11. dasLLAMA environment variables

GENERATED from the [EnvConfig] declarations in dasllama/dasllama_env.das - do not edit by hand. Regenerate with daslang modules/dasLLAMA/harness/gen_env_doc.das; a knob read anywhere in the tree without a declaration fails tests/test_env_registry.das.

Types: flag is unset-means-default, 0/false/off/no (any case) is false and anything else true; number falls back to the default when unset or unparseable, with a logged warning on garbage; text and path are taken verbatim. A SET-BUT-EMPTY variable counts as unset everywhere. Every knob loads ONCE, at context init, into the g_env_* globals - hot code reads struct fields, and set_env_variable after startup is invisible (arm a child process’s environment instead).

11.1. Engine

Read by the inference engine itself, so these affect any program that loads a model.

Variable	Type	Default	Effect
DASLLAMA_IMAGE	flag	on	Use the prepared-image file beside a source model for millisecond reloads. 0 skips the file: LLM ggufs take the full parse; audio carriers mint the image in memory and serve that (they are always image-backed).
DASLLAMA_IMAGE_SAVE	flag	on	Persist the prepared image beside the source; 0 builds it in memory and serves that, leaving no .dlim behind.
DASLLAMA_CPU_PREFILL	flag	off	Allow the CPU prefill path even when a GPU prefill override is registered.
DASLLAMA_PIN_PREFILL	text	unset	Pin prefill to one registered override by name (e.g. metal, vulkan); anything else declines before upload.
DASLLAMA_MTP_HEAD	path	unset	Split NextN/MTP head GGUF to attach to the trunk being loaded (the split mtp-.gguf form: its blk.<n_layers>.* tensors become the draft block; its copies of the trunk’s tables are ignored). Default: the mtp- sibling beside the trunk when present. The head rides the split-shard walk, so one prepared image carries trunk and head.
DASLLAMA_ROPE_SCALING	text	unset	RoPE scaling override for the load: yarn
DASLLAMA_ROPE_SCALE	number	0	The RoPE scaling factor s (the context multiplier) for the override; 0 = the file’s rope.scaling.factor, and yarn needs one.
DASLLAMA_YARN_ORIG_CTX	number	0	YaRN: the original training context the factor extends; 0 = the file’s original_context_length, else its context_length (llama.cpp –yarn-orig-ctx).
DASLLAMA_YARN_BETA_FAST	number	0	YaRN: the low correction dim beta_fast; 0 = the file’s, else 32.
DASLLAMA_YARN_BETA_SLOW	number	0	YaRN: the high correction dim beta_slow; 0 = the file’s, else 1.
DASLLAMA_PIN_BACKEND	text	unset	Pin the matmul backend by name, bypassing the measured auto-selection.
DASLLAMA_RARE_ONEPASS	flag	on	The streamed split’s host share runs each (expert, row) pair’s whole chain on one lane in one fork a batch - gate, up, act, requant, down, park; 0 = the three region-list GEMV passes with the act and requant between them (the A/B arm).
DASLLAMA_RARE_BATCHED	flag	on	The one-pass rare rail runs an expert’s rows through the batched expert cores (the plane read once a tile); 0 = the rows core a row (the A/B arm).
DASLLAMA_RARE_ARENA	number	1	The rare rail reads a streamed layer’s experts off the tier’s pinned device-layout planes through the format’s row-major kernel - the prompt’s host share and a token’s host picks alike. =1 (the default) takes the rail where the load kept no planar planes (a trimmed lane: one copy of the experts); where the planar planes are loaded (the CUDA home, a lane minted with DASLLAMA_TRIM=0, DASLLAMA_IMAGE=0) the tuned planar kernels serve them. =2 takes the rail on the planar planes too (the A/B arm of the row-major kernels, the environment’s twin of set_rare_arena(1)); =0 keeps every format on the planar planes.
DASLLAMA_MOE_HOST_FIRST	number	0	The meet hands the host this many of a streamed layer’s rarest experts before the device claims any (0 = none: the two sides claim at their own pace) - a measurement lever, so a carrier the device would take whole still runs the host rail.
DASLLAMA_IQ3S_SERVE	text	auto	How an IQ3_S weight is served: grid = the disk form (the 3.44-bit grid planes, decoded by a gather), lut = the iq3s4 form (a 4-bit codebook index a superblock, decoded by the LUT kernels - exact, 17% more bytes, the CPU decode several times faster), auto = lut on the CPU rails and the Vulkan tier, grid under the Metal tier (every home serves both forms at every site).
DASLLAMA_PIN_BATCH_BACKEND	text	unset	Pin the batched (prefill) matmul backend independently of the decode one.
DASLLAMA_EXPERT_REUSE	flag	off	Arm the MoE expert-reuse counter (probes and servers; benches use set_expert_reuse instead).
DASLLAMA_NOISY	flag	off	Print engine diagnostics (tier selection, upload plan, arm/decline reasons).
DASLLAMA_DEVICE_POOL	flag	on	A paged scheduler serving device-home streams keeps their openings in its prefix cache too - a recurrent model’s checkpoints and the finished turns’ rows copied down to host RAM under the cache’s budget, taken back up into a region when a request opens with one the device no longer holds; 0 keeps the regions’ parked claims and checkpoint rows alone - the A/B rail.
DASLLAMA_CALLER_PRIO	number	2 (the top notch)	A/B rail for the dispatch caller’s thread priority, -2..2; unset claims the top notch.
DASLLAMA_ALLOW_BARE_JOBQUE	flag	off	Let the engine dispatch on a job queue nobody configured with setup_dasllama_jobque: a loud warning instead of the panic. A bare queue decodes ~200x slower, so only a run measuring that regime wants it.
DASLLAMA_SINGLE_THREAD	flag	off	Run every kernel on the calling thread: no dispatch to the job queue at all, whatever queue exists (set_single_thread is the runtime spelling). The browser build’s baseline and the figure that says what the worker pool buys.
DASLLAMA_TRUTH_REFRESH	flag	off	Regenerate the stored parity truth files instead of comparing against them.
DASLLAMA_CONV_PROF	flag	off	Bucket gguf -> image conversion time by kind over the weight walk; one clock pair per tensor.
DASLLAMA_ALLOW_INTERP_LOAD	flag	off	Permit a big gguf load without -jit; the transforms run interpreted, so expect minutes per GB.
DASLLAMA_PREFETCH	flag	on	Advisory source-mapping readahead at gguf load (cold-conversion fix); =0 restores on-demand faulting.
DASLLAMA_PREFETCH_IMAGE	flag	off	Readahead of the prepared image at its map (madvise WILLNEED, synchronous on macOS): a purged page cache then reads the whole image at the SSD’s rate before the first token - 103 GB in 15 s on the M5 - instead of paying 2.5x on the routed expert stacks over the first 128 tokens, and a WARM map pays ~2.5 s for the page walk in place of 65 ms. A serving process that runs for hours wants it; a bench, a test or a short session does not.
DASLLAMA_VAD_STREAM_GPU	flag	off	The streaming silero VAD (vad_iter_feed, the live-mic callers) runs each feed’s whole chunks through the GPU driver’s VAD seat (Metal, or Vulkan where it serves) instead of the CPU chunk step; off, the measured-faster form for one 32 ms chunk. The batch VAD (vad_probs, find_speech_timestamps) takes the seat whatever this says; set_vad_stream_gpu is the in-process twin.
DASLLAMA_GPU	flag	off	One switch for the measured-best GPU rail set; any DASLLAMA_GPU_* knob still overrides individually. On a Vulkan build it is also the want the tower driver reads before any device init: the vision and audio towers’ block loops serve on the device only while it is on, and decline device otherwise.
DASLLAMA_CUDA	flag	off	The environment’s spelling of the GPU home (the program’s is set_gpu_home, dasllama-server’s its `–gpu vulkan
DASLLAMA_GPU_MOE_LAYERS	number	-1 (auto)	How many MoE expert layers to hold resident on the GPU; -1 lets the upload walk place the split.
DASLLAMA_GPU_MOE_STREAM	number	-1 (auto)	How many MoE layers to stream rather than hold resident; -1 is auto.
DASLLAMA_GPU_VRAM_MB	number	probed	Override the detected VRAM budget in MiB that sizes the resident expert stacks; a pinned cap takes no headroom and reads no OS room.
