29Environment and configuration reference

TandemLLM reads four families of settings. Engine switches are QWEN38_* variables, read through one registry. Server settings are command-line flags (Chapter 28); ops/start.sh builds them from ops/serve.env, and a few QSE_* variables set flags’ defaults or serve the scripts. The proxy, the collector and the plugins read TANDEM_* variables. And the conductor’s behaviour lives in its versioned documents, which the dashboard edits.

29.1 The served profile

ops/serve.env is the served configuration. On an installed engine it is ~/TandemLLM/run/ops/serve.env, rewritten by every run of the installer, and local.env beside it is read last for your own values.

Key	Value	Kind
HOME_DIR	(a path)	read by ops/start.sh
REPO	(a path)	read by ops/start.sh
PY	(a path)	read by ops/start.sh
PYTHONPATH	(a path)	read by ops/start.sh
HOST	0.0.0.0	read by ops/start.sh
PORT	8000	read by ops/start.sh
SERVED_MODEL	qwen38-spark-engine	read by ops/start.sh
NV	(a path)	read by ops/start.sh
HEAD	(a path)	read by ops/start.sh
CKPT8	(a path)	read by ops/start.sh
CKPT16	(a path)	read by ops/start.sh
CORPUS	(a path)	read by ops/start.sh
MAX_LEN	262144	read by ops/start.sh
DEFAULT_MAX_TOKENS	32768	read by ops/start.sh
PATTERN_STOP	(empty)	read by ops/start.sh
CONDUCTOR	on	read by ops/start.sh
REASONING_FORMAT	tags	read by ops/start.sh
REASONING_EFFORT	medium	read by ops/start.sh
LEN_FIXED	0	read by ops/start.sh
LEN_LATCH	1	read by ops/start.sh
BUDGET	16	read by ops/start.sh
MEM_CEILING_GIB	70	read by ops/start.sh
MEM_PEAK_GIB	77	read by ops/start.sh
RESERVE_SLACK_GIB	2	read by ops/start.sh
RESIDENT_GB	2.5	read by ops/start.sh
RESIDENT_STASH_GB	1.0	read by ops/start.sh
RESIDENT_GUEST_MAX	8192	read by ops/start.sh
RESIDENT_PIN_GB	4.0	read by ops/start.sh
RESIDENT_PARK_GB	4	read by ops/start.sh
CACHE_GB	2	read by ops/start.sh
STORE_ENTRY_GB	1	read by ops/start.sh
NARROW_ARM	1	read by ops/start.sh
END_ANCHOR	on	read by ops/start.sh
END_ALIGN	on	read by ops/start.sh
DROP_PAGE_CACHE	1	read by ops/start.sh
REQUEST_TIMEOUT	3600	read by ops/start.sh
MAX_QUEUE	8	read by ops/start.sh
QUEUE_TIMEOUT	240	read by ops/start.sh
FUSE_PROJ	0	read by ops/start.sh
QWEN38_DF2_TREE_MODE	nodes	engine switch
QWEN38_TREE_ALIAS_STATE	0	engine switch
DROP_IDLE	1	read by ops/start.sh
QWEN38_DRAFT_TEMP	0.3	engine switch
SAMPLED_TREE	det	read by ops/start.sh
QWEN38_FUSED_GDNPREFILL	1	engine switch
QWEN38_DECODE_ATTN	1	engine switch
QWEN38_DRAFT_NVFP4	1	engine switch
QWEN38_NVFP4_PREFILL_V2	1	engine switch
QWEN38_NVFP4_PREFILL_V2_UNTIL	1025	engine switch
QWEN38_NVFP4_SKINNY	1	engine switch
QWEN38_SKINNY_TILES	(a path)	engine switch
QWEN38_FUSED_COMMIT	1	engine switch
QWEN38_FUSED_GDNVERIFY	1	engine switch
QWEN38_DRAFT_FC_NVFP4	1	engine switch
QWEN38_FUSED_ADDNORM	1	engine switch
QWEN38_TREE_HOST_DEPTH	1	engine switch
QWEN38_HOST_WALK	1	engine switch
QWEN38_VERIFY_GRAPH	1	engine switch
QWEN38_COMMIT_IN_VERIFY	1	engine switch
QWEN38_DRAFT_GRAPH	1	engine switch
QWEN38_GDNV_ONE_WARP	1	engine switch
QWEN38_TREE_NODES_NARROW	16	engine switch
QWEN38_TREE_MS	8:77.9,16:78.9,24:87.0,32:97.0	engine switch
QWEN38_VERIFY_ROWS	32	engine switch
QWEN38_SKINNY_TILES_WIDE	(a path)	engine switch
QWEN38_TREE_NODES	24	engine switch
QWEN38_TREE_WIDE_AFTER	32	engine switch
QWEN38_DEEP	32	engine switch
QWEN38_DEEP_AFTER	2	engine switch
QWEN38_LEN_SWITCH	1	engine switch
QWEN38_LEN_MODE	wide	engine switch
QWEN38_STAIR_TABLES	(a path)	engine switch
QWEN38_STAIR_RHO	(a path)	engine switch
QWEN38_STAIR_SKIP	1	engine switch
QWEN38_SKINNY_SRUN	1	engine switch
QSE_USAGE_LEDGER	(a path)	server setting
QSE_CONDUCTOR_CONFIG	(a path)	server setting
QSE_BACKENDS	(empty)	server setting
QSE_PARALLEL	1	server setting
QSE_SIDE_MAX_LEN	65536	server setting
QSE_PARALLEL_FLOOR_GB	15	server setting
QWEN38_GDNV_CONV_BLOCK	32	engine switch
QWEN38_GDNV_CONV_WARPS	1	engine switch
QWEN38_GDNV_WY	1	engine switch
QWEN38_GDNV_WY_FUSED	1	engine switch
QWEN38_GDNV_WY_MAXT	16	engine switch
QWEN38_GDNV_WY_CHAIN_MAXT	16	engine switch
QWEN38_GDNV_WY_FUSED_MAXT	16	engine switch
LIVE_ACTIVITY	on	read by ops/start.sh
QSE_DASHBOARD_LOGIN	off	server setting
ROLES_AUTOLOAD	(empty)	read by ops/start.sh
QWEN38_DECIDER_MODELS	(empty)	engine switch
QWEN38_DECIDER_DIR	(a path)	engine switch
QWEN38_NARRATOR_MODEL	(empty)	engine switch
QWEN38_NARRATOR_DIR	(empty)	engine switch
QWEN38_SHADOW_MODEL	(empty)	engine switch
QWEN38_SHADOW_DIR	(empty)	engine switch
QWEN38_SHADOW_GB	4	engine switch
QWEN38_SHADOW_GRID	512	engine switch
QWEN38_SHADOW_DISK	auto	engine switch
QWEN38_SHADOW_DISK_GB	20	engine switch
QWEN38_PROBE_SNAPSHOTS	off	engine switch
QWEN38_PROBE_DIR	(a path)	engine switch
QSE_DISK_CACHE	(a path)	server setting
SESSION_SPILL	on	read by ops/start.sh
SESSION_SPILL_GB	64	read by ops/start.sh
SESSION_SPILL_CODEC	zstd	read by ops/start.sh
DISK_WRITE_BUDGET_GB	200	read by ops/start.sh
VISION	on	read by ops/start.sh
RESPONSE_CACHE	off	read by ops/start.sh
QSE_STATUS_PORT	(empty)	server setting
QSE_PROFILE_NAME	(a path)	server setting
QSE_SETUP_ENV_FILE	(a path)	server setting

