | functional | dump_debugger+scripts+test | "1" | "2" | $IT | 1 | 10 | 5 | scripts/longcontext/lib/hi_lb.py:20 |
| 6.64 | LABEL | functional | scripts | "hi" | 128K | AIME | hi | scripts/longcontext/lib/hi_lb.py:23 |
| 6.65 | LB_DIR | functional | scripts | os.path.join( | $D/ax_docs | $D/granite_docs | $DOCS | scripts/longcontext/lib/hi_lb.py:21 |
| 6.66 | LC_CS_EXTRA | functional | scripts | "" | dir1[,dir2]: | scripts/longcontext/lib/collect_natural.py:77 |
| 6.67 | LC_FONT_DIR | functional | scripts | "" | scripts/longcontext/lib/plot_long.py:22 | |
| 6.68 | LC_LB_DOCS | functional | scripts | /home2/ysshin/exp_cs/neo_1m/lb | /path | scripts/longcontext/lib/extract_lb_docs.py:17 | |
| 6.69 | LC_LB_MANIFEST | functional | scripts | "" | scripts/longcontext/lib/archive_results.py:40 | |
| 6.70 | LC_N_E2E | functional | scripts | "256" | scripts/longcontext/lib/collect.py:13 | |
| 6.71 | LC_TITLE | functional | scripts | "Qwen3-30B-A3B-Thinking (1M window | "Qwen3-4B-Thinking" | scripts/longcontext/lib/plot_long.py:151 | |
| 6.72 | LONGBENCH_ROOT | functional | test | _DEFAULT_ROOT | test/ruler/longbench_data.py:26 | |
| 6.73 | LONGBENCH_TRUNC | functional | test | "middle" | head | test/ruler/longbench_v2_data.py:124 |
| 6.74 | LONGBENCH_V2_DIFFICULTY | functional | test | test/ruler/longbench_v2_data.py:66 | ||
| 6.75 | LONGBENCH_V2_IDS | functional | scripts+test | "" | ",".join(ids) | scripts/longcontext/lib/lb_manifest.py:241 |
| 6.76 | LONGBENCH_V2_LENGTH | functional | test | long | test/ruler/longbench_v2_data.py:60 | |
| 6.77 | LONGBENCH_VER | functional | test | "v1" | v2 | test/ruler/longbench_gate.py:60 |
| 6.78 | MAX_CELL_S | functional | dump_debugger | "360" | 5400 | 800 | dump_debugger/sweep_test.py:48 |
| 6.79 | MAX_GEN_TOKENS | functional | dump_debugger+test | "128" | 128 | 512 | 6 | 64 | test/test_multi_bp.py:144 |
| 6.80 | MAX_TOKENS | functional | test | "32" | "512" | "64" | <int> | test/gpu_only_hybrid_debug.py:135 |
| 6.81 | MODEL_TYPE | functional | dump_debugger+test | os.environ.get("EXP_DTYPE", "bf16" | test/test_multi_bp.py:12 | |
| 6.82 | MULTIBP_MAX_BATCHED_TOKENS | functional | dump_debugger | str(cc.MAX_MODEL_LEN | 8192 | dump_debugger/sweep_test.py:81 |
| 6.83 | N_DECODE | functional | scripts | "16" | $DEC | 16 | scripts/longcontext/lib/hi_lb.py:17 |
| 6.84 | N_E2E | functional | scripts | str(N | 256 | scripts/longcontext/lib/hi_lb.py:97 |
| 6.85 | NEO_DENSIFY_AB | functional | scripts | "0" | 1: | scripts/longcontext/lib/neo_lb.py:117 |
| 6.86 | NEO_DENSIFY_PAST | functional | scripts | "1" if it % 2 else "0" | scripts/longcontext/lib/neo_lb.py:118 | |