DASLLAMA_GPU_MIN_CTX	number	built-in floor	Lower the context floor for arming the resident decode driver, for a short-context session on a small card.
DASLLAMA_GPU_CTX_MAX	number	the model’s context	Cap the resident driver’s KV mirror at this many positions (the card’s room and the 4 GiB binding cap on the widest layer’s rows still bound it); a cap under the arming floor lowers the floor to the cap (never under 256), since the cap is the shape the caller serves; 0 follows the model’s context; a server keeps the model’s.
DASLLAMA_GPU_CTX_STRICT	flag	off	The context pin (DASLLAMA_GPU_CTX_MAX, set_gpu_ctx_max) is the context asked, not a cap: a K/V mirror the card cannot hold at it declines the resident driver instead of shortening, and the per-op rails serve the asked context; the load log names the room.
DASLLAMA_GPU_PF_ROWS	number	0	The whole-model driver’s prompt window in rows: the prefill panels’ size, and the rows a hyper-connection model’s host expert step and the hot pool’s window chain take at once; a longer prompt runs as sequential windows, each paying every routed layer’s expert pass again. 0 (the default) lets the driver pick: 4096 where its plan fits with the wider window’s scratch, else 2048. Rounded down to a multiple of 128, never under 128; the plan charges the panels at this count, so a wider window shortens the K/V mirror or the pool.
DASLLAMA_GPU_PF_STREAM	flag	on	Stream each routed layer’s expert stack to the device for the prompt, where the routed experts serve on the host (a hyper-connection model, or a plain MoE’s host layers) - on the Vulkan tier over its transfer queue, on the CUDA home over its copy stream: pinned device-layout mirrors of every routed layer at load, a device slot (a layer’s stack, or under DASLLAMA_GPU_MOE_SPLIT the per-expert chain’s ring of experts), the copies landing while the chain computes, the host summing what the meet hands it. 0 keeps the host’s sums and the hot pool’s window chain.
DASLLAMA_GPU_PF_STREAM_LAYERS	number	0 (every host layer)	Under DASLLAMA_GPU_PF_STREAM, mirror only the first this many host layers’ expert stacks; the rest keep the host’s sums. A parity cell’s lever: a few layers’ mirrors beside a planar load, where every layer’s would not fit the box.
DASLLAMA_GPU_HOST_LAYERS	number	0 (the plan’s own count)	Serve at least this many routed layers’ experts on the host - the first in layer order - on the whole-model driver of any MoE that routes no more picks a row than the split command serves, the gemma-4 MoE with its parallel dense expert included, whatever the card holds (set_gpu_host_layers in process): the host-layers form, with its split token command and cut window chain, reached on any card. A hyper-connection model serves every routed layer on the host regardless.
DASLLAMA_GPU_KV	text	unset	The K/V codec the whole-model driver’s mirror is asked to arm at, where no caller set one in process (set_gpu_kv_dtype): f16 (the default), q8_0 (half the mirror’s bytes) or tq4 (a quarter). Sessions of the armed codec are served; any other spelling reads as f16.
DASLLAMA_GPU_RESIDENT	flag	on	The whole-model resident driver for a model that fits the card, MoE included; 0 keeps the per-op rails (the A/B lever).
DASLLAMA_GPU_KV_RING	flag	on	A sliding-window layer’s K/V mirror rows on a GPU home hold its window plus the rows one command writes, a position’s row at its slot of that ring; 0 keeps every position’s row (the flat slab, the A/B lever). A ringed history serves a continuation that starts no more than that slack under the rows it holds; one further back prefills again.
DASLLAMA_GPU_DN	flag	follows DASLLAMA_GPU	DeltaNet rail on the GPU.
DASLLAMA_GPU_DND	flag	follows DASLLAMA_GPU	DeltaNet decode rail on the GPU.
DASLLAMA_GPU_ATTN	flag	follows DASLLAMA_GPU	Attention rail on the GPU.
DASLLAMA_GPU_DENSE	flag	follows DASLLAMA_GPU	Dense-weight rail on the GPU.
DASLLAMA_GPU_DENSE_ATTN	flag	on	The dense rail’s attention half (per-layer q/o pairs); opt-OUT, measured a loss on gemma-4.
DASLLAMA_GPU_DENSE_SHEXP	flag	off	The dense rail’s gemma-4 shared-expert w1/w3/w2 triple; opt-IN, measured a wash-to-loss.
DASLLAMA_GPU_DENSE_FFN	flag	on	The dense rail’s FFN half (each dense layer’s w1/w3/w2 triple, the decode row’s FFN on the device); 0 keeps a dense layer’s FFN on the CPU (the A/B rail).
DASLLAMA_GPU_SHEXP	flag	follows DASLLAMA_GPU	Shared-expert rail on the GPU.
DASLLAMA_GPU_QKV	flag	off	Fused QKV rail on the GPU; stays off under DASLLAMA_GPU (measured a wash).
DASLLAMA_GPU_CLS	flag	on	Keep the classifier plane on the GPU - the best ms-per-GB region; 0 is the CPU A/B lever.
DASLLAMA_GPU_COMBINE	flag	on	Device-side routed MoE combine; 0 falls back to the host combine.
DASLLAMA_GPU_MOE_SPLIT	flag	on	Split each STREAMED expert layer’s prefill between the GPU (the experts it can stream in the CPU’s time) and the CPU (the rest, concurrently); 0 streams whole layers.
DASLLAMA_GPU_DEC_MEET	flag	off	A decode token’s picks on a streamed layer take the prefill’s route: the meet claims them as the GPU’s per-expert DMAs from the busy end and the CPU’s rail from the rare end, in place of the one-row host step. The experiment knob; the window floors lift under it.
DASLLAMA_GPU_HOT_DMA	flag	on	The hot expert pool fills by DMA: a token’s miss is copied off its pinned stack into the pool’s coldest unpicked slot on the home’s copy queue (the Vulkan tier’s transfer queue, the CUDA home’s copy stream), every miss a step and no host time, so the next token finds it; and a prompt’s streamed window serves the pool’s hits off their slots first and lands its busiest misses in the pool’s free and coldest slots, so the decode behind the prompt starts warm. 0 keeps the host’s gather-and-upload fill under its step budget and the window off the pool. The A/B knob.
DASLLAMA_GPU_HOT_SHARE	number	0	The share of a decode step’s pool misses the DMA fill admits past the heat gate - the warmest first, each into the coldest unpicked slot, landing for the next step while this step’s host sums them - so the pool’s working set follows the stream instead of the gate’s hot few. 0 keeps the heat gate alone; 1 admits every miss the tier can copy. Needs DASLLAMA_GPU_HOT_DMA.
DASLLAMA_GPU_DEC_TAIL	flag	on	Decode FFN tail: a resident expert layer’s FFN carries the combine, the residual and the next layer’s attention feed on the device, so the two submits share one fence wait; 0 collects every FFN on the host.
DASLLAMA_GPU_DEC_SPAN	flag	on	Whole-token decode span: every resident expert layer’s attention, router, top-k, FFN and combine as ONE recorded chain and one submit per token; 0 runs the per-layer submits.
DASLLAMA_GPU_HEAT	number	0	Hot expert slots a layer. On the per-op tier: hold the N hottest experts of a streamed layer resident (unset = none). On the whole-model driver with the routed experts on the host (a hyper-connection MoE): the slots of the hot expert pool - unset = every slot that fits beside the weights and the mirror, else up to 32 out of the mirror’s context; a count asks that many; 0 keeps every routed expert on the host.
DASLLAMA_GPU_PROF	flag	off	Report lifetime GPU queue submissions (real commands plus staging round-trips), and trace every routed row the host sums - the split token command’s host segment and the CPU forward alike - as one moe_k line (the layer, the session, the first pick, the L1 of the feed row and of the routed sum).
DASLLAMA_MOE_TRACE	text	unset	Record the routed picks of every layer’s host expert step under the whole-model driver - a layer, its rows, k expert ids and weights a row, in serving order - and write them to this file when the model drops: the hot expert pool’s policy experiments replay the trace offline (harness/hot_pool_sim.das). Unset: no trace.