29.2 Engine switches

engine/settings.py lists every engine switch with its default and the module that reads it; tests/test_settings.py fails when a module reads one another way. The table shows each switch, its code default, and the value the served profile sets when it sets one.

Variable	Code default	Served value • read in
QWEN38_ATTN_STAGES3	1	tools/attn_kernels.py
QWEN38_BLOCKING_SYNC	0	server/app.py
QWEN38_CLIENT_CWD	note	server/conductor/router.py
QWEN38_CLIENT_REASONING	*=reasoning_content,hermes=tags	server/compat.py
QWEN38_COMMIT_IN_VERIFY	0	1 • engine/model.py
QWEN38_CONDUCTOR_FIT	model	server/conductor/modelfit.py
QWEN38_CORPUS	none	engine/drafters/ngram.py
QWEN38_DECIDER_COLLATE	call	server/decider_api.py
QWEN38_DECIDER_DEVICE	auto	server/decider_api.py
QWEN38_DECIDER_DIR	~/.qwen38-spark-engine/deciders/laya	(a path) • server/decider_api.py
QWEN38_DECIDER_MAX_BATCH	64	server/decider_api.py
QWEN38_DECIDER_MAX_QUEUE	256	server/decider_api.py
QWEN38_DECIDER_MODELS	none	server/decider_api.py
QWEN38_DECIDER_WAIT_MS	2	server/decider_api.py
QWEN38_DECODE_ATTN	0	1 • engine/model.py
QWEN38_DEEP	0	32 • server/app.py
QWEN38_DEEP_AFTER	2	2 • server/app.py
QWEN38_DF2_TEMP	1.0	engine/drafters/dflash2.py
QWEN38_DF2_TREE_MODE	paths	nodes • engine/drafters/dflash2.py
QWEN38_DFLASH2	none	engine/drafters/dflash2.py
QWEN38_DRAFT_EXL3	none	engine/drafters/dflash2.py
QWEN38_DRAFT_FC_NVFP4	0	1 • engine/drafters/dflash2.py
QWEN38_DRAFT_GRAPH	0	1 • engine/drafters/dflash2.py
QWEN38_DRAFT_HEAD	none	engine/drafters/dflash2.py
QWEN38_DRAFT_HEAD_EXL3	none	engine/drafters/dflash2.py
QWEN38_DRAFT_HEAD_NVFP4	0	engine/drafters/dflash2.py
QWEN38_DRAFT_NVFP4	0	1 • engine/drafters/dflash2.py
QWEN38_DRAFT_TEMP	1.0	0.3 • engine/drafters/dflash2.py
QWEN38_DSPARK	none	engine/drafters/dspark.py
QWEN38_DSPARK_NO_YARN	none	engine/drafters/dspark.py
QWEN38_END_ALIGN_MOVED	on	server/endalign.py
QWEN38_END_ALIGN_REASONING	on	server/endalign.py
QWEN38_EXL3	none	engine/loader.py
QWEN38_EXL3_CHUNK_UNTIL	128	tools/exl3_linear.py
QWEN38_EXL3_FUSE	1	engine/loader.py
QWEN38_EXL3_MLP_BRIDGE	1	tools/exl3_linear.py
QWEN38_EXL3_ROWGRID	1	tools/exl3_linear.py
QWEN38_EXL3_SILU_FUSED	0	tools/exl3_linear.py
QWEN38_EXL3_STAGES	3	tools/exl3_linear.py
QWEN38_EXL3_TILES	none	tools/exl3_linear.py
QWEN38_FN_DENSE	0	engine/models/flashnext/engine.py
QWEN38_FN_DRAFT_HOT	32768	engine/models/flashnext/serving.py
QWEN38_FN_DROP_SHARD_CACHE	1	engine/models/flashnext/load.py
QWEN38_FN_FOLD_COMMIT	1	engine/models/flashnext/engine.py
QWEN38_FN_GEMV_SPLIT	1	tools/fn_kernels.py
QWEN38_FN_LOOKUP	0	engine/models/flashnext/serving.py
QWEN38_FN_MIX	int8	engine/models/flashnext/pack.py
QWEN38_FN_MOE_ST	3	tools/fn_kernels.py
QWEN38_FN_MOE_WARPS	4	tools/fn_kernels.py
QWEN38_FN_PDL	1	tools/fn_kernels.py
QWEN38_FN_PREFILL	kernels	engine/models/flashnext/engine.py
QWEN38_FN_PREFILL_CHUNK	4096	engine/models/flashnext/engine.py
QWEN38_FN_PREFILL_MM	fp32	tools/fn_prefill_kernels.py
QWEN38_FN_TABLE	warm	engine/models/flashnext/load.py
QWEN38_FN_TABLE_CACHE_GB	2	engine/models/flashnext/load.py
QWEN38_FN_WARM_TABLE	1	engine/models/flashnext/load.py
QWEN38_FP8_HEAD	none	engine/loader.py
QWEN38_FP8_TILES	none	tools/fp8_linear.py
QWEN38_FUSED_ADDNORM	0	1 • engine/model.py
QWEN38_FUSED_ATTN	0	engine/model.py
QWEN38_FUSED_ATTN_PREP	0	engine/model.py
QWEN38_FUSED_COMMIT	0	1 • engine/model.py