| 6.87 | NEO_KV_BACKEND | functional | scripts | "cpu_pool" | cpu_pool | gpu | scripts/longcontext/lib/neo_lb.py:82 |
| 6.88 | NEO_MODEL_NAME | functional | scripts | "llama3_8b" | scripts/longcontext/lib/neo_lb.py:21 | |
| 6.89 | NEO_OWNED_AB | functional | scripts | "0" | 1: | scripts/longcontext/lib/neo_lb.py:111 |
| 6.90 | NEO_OWNED_CPU_POOL | functional | scripts | "1" if it % 2 else "0" | scripts/longcontext/lib/neo_lb.py:112 | |
| 6.91 | NEO_THREAD_SWEEP | functional | scripts | "" | scripts/longcontext/lib/neo_lb.py:101 | |
| 6.92 | NSYS_CAPTURE | functional | test | test/ruler/longbench_gate.py:409 | ||
| 6.93 | NUM_CPU_BLOCKS | functional | test | "64" | test/gpu_only_debug.py:53 | |
| 6.94 | OUT | functional | dump_debugger | "/home2/qwg724/001_CompressSpec/dump_debugger/run10/babilong | "/home2/qwg724/001_CompressSpec/dump_debugger/run10/longbenc | "/home2/qwg724/001_CompressSpec/test/GSK8K" | $(ls | $AG/results/integration | $CLOUD_ROOT/run/${TAG}_$TS | $CS/jh_script/claude_script/re | dump_debugger/run10/gen_niah.py:19 |
| 6.95 | OUT_DUMP | log/debug | dump_debugger | "" | $OUT/$ARM/outputs.jsonl | $R/results/aime_tokens_${PHASE | dump_debugger/ruler_saturate_async_cs.py:78 |
| 6.96 | PACPU_SO | functional | scripts | os.path.join( | $NEO_SRC/pacpu/build/libpacpu- | /home2/qwg724/tmp/neo_ws/pacpu | scripts/longcontext/lib/neo_lb.py:22 |
| 6.97 | PREFIX_CACHING | functional | dump_debugger | "1" | 0 | dump_debugger/ruler_saturate_async_cs.py:72 |
| 6.98 | PROMPT_FILE | functional | test | <path> | test/gpu_only_debug.py:174 | |
| 6.99 | PROMPT_FILE_RAW | functional | test | "1" | test/gpu_only_debug.py:179 | |
| 6.100 | REALMODE_ACCEPT_TOLERANCE | functional | test | "0.5" | test/test_real_mode_e2e.py:44 | |
| 6.101 | REALMODE_ELAPSED_TOLERANCE_SEC | functional | test | "20" | test/test_real_mode_e2e.py:46 | |
| 6.102 | REALMODE_EXPECT_ACCEPT_PCT | functional | test | "100.0" | test/test_real_mode_e2e.py:43 | |
| 6.103 | REALMODE_EXPECT_ELAPSED_SEC | functional | test | "13" | test/test_real_mode_e2e.py:45 | |
| 6.104 | REALMODE_LOG_DIR | log/debug | test | str(PROJECT_ROOT / ".realmode-run" | test/test_real_mode_e2e.py:48 | |
| 6.105 | REF_TOKENS | functional | dump_debugger | "" | dump_debugger/ruler_saturate_async_mix.py:91 | |
| 6.106 | RELOAD_FORCE_SYNC | functional | dump_debugger | "0" | dump_debugger/ruler_saturate_async_cs.py:38 | |
| 6.107 | REP_MAX | functional | dump_debugger | "200" | dump_debugger/ruler_saturate_async_cs.py:86 | |
| 6.108 | REP_MODE | functional | dump_debugger | "shadow" | dump_debugger/ruler_saturate_async_cs.py:90 | |
| 6.109 | REP_NGRAM | functional | dump_debugger | "8" | dump_debugger/ruler_saturate_async_cs.py:85 | |