11.2. Vision

The image preprocessing rail: dynamic-resolution token bounds and the debug dump.

Variable	Type	Default	Effect
DASLLAMA_VISION_MIN_TOKENS	number	40 (the gemma-4 model default)	Minimum soft tokens per image; dynamic-resolution geometry upscales to reach it.
DASLLAMA_VISION_MAX_TOKENS	number	280 (the gemma-4 model default)	Maximum soft tokens per image; dynamic-resolution geometry downscales to fit.
DASLLAMA_VISION_DUMP	text	unset	Path prefix: dump every letterboxed image fed to a vision embedder as .ppm - the human smoke test; empty = off.
DASLLAMA_SPAN_FUSE	flag	on	Serve the image turn as ONE fused prefill (per-query span mask); 0 pins the three-eval splice (head / media rows / tail) - the parity A/B rail.

11.3. Metal backend

Apple GPU backend. Absent on non-Apple builds, where setting them does nothing.

Variable	Type	Default	Effect
DASLLAMA_METAL_MSL_DUMP	path	unset	Directory to write each kernel’s generated source as .metal when its pipeline is built - the SPIR-V dump’s twin, for diffing a kernel across a change.
DASLLAMA_METAL_LOGITS	flag	on	Produce logits on the GPU; 0 pulls the classifier back to the CPU. Blob-only models force it on.
DASLLAMA_METAL_ATTN	flag	on	Tiled QK/AV prefill attention (~10x the trio GEMMs); 0 pins the trio.
DASLLAMA_METAL_SPAN	flag	on	Serve the non-causal media span on the prefill driver (AttnArgs.uend); 0 declines it to the CPU arm, which then needs declared CPU intent.
DASLLAMA_METAL_TOWER	flag	on	Serve tower/embedder encodes on the Metal tower driver - the gemma4uv chain, the gemma4v ViT and gemma3v SigLIP block loops (their exact-plane lanes), the whisper-class block loops, the gemma4a Conformer and the canary / parakeet FastConformer block loops (their f32 lanes), the whole StyleTTS2 (kitten / kokoro) synthesis - PL-BERT, the text and duration encoders, the duration head, prosody, the decoder, the harmonic source, the generator and the inverse STFT - on both weight lanes, the Pocket TTS codec and frame loop, the conv frontends (whisper im2col + the qwen3a conv2d, which serves the q8 default), the GigaAM encode and CTC head (all three weight lanes: the f16 halfwords, the q8 slab, the kq lane’s own planes), and the silero VAD seat; 0 loads each family’s CPU image, the lane whole on the CPU.
DASLLAMA_METAL_TOWER_F16	flag	on	Serve a tower’s weighted GEMMs from a halfword twin of its weights on the crowned tensor mul_mm twins: the image’s baked f16 twin plane where the mint bakes one (only for f16-sourced GEMMs, so it holds the file’s own values), and the FastConformer chain’s device twin of parakeet’s f32 blob. Needs a tensor crown - a box without one, or an image without the plane, keeps the f32 route; 0 pins the f32 GEMM route - the A/B rail.
DASLLAMA_METAL_TOWER_FLASH	flag	on	Serve head-72 tower attention at %64 canvases on the lifted flash kernel (no score slab, no rowstat pass, no head restrides - the gemma3v-geometry ViTs); 0 pins the three-pass slab - the A/B rail.
DASLLAMA_METAL_WDEC	flag	on	Serve the whisper decoder side (cross-KV; the decode step under wdec_step) on the Metal ASR-decoder driver; 0 loads the lane’s CPU image, the whole lane on the CPU.
DASLLAMA_METAL_WDEC_STEP	flag	on	Serve the whisper DECODE STEP on the GPU too (needs wdec); 0 loads the lane’s CPU image, the whole lane on the CPU.
DASLLAMA_METAL_WDEC_SMALL_CPU	flag	on	Bake a whisper decoder narrower than 512 (tiny) into its Metal image with the decode STEP on the CPU - the device step’s per-dispatch latency loses at that width; cross-KV stays on the device; that image is its own file (-cstep), keyed on the box’s CPU configuration. 0 bakes the step onto the device at every width - the A/B rail.
DASLLAMA_METAL_RNNT	flag	on	Serve the transducer greedy decode (parakeet, GigaAM’s RNNT head) on the Metal ASR-decoder driver: the joint’s encoder projection off the encode’s device rows and the whole loop in command buffers of rounds, the tokens read back once; 0 pins the CPU decoder on a CPU image - the A/B rail - while a Metal image’s decode refuses: the image never runs on the CPU.
DASLLAMA_METAL_ATTN_D	flag	on	Fused single-pass decode attention (a head of 64, 96, 128, 256 or 512 on every KV codec; sliding windows, attention sinks, the logit soft cap and a second head class included), on the single row, the speculative verify’s rows and the batch; 0 is the A/B rail to the chunked pair.
DASLLAMA_METAL_ATTN_SINGLE	number	64	Row count below which attention uses the single-chunk kernel; clamped to 128.
DASLLAMA_METAL_MULMM	flag	on	The mul_mm prefill GEMM; 0 falls back to the legacy per-op path.
DASLLAMA_METAL_MM_TAIL	flag	on	GEMV-tail prefill dispatch: npos % 32 in [1,8] peels the last M tile’s real rows onto the fixed-B GEMV family instead of billing a full 32-row tile; 0 pins the padded-tile path (the A/B rail).
DASLLAMA_METAL_KQ_HALF	flag	on	The K-quant rows sites (two to eight rows: the verify, the batch step, the prefill’s short window) read x as the f16 twin the row producer stores beside it, for every format whose half-x route the box crowned (kq_rows_half_); 0 keeps the f32 forms under every crown (the A/B rail).
DASLLAMA_METAL_TENSOR	flag	on	Apply the box profile’s tensor-lane crowns (its runtime.metal_tensor set), so every family the box’s race crowned serves its tensor twin; 0 leaves the set empty and every family on its base form past the routed twins’ knee (set_metal_moe_kq_twin_avg) and the dev-W route, which ride the toolchain probe - what a GPU without the tensor lane (the M4 class) serves - the arm that proves a change, and the base kernel set, on that class from an M5.
DASLLAMA_METAL_DEVW	flag	off	The dense Q8_0 and K-quant prefill sites read a stored f16 converted copy of their planes (the dev-W panels the image bakes, or the resident ones); the A/B rail of the default - per-window scratch panels past each site’s row floor, its own bytes under it - which the image lane name carries as -devw.
DASLLAMA_METAL_QK_ROPE	flag	on	Fused prefill qk_norm+rope pass (one panel rewrite, q panel dual-stores its f16 twin); 0 restores the split norm/rope/cvt dispatches (the A/B rail).
DASLLAMA_METAL_GRID1D	flag	off	1-D-grid tall hmm GEMM twin: the threadgroup launch order is forced linear (M-fastest) instead of the driver’s 2-D walk; experiment rail, off by default.