QWEN38_FUSED_GDN	1	engine/model.py
QWEN38_FUSED_GDNBLOCK	1	engine/model.py
QWEN38_FUSED_GDNPRE	1	engine/model.py
QWEN38_FUSED_GDNPREFILL	0	1 • engine/model.py
QWEN38_FUSED_GDNTREE	1	engine/model.py
QWEN38_FUSED_GDNVERIFY	0	1 • engine/model.py
QWEN38_FUSED_HEAD	1	engine/model.py
QWEN38_FUSED_NORM	1	engine/model.py
QWEN38_FUSE_PROJ	0	tools/nvfp4_linear_v2.py
QWEN38_GDNV_BV	16	tools/gdn_verify_kernels.py
QWEN38_GDNV_CONV_BLOCK	256	32 • tools/gdn_verify_kernels.py
QWEN38_GDNV_CONV_WARPS	4	1 • tools/gdn_verify_kernels.py
QWEN38_GDNV_ONE_WARP	0	1 • tools/gdn_verify_kernels.py
QWEN38_GDNV_TREE_PF	0	tools/gdn_verify_kernels.py
QWEN38_GDNV_WARPS	4	tools/gdn_verify_kernels.py
QWEN38_GDNV_WY	0	1 • tools/gdn_verify_kernels.py
QWEN38_GDNV_WY_CHAIN_MAXT	32	16 • tools/gdn_verify_kernels.py
QWEN38_GDNV_WY_FUSED	0	1 • tools/gdn_verify_kernels.py
QWEN38_GDNV_WY_FUSED_MAXT	32	16 • tools/gdn_verify_kernels.py
QWEN38_GDNV_WY_KC	0	tools/gdn_verify_kernels.py
QWEN38_GDNV_WY_MAXT	32	16 • tools/gdn_verify_kernels.py
QWEN38_GDN_AB	0	engine/model.py
QWEN38_GDN_CHUNK	64	engine/model.py
QWEN38_GDN_MM	fp32	engine/gdn.py
QWEN38_GDN_MM_FROM	64	engine/gdn.py
QWEN38_GDN_PREFILL_BV	32	tools/gdn_prefill_kernels.py
QWEN38_GDN_PREFILL_NOREP	0	engine/model.py
QWEN38_GDN_PREFILL_PREC	bf16x3	tools/gdn_prefill_kernels.py
QWEN38_GDN_PREFILL_PREC_A	tf32	tools/gdn_prefill_kernels.py
QWEN38_GDN_PREFILL_PREC_SCAN	none	tools/gdn_prefill_kernels.py
QWEN38_GDN_PREFILL_SERIES	0	tools/gdn_prefill_kernels.py
QWEN38_GDN_PREFILL_STAGES	1	tools/gdn_prefill_kernels.py
QWEN38_GDN_PREFILL_WARPS_INTRA	8	tools/gdn_prefill_kernels.py
QWEN38_GDN_PREFILL_WARPS_SCAN	8	tools/gdn_prefill_kernels.py
QWEN38_GQA_FROM	64	engine/model.py
QWEN38_GRAPH_MAX_CTX	32768	engine/verify_graph.py
QWEN38_HEAD_GEMM	none	tools/head_gemv.py
QWEN38_HOST_ASYNC	0	engine/model.py
QWEN38_HOST_WALK	0	1 • engine/drafters/dflash2.py
QWEN38_KV_FP8	0	engine/model.py
QWEN38_LATCH_PRICE	0	engine/lenrouter.py
QWEN38_LAUNCH_FIRST	0	server/app.py
QWEN38_LEN_MODE	none	wide • engine/lenrouter.py
QWEN38_LEN_SWITCH	0	1 • engine/lenrouter.py
QWEN38_LOWER_RIGHT_FROM	64	engine/model.py
QWEN38_MEMORY_PROXIES	none	server/app.py
QWEN38_MEMORY_WRITER_TEXT	none	server/app.py
QWEN38_MODEL	none	engine/config.py
QWEN38_MODEL_NAME	none	server/conductor/schema.py
QWEN38_NARRATOR_DEVICE	auto	server/conductor/narrator.py
QWEN38_NARRATOR_DIR	none	server/conductor/narrator.py
QWEN38_NARRATOR_MODEL	none	server/conductor/narrator.py
QWEN38_NOTE_ANCHOR	on	server/conductor/router.py
QWEN38_NOTE_DEDUP	parts	server/conductor/router.py
QWEN38_NVFP4	none	engine/loader.py
QWEN38_NVFP4_BLOCK_TILES	0	tools/nvfp4_linear.py
QWEN38_NVFP4_DEQUANT_FROM	512	tools/nvfp4_linear.py
QWEN38_NVFP4_PREFILL_V2	0	1 • tools/nvfp4_linear.py
QWEN38_NVFP4_PREFILL_V2_UNTIL	1024	1025 • tools/nvfp4_linear.py
QWEN38_NVFP4_SKINNY	0	1 • tools/nvfp4_skinny.py
QWEN38_NVFP4_SPLITK	0	tools/nvfp4_linear.py
QWEN38_NVFP4_V2	1	tools/nvfp4_linear_v2.py
QWEN38_NVFP4_V2_BN	0	tools/nvfp4_linear_v2.py
QWEN38_NVFP4_V2_DOTS	1	tools/nvfp4_linear_v2.py
QWEN38_NVFP4_V2_MAX	32	tools/nvfp4_linear_v2.py
QWEN38_NVFP4_V2_MIN	1	tools/nvfp4_linear_v2.py
QWEN38_NVFP4_V2_PREFETCH	0	tools/nvfp4_linear_v2.py
QWEN38_NVFP4_V2_W	0	tools/nvfp4_linear_v2.py
QWEN38_PREFILL_CAUSAL	1	engine/model.py
QWEN38_PREFILL_WINDOW	8192	engine/cache.py
QWEN38_PREFILL_WINDOW_FREE	32768	engine/cache.py
QWEN38_PROBE_DAY_MB	100	server/probe.py
