7.21. dasllama-cli — dasLLAMA from the shell

dasllama-cli (utils/dasllama-server/cli.das) is the shell front end of dasLLAMA: the same engine, the same public facade and the same backend pick as dasllama-server, without the HTTP layer. One model, one conversation, one process - a script or a cron job asks a model, hears a clip or speaks a line with no server running. It ships beside the server in every bundle (dasllama-cli.exe on Linux and Windows, dasllama-server.app/Contents/MacOS/dasllama-cli on a Mac) and runs from the source tree under the JIT:

bin/daslang -jit utils/dasllama-server/cli.das -- chat --model <model.gguf>

7.21.1. Commands

Command	What it does
complete	Complete a prompt, streamed to stdout; the prompt is --prompt, --file or the first bare argument. --image <file> with --image-mmproj <gguf> asks about a picture. --quiet drops the token counters.
chat	A conversation from the terminal, or from --script <file> one message per line. /help lists the slash commands: /image and /audio attach a file to the next message (--image-mmproj / --audio-mmproj; the gemma-4 E-series mmproj carries both towers, pass it twice), /read sends a file’s text, /save and /load keep the transcript as JSON, /regen, /clear, /system, /stats, /speak on (with --tts, every reply lands in reply_<n>.wav under --out-dir), /exit. A line of """ opens and closes a multi-line message. A thinking model’s reasoning streams before its answer, a blank line between; --hide-thinking hides it, --no-think answers directly on a hybrid model.
transcribe	An audio file (wav, mp3, flac, ogg) to text through --asr (a GGUF pair takes --mmproj); --out writes it to a file, --lang hints the language where the model takes one - a model that detects it refuses a hint, and a code the model does not speak is refused with the codes it does.
speak	Text to a 16-bit WAV through --tts (--voice, --speed, --tts-lane q8 | f32). A voice the model does not carry is refused with the voices it does; a speed on a model that takes none is refused.
talk	The chain: --in <audio> heard by --asr, answered by --model, spoken by --tts into --out. --prompt <text> skips the hearing. The transcript, the reply and the three stage times are printed.
embed	A text’s embedding vector, one float per line, or --json.
tokenize	A text’s token ids and pieces, one id<TAB>piece per line, or --ids.
bench	The llama-bench rows on a model: pp512 and tg128 (-p, -n, -r), and tg128@N with --npl N; -o md prints the table. The standalone dasllama-bench remains the records instrument.

Every example runs as written on a box whose models directory carries the named files; utils/dasllama-server/test_cli.das runs each one:

dasllama-cli complete -m SmolLM2-135M-Instruct-Q8_0.gguf --quiet "The capital of France is"
dasllama-cli speak --tts kitten-nano.gguf -o hello.wav "Hello from the command line."
dasllama-cli transcribe --asr Qwen3-ASR-0.6B-Q8_0.gguf --mmproj mmproj-Qwen3-ASR-0.6B-bf16.gguf question.wav
dasllama-cli talk --asr Qwen3-ASR-0.6B-Q8_0.gguf --mmproj mmproj-Qwen3-ASR-0.6B-bf16.gguf --model Qwen3.5-0.8B-Q8_0.gguf --tts kokoro-82m.gguf --in question.wav --out answer.wav
dasllama-cli chat -m gemma-4-E2B-it-Q4_K_M.gguf --image-mmproj mmproj-gemma-4-E2B-it-bf16.gguf --audio-mmproj mmproj-gemma-4-E2B-it-bf16.gguf
dasllama-cli bench -m SmolLM2-135M-Instruct-Q8_0.gguf --npl 2 -o md

7.21.2. Flags

<command> --help prints the command’s own flags, then the groups it takes (from the source tree -? or --show-help, since the daslang host takes --help for itself). The shared flags are the server’s knobs under the server’s names; the config file fills whatever they leave empty (below).

Shared flags (every command)