DASLLAMA_METAL_LASTROW	flag	on	Last-layer FFN tail narrowing: the final dense layer’s FFN runs the last row alone (only the classifier reads past the final attention); 0 pins the full-panel tail (the A/B rail). Session keep_hidden opts a forward out.
DASLLAMA_METAL_DBUF	flag	on	Double-buffered W staging on the dense staged prefill GEMMs (q8 and kq stamps; the MoE family measured a loss and stays single-tile): two ping-pong tiles, one barrier per 64-chunk, staging overlaps the tensor op; 0 pins the single-tile forms (the A/B rail).
DASLLAMA_METAL_SPLIT_K	flag	on	Split-K on the short window’s prefill tiles (q8, every dense K-quant, grid and per-32 format, bf16): a 32-row tile grid of at most half the sidecar’s metal_split_k_tgs threadgroups divides its K across up to eight threadgroups a tile, the partial planes summed by one reduce dispatch; 0 pins the whole-K tiles (the A/B rail).
DASLLAMA_METAL_NCB	number	~4 layers/chunk	Command-buffer split: each chunk commits as encoded so the scheduler overlaps chunk k with k-1.
DASLLAMA_METAL_UNRETAINED	flag	off	Skip per-dispatch retain/release on the prefill command buffers.
DASLLAMA_METAL_DEVW_RESIDENT	number	0 (off)	MB budget for RESIDENT dev-W f16 panels under DASLLAMA_METAL_DEVW=1: static weights dequantize ONCE and the panels persist across forwards (vs the per-forward scratch re-dequant). Sites past the budget fall back to scratch.
DASLLAMA_METAL_PF_CAPTURE	flag	on	Record each prefill chunk as a step graph and replay on a concurrent encoder; 0 is serial-encode rollback.
DASLLAMA_METAL_SCHED	number	1	Graph scheduling mode for replay; 0 keeps capture order.
DASLLAMA_METAL_DECODE_CONCURRENT	flag	on	Concurrent encoder for the single-stream decode step, with barriers only at detected hazards.
DASLLAMA_METAL_FUSE	flag	on	Every decode sandwich seam as one residual step (the branch’s post norm, the add, the next norm); 0 restores the split chain for A/B.
DASLLAMA_METAL_SPEC	number	-1 (adaptive)	The single-row Metal decode’s greedy chain (the next step pre-encoded on the GPU’s own argmax): 0 off, 1 forced, -1 adaptive (backs off after a miss). A session sampled at temp > 0 never chains, whatever the mode.
DASLLAMA_METAL_GEMV_TG	number	4	Rows per MoE router/expert GEMV threadgroup, clamped 1..32 (the dense GEMV geometry is fixed).
DASLLAMA_METAL_WAIT_SPIN	flag	on	The single-row decode’s step wait spins on the command buffer’s GPU end time before the blocking wait: the calling thread owns the step and a blocking wake returns late on a box whose cores the jobque lanes fill (M4 Pro: 0.35 ms per token, a step every third step); 0 blocks in waitUntilCompleted alone - the A/B rail.
DASLLAMA_METAL_KQ_B8	flag	on	Single-pass B8 twin for K-quant small-batch mv at B=5..8; 0 is the A/B rail.
DASLLAMA_METAL_KV_MIRROR_MB	number	4096	The device-side KV mirror’s idle budget in MiB: the bytes idle mirrors evict down to once a step is served. A live step’s demand is bounded by the device’s largest buffer alone (unified memory: what fits, serves), and a step past that takes the CPU rails with one logged line.
DASLLAMA_METAL_MIRROR_ADOPT	flag	on	A released paged session’s KV mirror goes to the next session whose leading pages are the same pool groups, its rows not uploaded again; 0 uploads every attached row - the A/B rail.
DASLLAMA_METAL_PF_MIRROR	flag	on	An f16-KV prefill writes its K/V rows into the session’s KV mirror and its attention reads the cached rows there (the device pair, at any row count), so a continuation gathers nothing on the CPU and the first decode step uploads nothing; without the attn_dev crown the attention keeps its f32 panels and the gather, and the whole panel still converts into the mirror, so that first step uploads nothing there either; 0 keeps the panels and leaves the mirror to the decode step’s upload - the A/B rail.
DASLLAMA_METAL_RESIDENCY	flag	on	Pin the served working set (weight regions, planes, the recycled pool buffers) in an MTLResidencySet, macOS 15+: the per-commit residency pass stays a no-op, so the first submission after a CPU-only window stops repaying it (~15 ms/encode on the M1 Max released-rig image cell; 0 is the A/B rail.
DASLLAMA_METAL_HEARTBEAT_S	number	180	Seconds the residency-set heartbeat keeps re-requesting residency after the last served step (a requested set still un-wires over a long CPU-only window; 0 disables the heartbeat - the A/B rail.
DASLLAMA_METAL_HAZARD_PARANOID	flag	off	Barrier at every dispatch instead of at tracker-detected hazards (correctness bisect).
DASLLAMA_METAL_HAZARD_STRICT	flag	off	Treat every detected hazard as strict, widening barriers (correctness bisect).
DASLLAMA_METAL_PIPE_DEBUG	flag	off	Per-step pipeline trace: GPU envelope and true inter-step handoff idle, first steps plus outliers.
DASLLAMA_METAL_DECODE_SKIP	text	unset	Comma-separated dispatch names to skip in decode - a bring-up bisect that breaks correctness.
DASLLAMA_METAL_PREFILL_SKIP	text	unset	Prefill twin of DASLLAMA_METAL_DECODE_SKIP.
DASLLAMA_MTP_DEBUG	text	unset	MTP bring-up bisect: cold logs why the Metal seats left a round to the plain step (the forced reject is set_mtp_force_reject_every).
DASLLAMA_METAL_BATCH_GEMM_MIN	number	5	Batch size at which batched decode switches from GEMV to GEMM; floor 2.
DASLLAMA_METAL_BATCH_SK	flag	on	Split-K in the batched decode GEMM.
DASLLAMA_METAL_BATCH_NCB	number	4	Command buffers per batched decode step, clamped 1..16.
DASLLAMA_METAL_BATCH_CONCURRENT	flag	on	Concurrent encoder for batched decode.
DASLLAMA_METAL_BATCH_FUSE	flag	on	Epilogue fusion in batched decode.
DASLLAMA_METAL_BATCH_MV	flag	on	Fixed-B GEMV forms for small batched decode.
DASLLAMA_METAL_BATCH_MV8	flag	on	The column-tiling MV8 probe for batched decode.
DASLLAMA_METAL_BATCH_PIPE	flag	off	Pipelined batched decode submission.
DASLLAMA_METAL_BATCH_PRE	flag	off	The batched decode pre-encodes the next step under the current one’s GPU run and pokes its token inputs at commit. Off by default: it times the same as building at the call and holds a second step’s buffers.
DASLLAMA_METAL_BATCH_MM	flag	off	The mul_mm rail for batched decode; measured negative at M-pad-32, kept as the A/B rail.