QWEN38_PROBE_DIR	~/.qwen38-spark-engine/probe	(a path) • server/probe.py
QWEN38_PROBE_EVERY	1024	server/probe.py
QWEN38_PROBE_SNAPSHOTS	off	off • server/probe.py
QWEN38_RANKK	1	engine/model.py
QWEN38_REGENERATE	sample	server/app.py
QWEN38_REGEN_TEMPERATURE	0.7	server/app.py
QWEN38_REGEN_TOP_K	20	server/app.py
QWEN38_REGEN_TOP_P	0.95	server/app.py
QWEN38_REREAD_SAY_TOKENS	8192	server/app.py
QWEN38_SAMPLED_CUT	0	engine/router.py
QWEN38_SAMPLE_BATCH	1	engine/sample.py
QWEN38_SAMPLE_NUCLEUS	1	engine/sample.py
QWEN38_SAMPLE_TOPP_K	1024	engine/sample.py
QWEN38_SCALE_ON	weight	tools/fp8_linear.py
QWEN38_SESSION_GENERATIONS	on	server/conductor/router.py
QWEN38_SHADOW_DIR	none	server/conductor/shadow.py
QWEN38_SHADOW_DISK	none	auto • server/conductor/shadow.py
QWEN38_SHADOW_DISK_GB	20	20 • server/conductor/shadow.py
QWEN38_SHADOW_GB	4	4 • server/conductor/shadow.py
QWEN38_SHADOW_GRID	512	512 • server/conductor/shadow.py
QWEN38_SHADOW_MODEL	none	server/conductor/shadow.py
QWEN38_SHADOW_POOL_SLACK_GIB	2	server/conductor/shadow.py
QWEN38_SIDE_CARRY	on	server/app.py
QWEN38_SIDE_CONTINUE	view	server/app.py
QWEN38_SIDE_CONV	on	server/app.py
QWEN38_SIDE_DISK	read	server/app.py
QWEN38_SIDE_GUESTS	on	server/app.py
QWEN38_SIDE_HOME_WAIT_S	10	server/app.py
QWEN38_SIDE_MIN_ROOM	16384	server/app.py
QWEN38_SIDE_NEW_FIRST	on	server/app.py
QWEN38_SIDE_NEW_WAIT_S	20	server/app.py
QWEN38_SIDE_RESIDENT_GB	0.5	server/app.py
QWEN38_SIDE_SLOT0_WAIT_S	3	server/app.py
QWEN38_SIDE_STASH_GB	1.0	server/app.py
QWEN38_SKINNY_LDW	0	tools/nvfp4_skinny.py
QWEN38_SKINNY_PDL	0	tools/nvfp4_skinny.py
QWEN38_SKINNY_SRUN	0	1 • tools/nvfp4_skinny.py
QWEN38_SKINNY_SSTUB	0	tools/nvfp4_skinny.py
QWEN38_SKINNY_TILES	none	(a path) • tools/nvfp4_skinny.py
QWEN38_SKINNY_TILES_B	none	tools/nvfp4_skinny.py
QWEN38_SKINNY_TILES_C	none	tools/nvfp4_skinny.py
QWEN38_SKINNY_TILES_WIDE	none	(a path) • tools/nvfp4_skinny.py
QWEN38_SKINNY_TILES_WIDE_B	none	tools/nvfp4_skinny.py
QWEN38_SKINNY_XSTUB	0	tools/nvfp4_skinny.py
QWEN38_SPILL_PARK	on	engine/cache.py
QWEN38_STAIR_MAX_NODES	none	engine/lenrouter.py
QWEN38_STAIR_RHO	none	(a path) • engine/lenrouter.py
QWEN38_STAIR_SKIP	0	1 • engine/lenrouter.py
QWEN38_STAIR_TABLES	none	(a path) • engine/lenrouter.py
QWEN38_STATUS_CLIENTS	open-webui	server/conductor/states.py
QWEN38_SUFFIX_STORE	~/.qwen38-spark-engine/suffix	server/app.py
QWEN38_TANDEM_SCRATCH	none	server/conductor/packages.py
QWEN38_TANDEM_URL	none	server/conductor/packages.py
QWEN38_TOOL_FIT	on	server/conductor/router.py
QWEN38_TREE_ALIAS_STATE	0	0 • engine/model.py
QWEN38_TREE_CHAIN_DELEGATE	1	engine/model.py
QWEN38_TREE_HOST_DEPTH	0	1 • engine/model.py
QWEN38_TREE_MS	none	8:77.9,16:78.9,24:87.0,32:97.0 • engine/router.py
QWEN38_TREE_NODES	0	24 • engine/router.py
QWEN38_TREE_NODES_NARROW	0	16 • engine/router.py
QWEN38_TREE_WIDE_AFTER	0	32 • engine/lenrouter.py
QWEN38_TWO_STREAM	0	engine/model.py
QWEN38_UT_INVERSE	1	engine/gdn.py
QWEN38_VERIFY_GRAPH	0	1 • engine/model.py
QWEN38_VERIFY_ROWS	16	32 • engine/model.py
QWEN38_VERIFY_TILES	1	tools/nvfp4_verify_tiles.py
QWEN38_WARM_PREFILL_TOKENS	2600	server/app.py
QWEN38_WY_BLOCK	8	tools/gdn_wy_kernels.py
QWEN38_WY_BV	32	tools/gdn_wy_kernels.py
QWEN38_WY_DBG	0	tools/gdn_wy_kernels.py
QWEN38_WY_PREC	ieee	tools/gdn_wy_kernels.py
QWEN38_WY_WARPS	4	tools/gdn_wy_kernels.py
QWEN38_WY_WARPS_PREP	4	tools/gdn_wy_kernels.py