| 6.110 | ROUTER_RESERVE | functional | scripts+test | "32768" | $LC_ROUTER_RESERVE | scripts/saturate/lib/draft_saturate_async.py:466 |
| 6.111 | RULER_ARMS | functional | dump_debugger | "" | dump_debugger/ruler_saturate_async_mix.py:104 | |
| 6.112 | RULER_CHAT | functional | test | "1" | test/ruler/longbench_gate.py:155 | |
| 6.113 | RULER_CTX_LABEL | functional | dump_debugger | "16K" | ${LABEL:-$CTX} | dump_debugger/ruler_saturate_async_cs.py:8 |
| 6.114 | RULER_DATASET | functional | test | "ruler" | test/ruler/ruler_saturate_async.py:380 | |
| 6.115 | RULER_DUMP_LOGPROBS | log/debug | test | "2" | ${LC_DUMP_LOGPROBS:-2} | test/ruler/longbench_gate.py:371 |
| 6.116 | RULER_DUMP_LOGPROB_DIGITS | log/debug | test | "4" | test/ruler/longbench_gate.py:446 | |
| 6.117 | RULER_DUMP_OUT | log/debug | test | 1 | test/ruler/longbench_gate.py:370 | |
| 6.118 | RULER_DUMP_TOKENS | log/debug | test | "0" | 1: | test/ruler/ruler_saturate_async.py:442 |
| 6.119 | RULER_GPU_MEM_UTIL | functional | scripts+test | "0.8" | "0.9" | $GPU_MEM_UTIL | 0.8 | 0.85 | scripts/longcontext/lib/router.py:117 |
| 6.120 | RULER_HYBRID_KVCM | functional | test | test/ruler/ruler_saturate_async.py:159 | ||
| 6.121 | RULER_IGNORE_EOS | functional | test | "0" | "1" | 0 | 1 | test/ruler/longbench_gate.py:365 |
| 6.122 | RULER_KV_SEQ_LEN | functional | test | ctx + 1024 | test/ruler/ruler_saturate_async.py:193 | |
| 6.123 | RULER_LOGPROBS_MODE | functional | test | processed_logprobs | test/ruler/longbench_gate.py:206 | |
| 6.124 | RULER_MAX_BATCHED_TOKENS | functional | dump_debugger+scripts+test | "2048" | "8192" | max(max_model_len, 25300 | 8192 | scripts/longcontext/lib/router.py:115 |
| 6.125 | RULER_MAX_MODEL_LEN | functional | test | max(4096, ctx + 1024 | test/ruler/longbench_gate.py:167 | |
| 6.126 | RULER_MAX_NEW_TOKENS | functional | dump_debugger+test | "" | "512" | 512 | $OUT_TOKENS | none): | test/ruler/longbench_gate.py:362 |
| 6.127 | RULER_MODE | functional | test | "compress" | compress | vanilla | test/ruler/longbench_gate.py:149 |
| 6.128 | RULER_NO_TRUNC | functional | test | "0" | 1 | 1). | 1: | test/ruler/longbench_gate.py:351 |
| 6.129 | RULER_OUT_DIR | functional | dump_debugger+test | "." | _REPO | $WORK | test/ruler/longbench_gate.py:168 |
| 6.130 | RULER_PREFIX_CACHING | functional | test | 0 | test/ruler/longbench_gate.py:199 | |
| 6.131 | RULER_Q_FILES | functional | dump_debugger | "" | dump_debugger/ruler_saturate_async_mix.py:53 | |
| 6.132 | RULER_Q_LIST | functional | dump_debugger | "" | dump_debugger/ruler_saturate_async_cs.py:46 | |
| 6.133 | RULER_Q_START | functional | dump_debugger | "1" | dump_debugger/ruler_saturate_async_cs.py:43 | |