11.4. Vulkan backend

Vulkan GPU backend. Present only where the dasVulkan package is installed.

Variable	Type	Default	Effect
DASLLAMA_COOPMAT	text	auto	Cooperative-matrix mode (auto = cm2 where the device has NV_cooperative_matrix2, else mm, else sdot4; f16 and int8 pick the KHR f16 and int8 batch tiles, which auto never does); the flash-attention twin needs it even when the GEMM runs sdot4.
DASLLAMA_MM_SMALL	text	32	Small-batch tier: 32 = sdot4 (default, beats both coopmat tiles below the crossover), 64 = coopmat M, 128 = always-L.
DASLLAMA_MM_SMALLD	number	64	Small-d cutoff routing narrow roles (k/v) to the small tier; widening measured worse, so this is an instrument.
DASLLAMA_VK_FUSE	flag	on	Fused add+rms+requant on every site of the decode token command and the prefill’s batch pairs, plus the fused qk-norm+rope; 0 pins the split dispatches for a same-build A/B.
DASLLAMA_VK_CHAIN	number	0	The whole-model driver’s greedy chain: the next one-row step submitted behind the running one on the device’s own pick; 0 off, 1 every greedy step, -1 adaptive (backs off after a miss). A sampled session never chains.
DASLLAMA_VK_F16_FFN	flag	on	The resident prefill’s FFN GEMMs (gate, up, down) take the f16 activation feed where the tile family admits it; 0 pins the FFN group to the Q8 activation blocks the CPU chain quantizes, while the attention group keeps its own pick - the A/B rail that takes the feed’s rounding out of a resident-vs-CPU compare (gemma’s logits move with it).
DASLLAMA_VK_XFERQ	flag	on	Stream expert uploads on the dedicated transfer queue, overlapped via a timeline semaphore, and land the resident token command’s logits there behind a compute-timeline wait; 0 keeps the single-queue rail, where the token command copies its own logits under its fence.
DASLLAMA_VK_DECVEC	flag	on	Run the cm2 tiles’ four-wide decode callback where the device has VK_NV_cooperative_matrix_decode_vector; 0 strips it and serves the scalar callback - the same-build A/B and the fallback probe.
DASLLAMA_VK_PLE_F32_TILE	flag	on	The resident prefill window’s per-layer-embedding gate and projection serve a file’s f32 rows on the staged KHR tile (the rows’ words rounded to f16 on the tile’s stage ahead of the KHR multiply, at subgroup 32); 0 keeps them on the scalar f32 tile - the same-build A/B.
DASLLAMA_VK_IMPORT	flag	on	Stream mirrors import the mapped .dlim (VK_EXT_external_memory_host) instead of pinned copies, on the copy arm alone (DASLLAMA_VK_STREAM_FILE_READ=0): the file-read arm, the default, reads into mirrors of its own and never imports; =0 restores the copy path.
DASLLAMA_VK_STREAM_FILE_READ	flag	on	A streamed stack’s planes are read off the .dlim file straight into its pinned memory, the mapping’s expert pages never faulted; =0 copies them out of the mapping (the A/B arm).
DASLLAMA_EXPERT_CACHE_MB	number	0	The expert cache’s pinned budget in MiB for the streamed layers’ expert stacks: the resident slots of every streamed layer and the transient ring together; 0 holds every streamed expert resident (the whole stacks in pinned memory, as a box with the RAM does). Under it each streamed layer keeps a quota of its experts resident and the rest are read off the lane file on demand into the ring, through the format’s row-major kernels on the host and the DMA on the device alike; needs a mapped lane (a chunk-built image has no file to read from).
DASLLAMA_EXPERT_RING	number	0	The expert cache’s transient ring in experts a plane kind (gate, up, down): the slots the on-demand reads land in, freed at a window’s or a token’s end. 0 sizes it to the largest streamed layer’s non-resident count, so one window’s misses always fit.
DASLLAMA_EXPERT_READERS	number	8	The expert cache’s reader threads: the on-demand reads off the lane in flight at once, each thread one read at a time on its own file handle. 0 reads on the calling thread, one at a time.
DASLLAMA_VK_XF_BUNDLE	number	0	The experts one bundle of the streamed expert chain takes at most (0 = a sixteenth of the layer’s experts, 1 to 32). A bundle is the unit the device claims at the meet, at most two in flight, so the cap sets the device’s pace against the host’s share; the A/B lever.
DASLLAMA_TRIM	flag	on	Mint the TRIMMED GPU lane - the armed home’s, vulkan or cuda (the default; =0 mints the lane with its planar families): the device twin plus the packed emb region the CPU embed reads, the planar weight families dropped - a MoE’s expert stacks too, the device holding them or the host reading its streamed layers’ pinned mirrors through the rare rail (so the prompt stream and the rail’s arena arm must be on where routed layers serve on the host). A model the whole-model driver would not take whole keeps its families, and the load says so, as does one with per-layer embeddings, biased expert stacks or the gemma-4 MoE form (the host reads their planar planes); a call the driver passes to the CPU rails panics by name. Folded into the lane’s identity, so a flip re-bakes the lane; set_vulkan_trim is the in-process lever.
DASLLAMA_VK_MEMPRIO	flag	off	Tag every device allocation priority 1.0 (VK_EXT_memory_priority). Off by default: on the NVIDIA WDDM driver the tag makes the tagged weights the pageable set, and a resident model then decodes at PCIe speed once anything is demoted.
DASLLAMA_VK_FA	flag	on	Vulkan flash attention: the decode fa kernel pick AND the cm2 prefill fa tile; 0 falls back to the chunked/scalar paths.
DASLLAMA_VK_TOWER	flag	on	Serve the vision towers’ block loops (gemma4v, gemma3v, qwen3v, qwen25v), the audio towers’ (the whisper-class towers, gemma4a, canary) and the TTS seats (the StyleTTS2 families kitten and kokoro, Pocket) on the Vulkan tower and TTS drivers - the announce is dasLLAMA vulkan tower: <family> serves on the device, once per model when it first serves; 0 loads every tower lane’s CPU image, the whole lane on the CPU.
DASLLAMA_VK_GIGAAM_EXACT	flag	off	GigaAM’s lane policy on the Vulkan tower: 1 prefers the exact lane - the block GEMMs on the f16 GEMM over the carrier’s own halfwords - over the q8 image for an un-pinned load; 0 (the default, until the two lanes are timed on a Vulkan box) keeps the q8 image. Both lanes serve on the device on the f16 GEMM feed alone (off it, every lane loads the CPU image); the family’s lane pin overrides it, set_vulkan_gigaam_exact_lane is the in-process lever.
DASLLAMA_VK_WDEC	flag	on	Serve the whisper decoder side (the per-window cross-KV and the decode step) on the Vulkan ASR-decoder driver - the announce is dasLLAMA vulkan wdec: the whisper decoder serves on the device, once per model; 0 loads the lane’s CPU image, the whole lane on the CPU.
DASLLAMA_VK_RNNT	flag	on	Serve the transducer greedy decode (parakeet, GigaAM’s RNNT head) on the Vulkan ASR-decoder driver: the joint’s encoder projection off the encode’s device rows where the lane’s projection is f32, and the whole loop in command buffers of rounds, the tokens read back once; 0 pins the CPU decoder on a CPU image - the A/B rail - while a Vulkan image’s decode refuses: the image never runs on the CPU.