29.3 Server variables

Variable	Read in
QSE_ADMIN_TOKEN	ops/make-secrets.sh, ops/serve.env, ops/start.sh and 7 more
QSE_ANSWER_RESERVE	server/app.py
QSE_ASSETS	ops/profiles/_flashnext.env, ops/profiles/_lfm25.env, ops/use-profile.sh and 1 more
QSE_BACKENDS	ops/serve.env, ops/start.sh, server/app.py and 1 more
QSE_BARRIER_IDLE_OWNER	ops/serve.env, server/slots.py
QSE_CHAT_DATA	ops/chatd.sh, ops/sandbox/uninstall.sh, server/chat_forward.py
QSE_CHAT_FORWARD_KEY	ops/chatd.sh, ops/make-secrets.sh, server/chat_forward.py
QSE_CHAT_URL	server/chat_forward.py
QSE_COMMUNITY_DIR	server/setup/community.py
QSE_COMMUNITY_ENV_DIR	ops/serve.env, ops/use-profile.sh
QSE_COMMUNITY_URL	ops/serve.env, server/setup/community.py
QSE_COMPACTION	server/app.py
QSE_COMPACTION_IDLE_S	server/app.py
QSE_COMPACTION_MIN_TOKENS	server/app.py
QSE_COMPACTION_PREPARE	server/app.py
QSE_CONDUCTOR_CONFIG	ops/serve.env, ops/start.sh, ops/use-profile.sh and 3 more
QSE_CONDUCTOR_ROOT	ops/roles.sh, server/conductor/residency.py
QSE_DASHBOARD_LOGIN	ops/serve.env, ops/start.sh, proxy/__main__.py and 2 more
QSE_DISK_ADOPT_FROM	ops/serve.env, server/diskcache.py
QSE_DISK_CACHE	ops/profiles/_exl3.env, ops/profiles/_flashnext.env, ops/profiles/_lfm25.env and 6 more
QSE_ENGINE_FLAGS	ops/profiles/_flashnext.env, ops/profiles/_lfm25.env, ops/serve.env and 2 more
QSE_ENGINE_PORT	ops/engine_status.py
QSE_EXL3_BPW	ops/profiles/_exl3.env, ops/profiles/exl3-2.5.env, ops/profiles/exl3-3.0.env and 4 more
QSE_FN_BPW	ops/profiles/_flashnext.env, ops/profiles/flashnext-2.05.env, ops/profiles/flashnext-3.05.env and 3 more
QSE_FN_TABLE	ops/profiles/_flashnext.env, ops/profiles/flashnext-2.05.env, ops/profiles/flashnext-3.05.env and 3 more
QSE_FN_VISION_FILE	ops/profiles/_flashnext.env, ops/profiles/flashnext-4.05.env
QSE_IMAGE_BACKEND	server/app.py
QSE_KV_HARD_FLOOR_GB	ops/profiles/_flashnext.env, ops/serve.env, server/app.py
QSE_KV_MARGIN_ROWS	server/app.py
QSE_KV_POOL	engine/kvpool.py, ops/profiles/_flashnext.env, ops/profiles/_lfm25.env and 3 more
QSE_KV_POOL_MIN_ROWS	engine/kvpool.py
QSE_LFM_VARIANT	ops/profiles/_lfm25.env, ops/profiles/lfm25-bf16.env, ops/profiles/lfm25-exl3-2.5.env and 5 more
QSE_LIVE_DECODE_WARMUP_TOKENS	ops/serve.env, server/live.py
QSE_LIVE_HZ	ops/serve.env, ops/start.sh, server/app.py and 1 more
QSE_LIVE_MOSTLY_CACHED_PCT	ops/serve.env, server/live.py
QSE_LIVE_MOSTLY_CACHED_TOKENS	ops/serve.env, server/live.py
QSE_LIVE_NOW_WINDOW_S	ops/serve.env, server/live.py
QSE_LIVE_PREFILL_WARMUP_CHUNKS	ops/serve.env, server/live.py
QSE_MEM_NEED_GIB	ops/profiles/flashnext-2.05.env, ops/profiles/flashnext-3.05.env, ops/profiles/flashnext-4.05.env and 10 more
QSE_METRICS_TOKEN	ops/make-secrets.sh, ops/start.sh, proxy/app.py and 1 more
QSE_NO_ROLLBACK	ops/use-profile.sh, server/setup/apply.py
