Recipe file

A recipe is one model on one card. It names the exact engine image, the exact model file, the settings that were measured, and the numbers they gave. The launcher only runs what a recipe describes.

Location

recipes/<org>/<repo>/<card>.toml

<org>/<repo> is the model’s Hugging Face repository, and must equal model.hf_repo.
<card> is the card id, and must equal card.id.
The built-in recipes are the repository’s recipes/ directory, embedded into the launcher when it is built. --recipes DIR reads DIR/<org>/<repo>/<card>.toml instead.
Other files, and files and directories whose name starts with ., are ignored.
One invalid recipe fails the whole load.

Example

schema = 1
status = "released"

[model]
name       = "pn3"
title      = "Qwen3.8-27B-PN3 (P100-native 3-bit, mix P0k+19)"
hf_repo    = "iamMrDuncan/Qwen3.8-27B-PN3-GGUF"
revision   = "c701b6690ec360b16d40f85df1cb76062e30cd38"
file       = "Qwen3.8-27B-PN3-P0k19.gguf"
sha256     = "15ce2d4469673f1f1753c55de3780cfe04311dd6f05cb41633d73df128e91476"
size_bytes = 12954114496
license    = "Apache-2.0 (from Qwen/Qwen3.8-27B)"

[card]
id                 = "p100"
title              = "NVIDIA Tesla P100 16 GB"
compute_capability = "6.0"
min_memory_mib     = 16000
min_driver         = "525.60.13"

[engine]
release = "engine-v0.1.0"
arch    = "sm60"
image   = "ghcr.io/iammrduncan/tensorless-engine@sha256:35409ff97c2f518e468790c28bd19206ae777a945fcffe5775585e28fd9939a7"

[run]
alias       = "iamMrDuncan/Qwen3.8-27B-PN3-GGUF"
context     = 65536
threads_max = 8
args = ["-np", "1", "-ngl", "99", "-ot", "token_embd.weight=CUDA0", "-ctk", "q4_0", "-ctv", "q4_0",
        "-sm", "none", "-lm", "none", "-fit", "off", "--threads-http", "4", "-b", "512", "-ub", "512",
        "-fa", "on", "--jinja"]

[run.env]
GGML_CUDA_FATTN_SM60 = "1"
# ... the other engine switches

[receipt]
summary     = "Decode about 30 tok/s at short context, 1.14x the Q3 baseline model on the same card. Not yet measured at 32K or 65K."
decode_tps  = { short = 30.06 }
prefill_tps = { ctx_4k = 121.4 }

The full files are in recipes/. The schema’s own description is recipes/README.md.

Top level

Key	Type	Required	Meaning
`schema`	integer	yes	The recipe schema. Must be `1`; any other value fails with a request to update tensorless.
`status`	string	yes	`"draft"` or `"released"`. See Draft, released and unpublished.

[model]

Key	Type	Required	Meaning
`name`	string	no	A short alias, such as `pn3`, that works anywhere a repo id does. Lowercase letters, digits, `.`, `_` and `-`. It must not equal any repo id, nor another model’s short name.
`title`	string	no	A one-line human name, shown by `info` and `models --json`.
`hf_repo`	string	yes	The Hugging Face model repository, `org/name`. This is what users type, matched without case. Two repo ids that differ only in case are refused.
`revision`	string	yes	The repository commit: 40 lowercase hex characters. Never a branch or tag name.
`file`	string	yes	The weights file’s path inside the repository: relative, with forward slashes, clean (no `..`), without `:` or newlines.
`sha256`	string	yes	The file’s sha256, 64 lowercase hex characters. The launcher checks the download against it.
`size_bytes`	integer	yes	The file’s exact size, for the disk-space check and progress. Must not be negative. `0` marks the recipe unpublished.
`license`	string	yes	The model’s license.

[card]

Key	Type	Required	Meaning
`id`	string	yes	The card id, such as `p100`. Lowercase letters, digits, `.`, `_` and `-`. Must equal the file name without `.toml`.
`title`	string	no	The card’s human name, such as `NVIDIA Tesla P100 16 GB`.
`compute_capability`	string	yes	Must equal the `compute_cap` that `nvidia-smi` reports, such as `"6.0"`.
`min_memory_mib`	integer	yes	The GPU’s `memory.total`, in MiB, must be at least this. Positive.
`min_driver`	string	yes	The oldest NVIDIA driver the recipe runs on, a dotted version such as `"525.60.13"`.

A GPU matches a card when its compute capability equals compute_capability and its memory is at least min_memory_mib. A matching GPU with an older driver is refused, with the driver to install.

[engine]

Key	Type	Required	Meaning
`release`	string	yes	The engine release the recipe was measured on, such as `engine-v0.1.0`.
`arch`	string	yes	The GPU architecture of the image, such as `sm60`.
`image`	string	yes	The engine image, always by digest: `name@sha256:<64 hex>`. A tag is refused.

[run]

Key	Type	Required	Meaning
`alias`	string	yes	The served model name, as `/v1/models` lists it. Letters, digits, `.`, `_`, `-` and `/`; the recipes use the repo id. Passed as `--alias`.
`context`	integer	yes	The context length, passed as `-c`. Positive. `--context` overrides it and makes the run untuned.
`threads_max`	integer	yes	The most CPU threads to use. `-t`, `-tb` and `OMP_NUM_THREADS` are the smaller of this and the host’s CPU count. Positive.
`args`	array of strings	no	Every other `llama-server` argument, in order.
`env`	table of strings	no	The engine switches, exactly as measured. Each key must be a valid variable name. See Engine switches.

The launcher sets these itself, so args must not contain them: -m, --model, --host, --port, -c, --ctx-size, -a, --alias, -t, --threads, -tb, --threads-batch. env must not set OMP_NUM_THREADS.

[receipt]

What was measured, and where. Every key is optional. info shows them, and bench compares its result with decode_tps.short.

Key	Type	Meaning
`source`	string	Who measured the numbers, and on what build.
`summary`	string	One or two sentences on the result.
`decode_tps`	table of numbers	Decode speed in tokens per second. Keys: `short`, `ctx_32k`, `ctx_65k`. Not negative.
`prefill_tps`	table of numbers	Prefill speed in tokens per second, for example `ctx_4k`. Not negative.
`notes`	array of strings	The scope of each number, and anything else a reader needs.

info shows measurement keys in the order short, ctx_4k, ctx_32k, ctx_65k, then any other keys alphabetically. The full measurements of a release are in its receipts.

Checks

The launcher refuses a recipe file that:

has a key the schema does not have (unknown keys);
has a value that breaks a rule in the tables above;
sits at a path that does not match its model.hf_repo and card.id.

Draft, released and unpublished

State	How it is set	`pull`, `run`, `serve`	`info`	`--dry-run`	`models` shows
Released	`status = "released"`	Allowed.	Shown.	Shown.	`released`
Draft	`status = "draft"`	Refused unless `--allow-draft`.	Shown.	Refused unless `--allow-draft`.	`draft`
Unpublished	An all-zero `revision`, `sha256` or image digest, or `size_bytes = 0`	Refused, even with `--allow-draft`.	Shown, with `NOT PUBLISHED`.	Shown.	`draft, unpublished` or `released, unpublished`

A release moves a recipe to released only after the release gate has measured its image on its card. A pull request that changes [run] or [engine] needs a new receipt.