| 6.134 | RULER_REQS | functional | dump_debugger+test | "512" | _env_int("RULER_POOL", batch | # | test/ruler/ruler_saturate_async.py:371 |
| 6.135 | RULER_ROUTER | functional | test | "0" | 1 | 1, | test/ruler/longbench_gate.py:311 |
| 6.136 | RULER_SAMPLES | functional | test | None | $SAMPLES | ${FIXED_SAMPLES:-$B} | 2*B | 32 | test/ruler/longbench_gate.py:165 |
| 6.137 | RULER_SORT_BY_LEN | functional | test | 1 | 1: | test/ruler/longbench_gate.py:342 | |
| 6.138 | RULER_TAG | log/debug | dump_debugger+test | "" | f"b{B}" | $LC_DUMP_OUT | $tag | test/ruler/longbench_gate.py:173 |
| 6.139 | RULER_TASKS | functional | test | "all" | $TASK | all | code_repo | niah_single_1,qa_1 | test/ruler/longbench_gate.py:163 |
| 6.140 | RULER_WARMUP_TOKENS | functional | test | fixed_out | test/ruler/ruler_saturate_async.py:416 | |
| 6.141 | RUN_NUMA | functional | dump_debugger | "0,1,2,3" | dump_debugger/saturate_run.py:18 | |
| 6.142 | SAT_BATCH | functional | scripts | $B | batch(=max_num_seqs), | scripts/saturate/lib/draft_saturate_async.py:152 | |
| 6.143 | SAT_CTX | functional | scripts | "16384" | $ST_AIME_MAX_PROMPT | scripts/saturate/lib/draft_saturate_async.py:151 |
| 6.144 | SAT_DUMP_TOKENS | log/debug | scripts | "0" | 1: | scripts/saturate/lib/draft_saturate_async.py:491 |
| 6.145 | SAT_GPU_MEM_UTIL | functional | scripts | "0.8" | scripts/saturate/lib/draft_saturate_async.py:154 | |
| 6.146 | SAT_HYBRID_KVCM | functional | scripts | 0 | 1 | scripts/saturate/lib/draft_saturate_async.py:174 | |
| 6.147 | SAT_IGNORE_EOS | functional | scripts | "1" | $ST_IGNORE_EOS | scripts/saturate/lib/draft_saturate_async.py:402 |
| 6.148 | SAT_KV_SEQ_LEN | functional | scripts | ctx + 1024 | scripts/saturate/lib/draft_saturate_async.py:208 | |
| 6.149 | SAT_MAX_BATCHED_TOKENS | functional | scripts | max(max_model_len, 25300 | scripts/saturate/lib/draft_saturate_async.py:165 | |
| 6.150 | SAT_MAX_MODEL_LEN | functional | scripts | max(4096, ctx + 1024 | scripts/saturate/lib/draft_saturate_async.py:159 | |
| 6.151 | SAT_MAX_NEW_TOKENS | functional | scripts | 512 | $ST_OUT_TOKENS | scripts/saturate/lib/draft_saturate_async.py:401 |
| 6.152 | SAT_MODE | functional | scripts | "compress" | ${ST_MODE:-compress} | scripts/saturate/lib/draft_saturate_async.py:181 |
| 6.153 | SAT_OUT_DIR | functional | scripts | _REPO | $WORK | scripts/saturate/lib/draft_saturate_async.py:390 |
| 6.154 | SAT_OUT_NAME | functional | dump_debugger | "saturate" | dump_debugger/saturate_run.py:30 | |
| 6.155 | SAT_PREFIX_CACHING | functional | scripts | "0" | 1. | scripts/saturate/lib/draft_saturate_async.py:168 |
