Command-line tools

A release ships four binaries. This page shows what each prints for --help, captured from the binaries by website/scripts/capture-cli.mjs. The guides explain them in context: inference quickstart, compile and benchmark, benchmarking and kernel cache.

phobos-cli

Runs a GGUF model: on a prompt, as a REPL, or as a server with --listen. The options shown are the ones this site covers.

usage: phobos-cli --gguf FILE [OPTIONS] [PROMPT]

With a PROMPT: print the model's answer. Without one: start a REPL, which
answers each line on its own. A model whose file carries a chat template is
asked in it; --raw sends the text as typed and prints its continuation.

--gguf is required; there is no default model.

OPTIONS:
      --gguf FILE     run a GGUF model, dispatching on the architecture the
                      file declares, with the tokenizer the file carries
      --expert-cache SIZE
                      device memory for a streamed model's expert cache,
                      2g, 1500m or bytes; default what the resident weights
                      leave, less a reserve. Ignored without experts
  -n, --num N         generate up to N tokens (default: 200; stops early on an
                      end-of-turn token or the model's context limit)
  -t, --temp T        sampling temperature; 0 = greedy/argmax (default: 0)
  -k, --top-k K       sample only from the K highest-logit tokens (default: 0 = off)
  -p, --top-p P       nucleus sampling threshold (default: 1.0 = off)
      --min-p P       keep tokens at least P as likely as the best one
                      (default: 0.0 = off)
      --presence-penalty P
                      subtract P from the logit of every token generated so far,
                      the prompt excluded (default: 0.0 = off)
      --repetition-penalty P
                      scale the logit of every token already in the sequence,
                      prompt included, towards zero by P (default: 1.0 = off)
      --seed S        PRNG seed for sampling (default: 0)
      --show K        show the top-K candidates for the first token (default: 5)
      --raw           send the prompt as typed instead of as a user turn of
                      the model's chat template, and print its continuation;
                      for a base model, or text that is already a prompt
      --prefill-thinking TEXT
                      open the answer's think block on TEXT and let the
                      model reason on from it. When serving, this
                      applies to every chat request, whatever it asks for.
                      Ignored with --raw or a model with no chat template
      --listen ADDR   run an OpenAI compatible HTTP server on ADDR (e.g.
                      127.0.0.1:8080). The sampling options above become what a
                      request falls back to for every field it does not send.
      --no-tui        serve without the dashboard, printing the runtime's lines
                      instead. Implied when stdout is not a terminal, and the
                      reason is printed whenever that happens
      --tui           draw the dashboard even where the terminal was not
                      recognized as one
      --no-prefix-cache
                      start every request from an empty cache. By default a
                      finished request's session is kept and the next request
                      reuses as much of it as the two prompts agree on, which
                      is most of a chat. Turning it off costs that and gains
                      the cache's buffers back for a pass to use as scratch,
                      which is the difference on a card that only just fits
  -V, --version       print the version and the compiler fingerprint
  -h, --help          print this message

phobos-compile

Compiles a kernel source file to PTX. No GPU is needed. The PTX goes to standard output unless -o names a file. --chip defaults to sm_75, which every supported card runs, and --index-bits to 32, raised to 64 when the module needs it, as the runtime would. Set PHOBOS_PRINT_PHASES=1 to print the IR of each lowering phase.

usage: phobos-compile FILE.ph [--chip CHIP] [--index-bits 32|64] [-o OUT.ptx]

phobos-bench

Measures a GGUF model’s prompt processing (pp<N>) and text generation (tg<N>) in tokens per second, as llama-bench does. It needs a GPU.

usage: phobos-bench [OPTIONS]

OPTIONS:
  -m, --model FILE      the GGUF file to run (default: models/Qwen3.5-0.8B-Q8_0.gguf)
  -p, --n-prompt N,...  prompt-processing tokens, one pp<N> row each (default: 128)
  -n, --n-gen N,...     generated tokens, one tg<N> row each (default: 32)
  -d, --depth N,...     positions in the context before each tg row starts,
                        one set of tg rows per depth (default: 0)
  -r, --repetitions N   timed repetitions per row (default: 3)
      --no-warmup       skip the warmup pass, which leaves each kernel's first
                        compile inside the repetition it lands in
      --expert-cache SIZE
                        device memory for a streamed model's expert cache
                        (2g, 1500m, bytes); default, what the weights leave
  -h, --help            print this message

  -p and -n take a comma-separated list, and repeat, so -p 128,512 and
  -p 128 -p 512 both ask for the same two rows.

phobos-cache

Fills, lists and clears the compiled-kernel cache. The subcommands are explained on the kernel cache page.

usage: phobos-cache <warm|list|clear|prune> [--chip CHIP]... [--dir CACHE] [--manifest DIR]... [--jobs N] [--force] [--ptxas PATH | --no-ptxas] [KERNEL...]

All four print the version and the compiler fingerprint with --version.