Flag	Meaning
--config / -c	TOML config file (default: dasllama-server.toml in the cwd, else in ~/.dasllama, else beside the program); keys mirror the long flag names, explicit flags win
--model / -m	The LLM GGUF: a path, or a file name under --models-dir
--quant / -q	Weight serving mode: fp32 | q8 (default; the file-format spellings serve natively under q8) | q4 = the legacy requant tier
--kv-dtype	KV cache codec: f32 | f16 (default) | q8_0 | tq4 (rotated 4-bit; needs a power-of-two head size); under --gpu vulkan the whole-model driver holds its cache in the same codec for f16, q8_0 and tq4
--gpu	GPU backend: auto (default: metal or vulkan when detected, else CPU) | off | metal | metal-required | vulkan; vulkan serves a model that fits the card whole from the device, the conversation’s cache with it
--metal	Metal serving mode when --gpu is unset: off | auto | required
--gpu-layers	vulkan: resident MoE expert-stack layers, offloaded from the end (default auto under --gpu vulkan)
--gpu-stream	vulkan: streamed prefill MoE layers below the resident set (default auto under --gpu vulkan)
--gpu-dn	vulkan: deltanet (recurrent) layers through the device chain (default on under --gpu vulkan)
--gpu-attn	vulkan: full-attention layers through the device chain (default on under --gpu vulkan)
--gpu-dense	vulkan: dense attention-side planes resident (default off)
--gpu-vram-mb	vulkan: resident-weight VRAM cap override in MB (default: query the device)
--ctx	Context length in tokens, served whole: the whole-model GPU driver holds it or declines to the per-op rails, and the load log names the room. Default: the model’s trained context_length, shortened to what the card holds
--rope-scaling	RoPE scaling override: yarn | linear | none (default: the model file’s own rope.scaling keys; none drops them, a file’s per-pair factor tensors stay). The Qwen families publish the YaRN recipe and enable it as a setting, for long conversations only: --rope-scaling yarn --rope-scale 4; no other vendor validates it, and a non-Qwen file logs a warning. The override is baked into the prepared image under its own lane
--rope-scale	RoPE scaling factor for the override (default: the file’s rope.scaling.factor; yarn needs one)
--yarn-orig-ctx	YaRN: the original training context the factor extends (default: the file’s original_context_length, else its context_length)
--threads / -t	Worker-lane cap for the matmul dispatch (default 16; -1 = all cores)
--models-dir	Where a bare model name resolves (default ~/.dasllama/models, where the server’s catalog downloads land; DASLLAMA_MODELS_DIR overrides)
--tune	Re-tune this box’s dasLLAMA kernels, then relaunch (a fat build carries no tuner and ignores it)
--verbose	Echo the engine’s log records (dasllama-cli.log - ~/.dasllama/logs from an installed build, the das root’s logs/ from the source tree: a missing Metal profile, a GPU decline, a model load) to stderr as they land; stdout stays the answer
--help / -?	The command’s help (--show-help from the source tree)

Sampling (complete, chat, talk)

Flag	Meaning
--temp	Sampling temperature; 0 = greedy. Unset, complete is greedy and chat / talk sample at 0.7 - under greedy decoding a thinking model loops on its own draft and never reaches its answer
--top-k	Top-k cutoff (unset: 20 on the sampled default, off under greedy)
--top-p	Nucleus cutoff (unset: 0.95 on the sampled default)
--min-p	Drop tokens below this fraction of the top probability (default off)
--repeat-penalty	Multiplicative repetition penalty over the recent window (default 1 = none)
--presence-penalty	Subtracted from every token already in the window (default 0)
--frequency-penalty	Subtracted per occurrence in the window (default 0)
--seed / -s	The sampler’s seed for a reproducible run (default: the session’s fixed seed)
--max-tokens / -n	Reply token budget (default 256 for complete, 16384 for chat and talk)

Speaker (chat, speak, talk)

Flag	Meaning
--play	Also play each spoken reply through the speaker once its WAV has landed; the run blocks until the clip ends. A box with no audio device says so once and keeps writing the files
--null-audio	With --play: drive the null audio backend instead of a device (a test box, a CI runner)

complete

Flag	Meaning
--prompt / -p	The prompt (or --file, or the first bare argument)
--file / -f	Read the prompt from this file
--image	A picture to ask about (needs --image-mmproj); the family’s own thinking default applies
--image-mmproj	The vision mmproj GGUF for --image
--quiet	Print the completion only - no token counts or rates after it

chat

Flag	Meaning
--system	The system prompt
--script	Read the turns from this file instead of the terminal, one message per line (slash commands work); the reply follows each
--image-mmproj	The vision mmproj GGUF: arms /image
--audio-mmproj	The audio mmproj GGUF: arms /audio (the E-series mmproj serves both towers, pass the same file twice)
--tts	A TTS model GGUF: /speak on writes every reply to a WAV file
--voice / -v	Voice name for --tts (default: the model’s first voice)
--speed	Speech speed multiplier for --tts (default 1.0)
--out-dir	Where the spoken replies land, reply_<n>.wav each (default: the cwd)
--hide-thinking	Hide a thinking model’s reasoning (default: it streams before the answer, a blank line between)
--no-think	Answer directly on a hybrid thinking model (default: the model’s own default)
--quiet	Print the reply only - no token counts or rates after it

transcribe

Flag	Meaning
--asr / -a	The ASR model (GGUF or .bin): a path, or a file name under --models-dir
--mmproj	mmproj for a GGUF ASR model (paired with --asr)
--file / -f	The audio file to transcribe (wav / mp3 / flac / ogg; or the first bare argument)
--out / -o	Write the transcript here instead of stdout
--lang / -l	Language hint (default: auto for a model that detects it, else en); a model that detects the language refuses a hint, a code the model does not speak is refused with the codes it does