| 6.156 | SAT_REQS | functional | scripts | _env_int("SAT_POOL", batch | # | $_reqs | scripts/saturate/lib/draft_saturate_async.py:389 |
| 6.157 | SAT_ROUTER | functional | scripts | "0" | scripts/saturate/lib/draft_saturate_async.py:454 | |
| 6.158 | SAT_TAG | log/debug | scripts | "" | $tag | scripts/saturate/lib/draft_saturate_async.py:391 |
| 6.159 | SAT_UNIQ_PREFIX | functional | scripts | "0" | 1 | scripts/longcontext/lib/hi_lb.py:47 |
| 6.160 | SAT_WARMUP_TOKENS | functional | scripts | fixed_out | scripts/saturate/lib/draft_saturate_async.py:430 | |
| 6.161 | SEQUENTIAL | functional | dump_debugger | "0" | 1 | dump_debugger/ruler_saturate_async_cs.py:22 |
| 6.162 | SERVER_ENV | functional | dump_debugger | "verify_L40S" | ${SERVER_ENV:-verify} | dump_debugger/saturate_run.py:53 |
| 6.163 | SNAPKV_AGG | functional | scripts+test | "mean" | scripts/saturate/lib/draft_saturate_async.py:249 | |
| 6.164 | SNAPKV_KERNEL | functional | scripts+test | "7" | scripts/saturate/lib/draft_saturate_async.py:248 | |
| 6.165 | SNAPKV_POOLING | functional | scripts+test | "maxpool" | scripts/saturate/lib/draft_saturate_async.py:250 | |
| 6.166 | SNAPKV_SINK_BLOCKS | functional | scripts+test | "0" | scripts/saturate/lib/draft_saturate_async.py:251 | |
| 6.167 | SNAPKV_WINDOW | functional | scripts+test | "32" | scripts/saturate/lib/draft_saturate_async.py:247 | |
| 6.168 | SPLIT | functional | dump_debugger | "test" | dump_debugger/run10/gen_gsm8k.py:15 | |
| 6.169 | SRV_BIND | functional | dump_debugger | "0-7,16-23,32-39,48-55" | ${SRV_BIND:-0-15} | ${SRV_BIND:-0-5,10-11,16-23,32 | 0-5,10-11,16-23,32-37,42-43,48 | 24-55 | dump_debugger/saturate_run.py:16 |
| 6.170 | STALL_S | functional | dump_debugger | "180" | 400 | 480 | 500 | 5400 | dump_debugger/sweep_test.py:47 |
| 6.171 | START_ITER | functional | dump_debugger | "0" | 0 | 4 | dump_debugger/sweep_test.py:96 |
| 6.172 | STREAMINGLLM_SINK | functional | test | "64" | test/chat.py:91 | |
| 6.173 | SWEEP_BATCHES | functional | test | "1,2,4,8" | test/throughput_sweep.py:24 | |
| 6.174 | SWEEP_DRV_CPUS | functional | dump_debugger | "" | 56-63 | 8-15,24-31,40-47,56-63 | dump_debugger/sweep_test.py:236 |
| 6.175 | SWEEP_DRV_OMP | functional | dump_debugger | dump_debugger/sweep_test.py:275 | ||
| 6.176 | SWEEP_DUMP_TOKENS | log/debug | dump_debugger | dump_debugger/sweep_test.py:113 | ||
| 6.177 | SWEEP_MAXTOK | functional | test | "128" | test/throughput_sweep.py:26 | |
| 6.178 | SWEEP_OUT | functional | test | os.path.join( | test/throughput_sweep.py:27 | |
| 6.179 | SWEEP_OUT_ROOT | functional | dump_debugger | $OUT | $R512 | dump_debugger/sweep_test.py:42 | |
| 6.180 | SWEEP_SEQS | functional | test | "4096,8192,16384" | test/throughput_sweep.py:25 | |