QSE_OPS_DIR	ops/start.sh, server/setup/apply.py, server/setup/device.py
QSE_PAGE_CACHE_DROP	ops/profiles/_flashnext.env, ops/profiles/_lfm25.env, ops/serve.env
QSE_PAGE_CACHE_KEEP	ops/profiles/_flashnext.env, ops/serve.env
QSE_PARALLEL	ops/profiles/_flashnext.env, ops/profiles/_lfm25.env, ops/serve.env and 2 more
QSE_PARALLEL_BATCH	server/app.py
QSE_PARALLEL_CUT	server/app.py
QSE_PARALLEL_DRAFT	server/app.py
QSE_PARALLEL_FLOOR_GB	engine/kvpool.py, ops/profiles/_flashnext.env, ops/serve.env and 1 more
QSE_PARALLEL_LAYER_YIELD	server/app.py
QSE_PARALLEL_SWITCH_MS	server/app.py
QSE_PARALLEL_TIMING	server/app.py
QSE_PARALLEL_WINDOW_ROWS	server/app.py
QSE_PRICE_TABLE	server/app.py
QSE_PROFILE_NAME	ops/serve.env
QSE_PROFILE_OVERRIDE	ops/serve.env, ops/use-profile.sh, server/setup/apply.py and 1 more
QSE_ROLES_DYNAMIC	ops/serve.env, server/rolemem.py
QSE_ROLE_MAX_HOLD_S	ops/serve.env, server/rolemem.py
QSE_ROLE_RELOAD_MARGIN_GB	ops/serve.env, server/rolemem.py
QSE_ROLE_RELOAD_S	ops/serve.env, server/rolemem.py
QSE_ROLE_THRASH_S	ops/serve.env, server/rolemem.py
QSE_ROLE_TICK_FAKE	server/app.py
QSE_ROLE_TICK_S	server/rolemem.py
QSE_ROLE_UNLOAD_LAYA	ops/serve.env, server/app.py, server/rolemem.py
QSE_SANDBOX_SOCKET	ops/sandbox/install.sh
QSE_SECRETS	install.sh, ops/start.sh, server/setup/apply.py
QSE_SESSION_S	server/auth.py
QSE_SETUP_ENV	ops/serve.env, server/setup/apply.py
QSE_SETUP_ENV_FILE	ops/serve.env
QSE_SETUP_MODE	ops/start.sh
QSE_SIDE_GRAPHS	ops/serve.env, server/app.py
QSE_SIDE_MAX_LEN	ops/profiles/_flashnext.env, ops/serve.env, server/app.py
QSE_SIDE_RESIDENT_GB	server/app.py
QSE_SIDE_SLOT_CONST_GB	ops/serve.env, server/app.py
QSE_STATE_DIR	ops/chatd.sh, ops/make-secrets.sh
QSE_STATUS_FILE	ops/engine_status.py
QSE_STATUS_KEEP	ops/hold.sh, ops/stop.sh
QSE_STATUS_PORT	ops/engine_status.py, ops/hold.sh, ops/serve.env and 1 more
QSE_TEST	proxy/app.py, server/app.py, server/ledger.py
QSE_USAGE_DEFAULT	server/app.py
QSE_USAGE_LEDGER	install.sh, ops/serve.env, ops/start.sh and 3 more
QSE_WEIGHTS	engine/models/flashnext/serving.py, engine/models/lfm2_moe/serving.py, ops/engine_status.py and 9 more
QSE_WEIGHTS_ALSO	ops/profiles/_flashnext.env, ops/use-profile.sh
QSE_WEIGHTS_DIR	ops/profiles/_exl3.env, ops/profiles/_flashnext.env, ops/profiles/_lfm25.env and 5 more
QSE_WEIGHTS_FILES	ops/profiles/_exl3.env, ops/profiles/_flashnext.env, ops/profiles/_lfm25.env and 4 more
QSE_WEIGHTS_GB	ops/profiles/exl3-2.5.env, ops/profiles/exl3-3.0.env, ops/profiles/exl3-3.5.env and 18 more
QSE_WEIGHTS_HF	ops/profiles/_exl3.env, ops/profiles/_flashnext.env, ops/profiles/lfm25-bf16.env and 10 more
QSE_WEIGHTS_REVISION	ops/profiles/_flashnext.env, ops/profiles/flashnext-2.05.env, ops/profiles/flashnext-3.05.env and 13 more
QSE_WREN_MODEL	server/setup/wren.py