DASLLAMA_VK_KV_MERGE	flag	on	Merged k
DASLLAMA_VK_FFN_SLICE	flag	on	The last layer’s FFN runs on the window’s last 32 rows only (the classifier reads one); 0 runs it over the whole window for a same-build A/B.
DASLLAMA_VK_OVERLAP	flag	on	Prefill record/execute overlap: the window chain submits in ramped chunks (1,2,4,8 layers) so the GPU starts while the CPU still records; 0 pins the single fenced submit (the per-role GPU profile pins it too, so a chunk gap never bills to a role).
DASLLAMA_VK_CHUNK	number	0	Layers one prefill submit holds at most (the overlap ramp’s cap, 1..8); 0 = 1 on an integrated GPU, 8 on a discrete one. A Linux amdgpu resets the device under a job past its watchdog (about 10 s), which an iGPU’s 8-layer chunk reaches on a long prompt.
DASLLAMA_VK_GPU_EMBED	flag	on	Device-side token-embedding gather for the resident prefill and every token command (the one-row step, the N-row command, the NextN verify and draft): a q8 or K-quant table off its own plane (the tied classifier’s, or an untied table’s where it fits the room past the weights and the K/V), an F32/F16/BF16 table off its exact f32 rows uploaded under a 512 MB cap; 0 keeps the CPU embed loop. set_gpu_resident_embed pins it for the next loads in process.
DASLLAMA_VK_FULLSG	flag	off	Pin REQUIRE_FULL_SUBGROUPS on every class pipeline (instrument; measured slower than plain pipelines on the mm_a gate shape, so those are the default).
DASLLAMA_VK_SUBGROUP	number	32	The subgroup width every class pipeline pins where the device’s own differs and its compute stage may pin it (Vulkan 1.3 subgroupSizeControl; RDNA runs 32 or 64 lanes and defaults to 64): the coopmat tiles, the h128 attention twin and the tower tiles own four 32-lane fragments a workgroup, so a device left at 64 serves the sdot4 tiles and the scalar attention instead. 0 = the device’s own width.
DASLLAMA_VK_REBAR	flag	off	Place the planes in the ReBAR device-local host-visible memory type (one larger than 1 GB present) and write uploads straight into them. Off by default: on the NVIDIA WDDM driver a mapped plane is a host-side shadow the manager parks in system memory under pressure, and a host-experts plan then decodes at PCIe speed (gemma-4 26B Q8_0 8.1 tok/s mapped against 32.7 staged, the 35B-A3B 38 against 71: debug-jit readings, the ReBAR entry in modules/dasLLAMA/PERF_LEDGER.md) - the default stages every upload into unmapped device-local planes.
DASLLAMA_VK_KV32	number	0	Arm the resident driver with f32 KV mirrors instead of the f16 default (A/B instrument; only sessions of the armed codec are served).
DASLLAMA_VK_NROW_BISECT	number	0	benchmarks/lcpp_bench.das --npl N only: drops a class of dispatch from the recorded N-row token command so a profile prices it - bits 1 = no attention, 2 = the one-column GEMVs (row 0 alone), 4 = no q/k norm and no rope, 8 = no activation requant; the logits are garbage under any bit. The bits reach the head and the dense triple; the routed block and the per-layer-embedding branch run their one-row leaves and column forms whole.
DASLLAMA_VK_FUSE_BISECT	number	0	Keeps a fused form of the token command on its split dispatches (a same-build A/B; the logits stay right; the N-row command shares the gate-up and the fused q/k norm through the layer flags read at load) - bits 1 = the wo GEMV’s residual epilogue, 2 = the down GEMV’s, 4 = the fused gate-up twin, 8 = the prologue’s add+rms+requant stamp, 16 = the post-attention one (the wo epilogue too), 32 = the post-FFN one (the down epilogue too), 64 = the fused qk-norm+rope; the load log names each bit as taken or inert on the model
DASLLAMA_CM2_TILE	number	0	cm2 prefill tile pick: 0 = the wave model’s pick (cm2_gemm_pick), 128 = force the m tile, 256 = force the l tile (A/B instrument). Inert on the KHR arm (DASLLAMA_COOPMAT=mm, or a device without NV_coopmat2), whose kq tile has one geometry.
DASLLAMA_CM2_SPLITK	number	0	cm2 split-k: 0 = the pick’s own (the Vulkan wave model cm2_gemm_pick; the CUDA home’s cud_cm2_pick splits a grid short of two blocks an SM until it fills the card and runs a full grid whole), 1 = off, N = force N k-chunks on either home (A/B instrument; shrinks if N strands an empty tail). On the KHR arm the chunks ride its one 128-column geometry.
DASLLAMA_CUDA_MMQ	number	1	The CUDA home’s window chain on the MMQ GEMMs: the int8 one (q8_mmq and the leaves of Q4_K, IQ4_XS, Q4_0, IQ4_NL, Q5_0, Q5_1), the activations requantized to Q8_0 blocks between the GEMMs, as the token step’s feeds are; and NVFP4’s FP4 tensor leaf (nv4_mmq, W4A4: the activations quantized to e2m1 per token); a routed layer’s MXFP4 or NVFP4 expert stacks ride the same rung fed from the expert schedule (mx4_fp4_mmq_moe / nv4_fp4_mmq_moe, the gathered rows as the stacks’ FP4 feed; 0 keeps them on the cm2 s and e tiles). 1 = every plane whose format’s leaf beat the cm2 tiles in the measured cards’ windows, where its shape fits the tile (the output width a multiple of 128, the reduction of 256), runs on that leaf, the other planes on the f16-fed cm2 tiles; 2 = every leaf, the ones the cards split or lost on too (Q5_0, Q5_1) - a new card’s own measurement; 0 = the cm2 tiles everywhere (A/B instrument). The FP4 leaf needs compute capability 120; below it an nv4 plane stays on the tiles under every value. The first window reads the value once and keeps it. The load log names the pick, and the value with the setter that gave it (set_cuda_mmq_pin under a test’s pin).
DASLLAMA_CUDA_PDL	number	1	The CUDA home’s launches under programmatic dependent launch: a kernel is launched while the one before it in the stream still runs, and its entry wait (griddepcontrol.wait, in every emitted kernel on sm_90 and up) holds its reads until that one completes, so the launch latency rides under the predecessor’s tail. 1 = on where the card is sm_90 or newer, 0 = plain stream order (A/B instrument). The device-ready log line names the state.
DASLLAMA_DN_CHUNK	number	1	The window chain’s deltanet scan on the discrete homes: 1 = the chunked form on the 16 x 16 cooperative tiles (dn_scan_chunk_cls: sixteen tokens a chunk, the state advanced once a chunk) wherever the device has them - CUDA, and Vulkan with KHR cooperative matrices at subgroup 32 - the per-token scan elsewhere; 0 = the per-token scan everywhere (A/B instrument). The token step’s fused scan is not this knob’s.
DASLLAMA_CUDA_CHAIN	number	0	The CUDA home’s greedy chain: the next one-row step launched behind the running one on the device’s own pick; 0 off, 1 every greedy step, -1 adaptive (backs off after a miss). A sampled session never chains.
DASLLAMA_VK_SPV_OVERRIDE	path	unset	Directory of .spv files served instead of the emitted words at pipeline creation (offline spirv-opt / hand-patched A/B instrument).
DASLLAMA_VK_SPV_DUMP	path	unset	Directory to write each kernel’s emitted words as .spv at pipeline creation (the override instrument’s capture half).
DASLLAMA_VK_HAZARD_PARANOID	flag	off	Barrier at every dispatch (correctness bisect).
DASLLAMA_VK_HAZARD_TRACE	flag	off	Log every detected hazard and the barrier it produced.