speak

Flag	Meaning
--tts	The TTS model GGUF: a path, or a file name under --models-dir
--text	The text to speak (or --file, or the first bare argument)
--file / -f	Read the text from this file
--out / -o	The WAV file to write (16-bit PCM, the model’s rate; default out.wav)
--voice / -v	Voice name or alias (default: the model’s first voice); a voice the model does not carry is refused with the voices it does
--speed	Speech speed multiplier (default 1.0); a speed on a model that takes none is refused
--tts-lane	TTS weight lane: q8 (default; the prepared quant image beside the GGUF) | f32 (the file’s own planes)
--prof	Print the generator’s per-op profile

talk

Flag	Meaning
--asr / -a	The ASR model (GGUF or .bin) that hears --in
--mmproj	mmproj for a GGUF ASR model (paired with --asr)
--in / -i	The audio file to answer (wav / mp3 / flac / ogg; or the first bare argument)
--prompt / -p	A text prompt to answer instead of --in
--system	The system prompt
--lang / -l	Language hint for the ASR model (default: auto for a model that detects it, else en)
--tts	The TTS model GGUF that speaks the reply
--voice / -v	Voice name for --tts (default: the model’s first voice)
--speed	Speech speed multiplier for --tts (default 1.0)
--out / -o	The WAV file the reply lands in (default reply.wav)
--no-think	Answer directly on a hybrid thinking model (default: the model’s own default)
--quiet	Print the reply only - no transcript, stage times or token counts

embed

Flag	Meaning
--text	The text to embed (or --file, or the first bare argument)
--file / -f	Read the text from this file
--json	Print the vector as a JSON array (default: one float per line)

tokenize

Flag	Meaning
--text	The text to tokenize (or --file, or the first bare argument)
--file / -f	Read the text from this file
--ids	Print the ids only, space-separated (default: one id<TAB>piece per line)
--no-special	Do not add the model’s BOS / leading specials

bench

Flag	Meaning
--plen / -p	Prompt tokens per pp rep (default 512, llama-bench -p; 0 skips the row)
--ngen / -n	Generated tokens per tg rep (default 128, llama-bench -n; 0 skips the row)
--reps / -r	Timed repetitions per row after an untimed warmup (default 5, llama-bench -r)
--npl	Streams of the batched tg row, tg<n>@<npl>: that many streams through the scheduler, the summed rate (default 0 = no batched row)
--output / -o	txt (default) | md (a llama-bench-style table)

7.21.3. The config file

A bare model name resolves under --models-dir (~/.dasllama/models, where the server’s catalog downloads land; DASLLAMA_MODELS_DIR overrides).

The dasllama-server.toml in the current directory, else in ~/.dasllama (where the server’s control page saves it), else beside the program - the server’s own lookup - fills whatever the flags leave empty: the model (a [[models]] roster’s default entry included), its image_mmproj and audio_mmproj, the asr and mmproj pair, the tts model and its lane (an [[asr]] / [[tts]] roster’s first table), gpu and the Vulkan detail keys, threads, ctx, models_dir. On a box the server’s setup page configured, dasllama-cli chat with no flags talks to the served model on the served backend. Explicit flags win; --config names another file.

Output goes three ways: the answer - the completion, the reply, the transcript, the vector, the ids - on stdout alone, so it pipes; the CLI’s progress lines and the token counters (prompt and generated tokens, time to first token, prefill and decode rates) on stderr, --quiet dropping the counters; and the engine’s own notices - a missing Metal profile, a GPU decline - in dasllama-cli.log (~/.dasllama/logs from an installed build, the das root’s logs/ from the source tree), where the server’s land too, echoed to stderr as they land under --verbose (from the source tree the tune policy guard still prints its one status line first). A refusal - no model, a file that does not exist, a flag value that does not parse - names what is missing and exits non-zero.

7.21.4. Tuning

The CLI shares the box’s tune winners with the server: requiring dasLLAMA pulls in its [tune_scope], and the policy is auto - the first run from the source tree on an untuned box tunes and relaunches, every later run is instant, --tune forces a re-tune. A fat bundle carries no tuner: its kernels are the class clones, and its first dasllama-cli run on a Mac races the Metal crowns once into ~/.dasllama/tune/dasllama-cli.tune.json, as the server’s first run does into its own. See Kernel tuning.

7.21.5. Testing

utils/dasllama-server/test_cli_args.das is the model-free half, run everywhere: the plan a command line parses to, the help surfaces, the config file’s keys and roster, the model-path resolution, and the chat loop’s line logic. utils/dasllama-server/test_cli.das runs every command end to end, one child process per cell, on the small models a stocked box carries - the talk chain on two model columns and the picture and the clip through the gemma-4-E2B towers - and skips by name where a model is absent.

See also

dasllama-server — an OpenAI-compatible server over dasLLAMA - the OpenAI-compatible server the CLI sits beside

dasLLAMA-08 — Audio Chat - the audio turn the /audio command and talk are built on