The most used: QSE_CONDUCTOR_CONFIG, QSE_USAGE_LEDGER, QSE_DISK_CACHE and QSE_BACKENDS name the conductor’s directory, the ledger, the disk tier’s directory and the proxy’s config database; ops/start.sh passes them to the service only, so a benchmark server started with serve.env exported never writes them. QSE_DASHBOARD_LOGIN turns the dashboard login on. QSE_ADMIN_TOKEN and QSE_METRICS_TOKEN live in secrets.env, which ops/make-secrets.sh writes.

29.4 Proxy, collector and plugin variables

Variable	Example	What it sets
TANDEM_BACKENDS	deepseek,vllm1	the backends the environment declares
TANDEM_BACKEND_<ID>_KIND	deepseek	generic | tandemllm | vllm | sglang | openai | deepseek
TANDEM_BACKEND_<ID>_URL	https://api.deepseek.com/v1
TANDEM_BACKEND_<ID>_KEY	(a key)	held in memory, never stored
TANDEM_BACKEND_<ID>_LABEL	DeepSeek
TANDEM_BACKEND_<ID>_MODELS	deepseek-flash,deepseek-v4-pro
TANDEM_BACKEND_<ID>_MODELS_MODE	discover	discover (GET /models, plus MODELS) | list
TANDEM_BACKEND_<ID>_FALLBACK	openrouter	backends tried after this one, in order
TANDEM_BACKEND_<ID>_MAX_INFLIGHT	none	requests at once, past it a fallback or a 503; none (the default): no limit from the proxy, the backend takes what it accepts
TANDEM_BACKEND_<ID>_ENABLED	1
TANDEM_BACKEND_<ID>_INJECT_USAGE	1	ask a stream for its usage frame
TANDEM_BACKEND_<ID>_HARNESS	off	off | shadow | on
TANDEM_BACKEND_<ID>_HARNESS_PROFILE	coding	the conductor profile of its requests
TANDEM_BACKEND_<ID>_WRITER	auto	the background memory writer on this backend (proxy/memwriter.py): auto (on, off for a paid cloud model) | off | shadow | on
TANDEM_BACKEND_<ID>_WRITER_USD_DAY	0.25	its spend cap a day on a paid model
TANDEM_BACKEND_<ID>_WRITER_RUNS_DAY	48	its runs a day
TANDEM_BACKEND_<ID>_WRITER_PROMPT_TOKENS	32000	its prompt at most (older messages dropped)
TANDEM_BACKEND_<ID>_TIMEOUT_CONNECT	10	seconds; also _TIMEOUT_FIRST_BYTE, _TIMEOUT_IDLE
TANDEM_BACKEND_<ID>_HOST	auto	where the System view reads the backend’s machine: auto (this machine when the URL is), same, collector, exporter, off
TANDEM_BACKEND_<ID>_HOST_URL	(a URL)	the remote collector (python -m proxy.hostcollect), or a node_exporter page
TANDEM_BACKEND_<ID>_HOST_DCGM_URL	(a URL)	a DCGM exporter page (mode exporter)
TANDEM_COLLECTOR_TOKEN	...	the remote collectors’ bearer token, in memory only (proxy/hostcollect.py)
TANDEM_ROUTE_<n>	gpt-*=openai	routing rules, ordered by n
TANDEM_PRICE_<n>	deepseek/deepseek-flash=0.30,0.006,1.2	USD per million: input, cached input, output, and optionally cache write
TANDEM_PRICE_<n>_OFFPEAK	0.15,0.003,0.6	the same model outside the provider’s peak hours (DeepSeek)
TANDEM_API_KEYS	(a key)	the proxy’s own keys, in memory only
TANDEM_SECRETS_KEY	<base64	of 32 bytes> the master key (proxy/secrets.py)
TANDEM_ADMIN_TOKEN		read in ops/dash.env.example, proxy/__main__.py
TANDEM_BACKEND_DEEPSEEK_KEY		read in ops/dash.env.example
TANDEM_BACKEND_DEEPSEEK_KIND		read in ops/dash.env.example