11.5. MoltenVK (macOS)

MoltenVK’s own configuration - listed because the vulkan tier arms it on Apple.

Variable	Type	Default	Effect
MVK_CONFIG_USE_METAL_ARGUMENT_BUFFERS	flag	unset (tier forces 0)	MoltenVK’s own knob. Its argument-buffers default miscompiles the batch tile kernels (stores never land, silently); the tier forces discrete bindings when this is unset.

11.6. Accelerate / AMX backend

Apple Accelerate / AMX float lane. DASLLAMA_ACCEL arms the whole group.

Variable	Type	Default	Effect
DASLLAMA_ACCEL	flag	off	Arm the Apple Accelerate / AMX float-batch override.
DASLLAMA_ACCEL_F16	flag	off	Use the BNNS f16 lane inside the Accelerate backend.
DASLLAMA_ACCEL_F16_STRIPS	number	0	Strip count for the f16 lane; 0 lets the backend choose.
DASLLAMA_ACCEL_STRIPS	number	d/256, max 8	Strip count for the f32 sgemm lane; 0 keeps the shape-derived default.
DASLLAMA_ACCEL_MIN_MMAC	number	backend default	Minimum MMAC count below which Accelerate declines and the daslang kernel runs.
DASLLAMA_ACCEL_MIN_NTOK	number	32	Minimum token count for the Accelerate float-batch override, floor 1.

11.7. Harness - tuner and probes

Variable	Type	Default	Effect
DASLLAMA_MODELS_DIR	path	unset	Directory holding the .gguf models the probes, benches and tests load. dasllama-server’s model catalog downloads here too when set.
DASLLAMA_CONFIRM_MODEL	path	auto-resolved from the models dir	Model used by the tuner’s confirm gate (FULL path, not a bare filename). Unset: the gate auto-resolves from the models dir - the preferred confirm carrier, else the largest present q8 gguf; the fallback pins only when the box has no q8 model at all.
DASLLAMA_CONFIRM_IQ2XXS	path	auto-resolved from the models dir	IQ2_XXS vehicle for the tuner’s serving-crown confirm (FULL path). Unset: the confirm auto-resolves the first IQ2_XXS.gguf in the models dir; none found = no crown (the base kernel serves).
DASLLAMA_CONFIRM_MTP_GEMMA	path	auto-resolved from the models dir	gemma-4 vehicle for the tuner’s assistant-depth confirm (FULL path; its mtp- head must sit beside it). Unset: the confirm auto-resolves the first gemma-4*.gguf with a head in the models dir; none found = depth 1 serves
DASLLAMA_BATCH_CHUNKS	text	unset	Override the batched-dispatch chunk count in the 1-core GEMM probe.
DASLLAMA_PROBE_MODEL	text	Qwen3VL-8B-Instruct-Q8_0.gguf	Model filename (inside the models dir) the image-turn attribution probe loads.
DASLLAMA_BATCH_GRID_2D	number	unset	Use the 2D batch grid in the parity probe.
DASLLAMA_FOCUS_BACKEND	text	unset	Restrict the 1-core GEMM probe to one backend.
DASLLAMA_FOCUS_SHAPE	text	unset	Restrict the 1-core GEMM probe to one shape.
DASLLAMA_FOCUS_NTOK	number	unset	Restrict the 1-core GEMM probe to one token count.
DASLLAMA_FOCUS_FORKED	flag	off	Run the 1-core GEMM probe through the fork pool rather than inline.
GEN_TUNE_COMPILE_ONLY	flag	off	Compile the generated tune probe without running it.
SMMLA_PROBE_COMPILE_ONLY	flag	off	Compile the smmla probe without running it.
METAL	flag	off	Run the Accelerate contention probe against the Metal backend.
DAS_DETAIL	flag	off	Per-job dump in the dispatch probe.
DAS_SPIN_NS	number	unset	Synthetic per-job spin in nanoseconds for the dispatch probe.
DAS_BATCH_DISPATCH	flag	off	Exercise batched dispatch in the dispatch probe.
DAS_TWO_WAVE	flag	off	Exercise the two-wave dispatch shape in the dispatch probe.
DAS_FORK_SKIP_HEAP_RESET	flag	off	Skip restartHeaps on pooled fork-context reuse (A/B for the fork-pool heap cost).
DAS_WORKER_SPIN_US	number	unset	Worker spin-before-park window in microseconds, applied via set_jobque_spin_us.
DAS_JOBQUE_JOIN_SPIN	number	unset	Join-side spin window for the 1-core GEMM probe.

11.8. Benchmarks

Variable	Type	Default	Effect
DASLLAMA_ALLOW_UNTUNED	flag	off	Let a bench cell measure on fallback/absent tune winners, and skip the auto-tune mint entirely (untuned kernels keep fallback stamps instead of paying the minutes-long tune). The gate exists because a stale sidecar silently downgrades every kernel and the whole board sandbags; only set for dev runs whose numbers are never published.
DASLLAMA_BENCH_MODEL	text	tinyllama	Model name for the isolated GEMM bench.
DASLLAMA_BENCH_NTOK	number	unset	Token count for the isolated GEMM bench.
DASLLAMA_BENCH_SKIP_ROWMAJOR	flag	off	Skip the row-major arm of the isolated GEMM bench.
DASLLAMA_BENCH_VERBOSE	flag	off	Per-run detail from the comparison bench (lcpp_bench).
DASLLAMA_MIN_CHUNK_ROWS	number	unset	Override the minimum rows per dispatch chunk in the GEMM bench.
DASLLAMA_FUSED_DECODE	flag	unset	Exercise the fused decode arm in the emission bench.
DASMETAL_LAB_ROUNDS	number	3	Timing rounds per cell in the Metal kernel labs.
DASMETAL_LAB_PASSES	number	4	Passes per round in the Metal MoE lab.
DASMETAL_LAB_ATTN_NSGS	number	16	Simdgroups per threadgroup in the Metal attention lab.
DASMETAL_LAB_DUMP_MSL	text	unset	Dump generated MSL from the Metal labs instead of running it.
DASMETAL_LAB_ARMS	text	unset	Substring filter restricting which arms the Metal MoE lab runs.
DASMETAL_LAB_SHAPES	text	unset	Substring filter restricting which shapes the Metal GEMV/GEMM labs run.
DASMETAL_LAB_VARIANTS	text	unset	Substring filter restricting which kernel variants the Metal GEMV/GEMM labs run.
PROBE_PATH	path	_wcliff.bin	Scratch file the write-cliff probe writes, cwd-relative by default; point it at the drive under test.
PROBE_GB	number	50	Gigabytes the write-cliff probe writes before reporting.