| 6.181 | SWEEP_SERVER_ENV | functional | dump_debugger | "verify_L40S" | dump_debugger/sweep_test.py:43 | |
| 6.182 | SWEEP_SRV_BIND | functional | dump_debugger | "" | 0-55 | 0-7,16-23,32-39,48-55 | dump_debugger/sweep_test.py:227 |
| 6.183 | SWEEP_SRV_ISA | functional | dump_debugger | dump_debugger/sweep_test.py:254 | ||
| 6.184 | SWEEP_SRV_OMP | functional | dump_debugger | dump_debugger/sweep_test.py:253 | ||
| 6.185 | SWEEP_SRV_SPIN | functional | dump_debugger | dump_debugger/sweep_test.py:255 | ||
| 6.186 | SWEEP_TEMPERATURE | functional | dump_debugger | "0.0" | dump_debugger/sweep_test.py:92 | |
| 6.187 | TASKS | functional | dump_debugger | "qa2,qa3,qa4,qa5" | dump_debugger/run10/gen_babilong.py:14 | |
| 6.188 | VLLM_CPU_VERIFY_TRACKER_JSON_DIR | functional | dump_debugger | dump_debugger/logprobs_debugger.py:30 | ||
| 6.189 | WAIT_READY_S | functional | dump_debugger | "420" | 3600 | dump_debugger/saturate_run.py:57 |
| 6.190 | WARMUP | functional | scripts+test | "0" | "1" | $W | 0 | 0으로 | 1 | scripts/longcontext/lib/hi_lb.py:20 |
| 6.191 | WARMUP_COUNT | functional | test | "0" | test/test_multi_bp.py:48 | |
| 6.192 | WARMUP_CTX | functional | test | "" | 4096 | test/test_multi_bp.py:56 |
| 6.193 | WARMUP_MAX_TOKENS | functional | test | str(sampling_params.max_tokens | 6 | test/test_multi_bp.py:152 |
| 6.194 | WINDOW_S | functional | breakdown_debugger | "0" | breakdown_debugger/breakdown_debugger.py:169 |
7엔진 변수 접두어별 개수
| 접두어 | 개수 |
|---|---|
| COMPACT | 37 |
| VERIFY | 19 |
| COMPRESS | 15 |
| CPU | 14 |
| RS | 12 |
| COMPSPEC | 11 |
| BREAKDOWN | 9 |
| LOGPROB | 9 |
| LS | 8 |
| RULER | 8 |
| SCORE | 7 |
| ATTN | 6 |
| DUMP | 6 |
| CS | 5 |
| KV | 5 |
| LAYER | 5 |
| MULTIBP | 5 |
| REJECT | 5 |
| COMPRESSION | 4 |
| DYN | 4 |
| NUMA | 4 |
| COMPACTION | 3 |
| DIFF | 3 |
| DRAFT | 3 |
| KEEPMAP | 3 |
| PR4 | 3 |
| RECOMPACT | 3 |
| SNAPKV | 3 |
| VLLM | 3 |
| BENCH | 2 |
| EXP | 2 |
| FAKE | 2 |
| GPU | 2 |
| HEADINFER | 2 |
| LOG | 2 |
| MAX | 2 |
| RELAX | 2 |
| RESERVE | 2 |
| TAS | 2 |
| ACCEPT | 1 |
| ALLOC | 1 |
| ARM | 1 |
| BLOCK | 1 |
| CACTUS | 1 |
| CP | 1 |
| DEBUG | 1 |
| DELTA | 1 |
| DENSE | 1 |
| DUMMY | 1 |
| EXAMPLE | 1 |
| FORCE | 1 |
| FREE | 1 |
| FSM | 1 |
| FWD | 1 |
| HELD | 1 |
| HI | 1 |
| INIT | 1 |
| INPUT | 1 |
| LAST | 1 |
| LSE | 1 |
| MAXPOS | 1 |
| MINIFORGE | 1 |
| MODEL | 1 |
| NAIVE | 1 |
| NUM | 1 |
| OFFLOAD | 1 |
| OUT | 1 |
| PAYLOAD | 1 |
| PORT | 1 |
| PRESERVE | 1 |
| READY | 1 |
| REUSE | 1 |
| SAVE | 1 |
| SEED | 1 |
| SHARED | 1 |
| SKIP | 1 |
| SMALL | 1 |
| SPECULATIVE | 1 |
| STREAMINGLLM | 1 |
| TASK | 1 |
| TEMPBUF | 1 |
| TOKEN | 1 |
| UPLOAD | 1 |