TANDEM_BACKEND_DEEPSEEK_LABEL		read in ops/dash.env.example
TANDEM_BACKEND_VLLM1_FALLBACK		read in ops/dash.env.example
TANDEM_BACKEND_VLLM1_HOST		read in ops/dash.env.example
TANDEM_BACKEND_VLLM1_HOST_DCGM_URL		read in ops/dash.env.example
TANDEM_BACKEND_VLLM1_HOST_URL		read in ops/dash.env.example
TANDEM_BACKEND_VLLM1_KIND		read in ops/dash.env.example
TANDEM_BACKEND_VLLM1_MAX_INFLIGHT		read in ops/dash.env.example
TANDEM_BACKEND_VLLM1_TIMEOUT_FIRST_BYTE		read in ops/dash.env.example
TANDEM_BACKEND_VLLM1_URL		read in ops/dash.env.example
TANDEM_COLLECT		read in proxy/launch.py
TANDEM_COLLECT_BASE		read in proxy/launch.py
TANDEM_COLLECT_TOKEN		read in plugins/sglang/tandem_sglang_plugin/_push.py, plugins/vllm/tandem_vllm_plugin/__init__.py
TANDEM_COLLECT_URL		read in plugins/sglang/tandem_sglang_plugin/_push.py, plugins/vllm/tandem_vllm_plugin/__init__.py
TANDEM_CONDUCTOR		read in proxy/__main__.py
TANDEM_CONFIG_DB		read in ops/dash.env.example, proxy/__main__.py
TANDEM_DASHBOARD_LOGIN		read in ops/dash.env.example, proxy/__main__.py
TANDEM_DIR		read in install.sh
TANDEM_HEALTH_TIMEOUT		read in install.sh
TANDEM_HOST		read in install.sh, ops/dash.env.example
TANDEM_MIN_FREE_GIB		read in install.sh
TANDEM_MODEL_NAME		read in install.sh
TANDEM_NOTIFY_URL		read in server/conductor/monitor/notify.py, server/conductor/monitor/sessiontasks.py
TANDEM_PORT		read in install.sh, ops/dash.env.example
TANDEM_PRICE_1		read in ops/dash.env.example
TANDEM_PRICE_1_OFFPEAK		read in ops/dash.env.example
TANDEM_REF		read in install.sh
TANDEM_REPO		read in install.sh
TANDEM_ROUTE_1		read in ops/dash.env.example
TANDEM_SCRATCH		read in server/conductor/packages.py
TANDEM_SKIP_MEMCHECK		read in install.sh
TANDEM_URL		read in server/conductor/packages.py, server/laya_kit/laya_drive.py
TANDEM_USAGE_LEDGER		read in ops/dash.env.example, proxy/__main__.py

29.5 The conductor’s documents

document	what it holds	edited on
profiles	one document per profile: mode, decisions, task modes and caps, thinking, ui, outcomes, library, memory, shadow	Profiles
router	the classes and their policy, the classifier stages and thresholds, the structural rules, side shapes, outcomes, library	Routing
packages	every package: source, switch, limits, flags	Packages
roles	every role’s contract	Profiles, Roles
settings	the default profile, transcripts, the disk tier, memory notes and the writer	Memory, Profiles
deciders	the decision models, their questions, Laya for scripts	Deciders
residency	the floor, the margin, the settle time, the load order, the claims	Profiles, Residency
monitor	the session monitor’s policy (version 0 is ops/supervisor.yaml)	Monitor
harness	the harness’s settings per backend, behind the proxy	Backends

A document is JSON, checked against the conductor’s schema (GET /v1/conductor/meta serves it, and docs/contract/conductor-v1/ holds a copy). A version stored by an older release loads with defaults for every field it lacks, and the journal says so once.