11.9. Profiling and baselines

Variable	Type	Default	Effect
DASLANG_BIN	path	bin/daslang	daslang binary the profiling harness shells out to.
DASLLAMA_BOX	text	probed	Override the box identifier recorded in a profile.
DASLLAMA_CPU	text	probed	Override the CPU name recorded in a profile.
DASLLAMA_OS	text	probed	Override the OS name recorded in a profile.
DASLLAMA_RAM_GB	number	probed	Override the RAM size in GiB recorded in a profile.
DASLLAMA_THREADS	number	probed	Override the thread count recorded in a profile.
DASLLAMA_GPU_NAME	text	probed	Override the GPU name recorded in a box profile.
LLAMA_BENCH	path	unset	Path to the llama-bench binary used to establish reference baselines.
LLAMA_BENCH_CLEAN	path	unset	Path to a clean-build llama-bench for the reference column.
LLAMA_BENCH_STOCK	path	unset	Path to a stock-build llama-bench for the reference column; its sibling llama-mtmd-cli serves as the metal/amx image-chat reference.
LLAMA_BENCH_VULKAN	path	unset	Path to a Vulkan-build llama-bench (setup_lcpp_ref’s vulkan flavor) for the vulkan leg’s reference column.
LLAMA_BENCH_CUDA	path	unset	Path to a CUDA-build llama-bench (-DGGML_CUDA=ON at the reference tag) for the cuda leg’s reference column.
WHISPER_CPP	path	../whisper.cpp	whisper.cpp checkout used for ASR reference baselines.
WHISPER_CPP_MODELS	path	unset	Directory of whisper.cpp models for ASR reference baselines.
CRISPASR	path	../CrispASR	CrispASR checkout used for the GigaAM ASR reference baselines (its Metal build in build/, its CPU-only build in build-cpu/).
MARIAN	path	../translations	Mozilla translations checkout whose patched inference/build/marian-decoder is the Bergamot translation reference.
MTMD_BIN	path	unset	llama-mtmd-cli binary for multimodal reference baselines.
DASLLAMA_BASE_PYTHON	path	unset	Base CPython 3.10-3.12 for creating the ASR oracle venvs, when neither PATH names nor uv resolve one in range.
NEMO_PY	path	unset	Python interpreter of the NeMo oracle venv, for canary/parakeet baselines.
ONNX_PY	path	unset	Python interpreter of the ONNX oracle venv, for parakeet baselines.
MLX_PY	path	unset	Python interpreter of the MLX reference venv (mlx-audio, mlx-lm), for the Apple-GPU baselines.
TTS_PY	path	unset	Python interpreter of the TTS reference venv (kokoro, kittentts, onnxruntime, pocket-tts), for the text-to-speech baselines.
GIGAAM_PY	path	unset	Python interpreter of the GigaAM oracle venv (upstream gigaam + torch), for the GigaAM external-reference dumps.
NMT_PY	path	unset	Python interpreter of the translation oracle venv (numpy, sentencepiece), for the Bergamot marian stage dumps.
OS	text	set by Windows	Read to detect Windows (Windows_NT); set by the OS, not by dasLLAMA.
PROCESSOR_ARCHITECTURE	text	set by Windows	Read for the CPU architecture on Windows; set by the OS, not by dasLLAMA.
NUMBER_OF_PROCESSORS	number	set by Windows	Read for the core count on Windows; set by the OS, not by dasLLAMA.

11.10. Tests

Variable	Type	Default	Effect
DASLLAMA_PARITY_FULL	flag	off	Run the FULL parity matrix. Without it the big-model rows silently skip.
DASLLAMA_TEST_ARMS	text	unset	Comma-separated filter restricting which parity arms run.
DASLLAMA_TEST_FAMILY	text	unset	Comma-separated filter restricting which model families run.
DASLLAMA_LLAMA2C_DIR	path	unset	Directory of llama2.c reference checkpoints for the forward/decode parity tests.
DASLLAMA_WHISPER_DIR	path	unset	Directory of whisper models for the audio tests.
DISPLAY	text	unset	Ambient platform variable; read only by the browser examples’ smoke tests (storyteller, storywish, parrot) to tell whether a Linux box has a window server for the example’s window.
WAYLAND_DISPLAY	text	unset	Ambient platform variable (the Wayland twin of DISPLAY); read only by the browser examples’ smoke tests.
DASLLAMA_CORPUS_DIR	path	unset	Directory of audio corpus files for the transcription tests.
TMPDIR	path	/tmp	Scratch directory for test artifacts; set by the OS on macOS.
TEMP	path	unset	Windows scratch-directory fallback when TMPDIR is unset (the test runner’s log dir).

11.11. Examples and tutorials

Variable	Type	Default	Effect
DASLLAMA_MODEL	path	unset	Model path for the dasLLAMA tutorials when none is passed on the command line.

11.12. Sidecar exchange

Overrides for the dasllama.io exchange client.

Variable	Type	Default	Effect
DASLLAMA_EXCHANGE_URL	text	unset	Sidecar exchange base URL override (tests, mirrors); unset = the baked-in dasllama.io. Setting it is an expressed choice to use the exchange - the first-contact consent question is skipped (announced at boot).
DASLLAMA_EXCHANGE_ACCEPT	text	unset	One-shot exchange accept-policy override: verified

11.13. daslang core knobs dasLLAMA honours

Owned by daslang, not by dasLLAMA - listed because dasLLAMA’s behaviour depends on them. DAS_TUNE_* is covered in skills/tune.md.

Variable	Type	Default	Effect
HF_HOME	path	~/.cache/huggingface	Hugging Face hub root; fetch_models –convert resolves checkpoint blobs under <HF_HOME>/hub. Falls back to HOME’s default cache location when unset.
HOME	path	unset	Ambient platform variable; read to derive the per-user ~/.dasllama directory (when USERPROFILE is unset) and the default Hugging Face cache when HF_HOME is unset.
USERPROFILE	path	unset	Ambient platform variable (Windows twin of HOME); read only to derive the per-user ~/.dasllama directory.
DAS_JOBQUE_THREADS	number	conservative default	Total compute lanes for job queues (N-1 workers plus the caller). Overrides set_jobque_threads_cap; see skills/internal/environment_variables.md.
DAS_JOBQUE_AFFINITY	number	2 on darwin (engine [init]), else 0	Worker affinity: 0 off, 1 ideal-processor hint, 2 hard mask - on darwin mode 2 is QoS classification (no pin API) and dasLLAMA defaults it on. Matters on big SMT boxes.
DAS_JOBQUE_TEAM_RANK_GATE	number	profile-driven	Team-dispatch rank gate. When set, it suppresses the box profile’s own team_rank_gate knob.
DAS_TUNE_MANIFEST	path	.tune.json	Kernel-tuning sidecar to read/write. Point it somewhere writable when the app dir is read-only.
DAS_TUNE_MODE	text	unset	Kernel-tuning mode. The [tune] framework owns these; see skills/tune.md.
DAS_LOG_LEVEL	text	warning (daslang’s to_log floor); tests/run.das sets info for itself and its children when unset	The to_log floor daslang reads per call (trace, debug, info, warning, error or critical); see skills/internal/environment_variables.md.
DAS_LOGGER_LEVEL	text	DAS_LOG_LEVEL, then warning	The daslib/logger floor, read once at init; it wins over DAS_LOG_LEVEL where both are set. See skills/internal/environment_variables.md.
DAS_TUNE_POLICY	text	unset	Kernel-tuning policy override. The [tune] framework owns these; see skills/tune.md.
DAS_TUNE_COMPILE_FALLBACKS	text	unset	Semicolon-separated kernel fallbacks for the generated tune probe.
JOBQUE_PROFILING	flag	off	Compile the jobque marker rail in (read at COMPILE time by daslib/build_const). Without it trace_tag/trace_marker erase from the program; decode_prof –trace refuses on a build without it.