Skip to content

model-cli.md

Latest commit

 

History

History
86 lines (72 loc) · 4.57 KB

model-cli.md

File metadata and controls

86 lines (72 loc) · 4.57 KB
Raw

Model command reference

Reference for jetlink models, the installer's command, which runs jetlink-server models on the server's models folder. To choose or prepare a model before a drive, start with model management.

Model identifiers and storage

  • A ref is a 40-character commit hash from comma's openpilot repository. It names a big model, as listed under Settings > Models > Big Model on the comma.
  • A SHA-256 is the 64-character hash of the ONNX file. One ref resolves to exactly one SHA-256, and that never changes.
  • Both come from the model index, zoompilot/jetlink-model-index, which works out every big model's ONNX once an hour and publishes index/v1.json: the models to list, in order and by folder, and the ONNX and size of every ref it has seen, so an old pick still resolves. New models appear without a Jetlink update. The server keeps the last index that parsed and works offline with it; one it cannot use (another schema, or nothing listed) is refused and the previous one kept.
  • A ref the kept index does not know, or a SHA-256 it cannot size, fetches the index again once before the command gives up.
  • The ONNX itself comes from comma's LFS servers, asked in turn: comma's Hugging Face bucket by SHA-256, then the GitLab and Hugging Face LFS servers, then comma's model repo (commaai/openpilot_driving_models).
Path What is in it
<cache>/models/ The downloaded ONNX files, about 766 MB each
<cache>/engines/ The prepared engines and their sidecar files, one per backend and device
<cache>/registry/ index.json, the last model index that parsed, and local-models.json, records of models you imported

Default cache folders: model management. --cache DIR overrides it on every command.

Commands

jetlink models list      [--refresh] [--json]
jetlink models resolve   REF [--json]
jetlink models fetch     REF_OR_SHA256
jetlink models import    PATH [--name NAME]
jetlink models inventory [--json]
jetlink models rm        SHA256 [--artifacts] [--model]
jetlink models prepare   REF_OR_SHA256 [--backend auto|trt|ort] [--device D]
  • From a checkout, run jetlink-server models ... with the same arguments; each takes --cache DIR. jetlink models prepare passes a PC's --tensorrt-libs itself; jetlink-server models prepare on a PC needs it (the server command).
  • REF_OR_SHA256 is a ref, or the SHA-256 of a model the model index knows.
  • Output a script reads goes to standard output; progress and warnings go to standard error.
Command Does
list The models the index lists, in its order, with each model's size and state (downloaded, prepared). Uses the kept index if under an hour old; --refresh fetches it again. The table shortens refs; --json prints the control protocol's catalog event fields, each model's folder included.
resolve Prints the SHA-256 and size in bytes the index gives a ref, listed or not. A lookup in the kept index, so it needs no network.
fetch Downloads the ONNX to a .part file, checks size and hash, renames it into place, and prints its path. An interrupted download starts over.
import Hashes and copies an ONNX file you already have into <cache>/models/, records its name, and prints the SHA-256 and path.
inventory Downloaded models, prepared engines with backend and device, and disk use. * marks the engine a server started here would load. --json prints the inventory event fields.
rm --model deletes the download, --artifacts every prepared engine; name at least one. A prepared engine keeps working after its download is removed.
prepare Fetches the model if needed, then builds an engine and loads it once, as jetlink-server build does. That makes it the model loaded last, which a server started on this cache preloads.

Do not run prepare while a server uses the same cache. Concurrent builds are unsupported and can exhaust memory. On a Jetson or PC, jetlink models prepare refuses while the service runs:

jetlink stop
jetlink models prepare f877d7a0ccc3cce943c76e285214c020cd65c899
jetlink start

The apps prepare through their running server instead.

Exit codes

Code Meaning
0 Success
1 Refused, or the ref or model was not found
2 A network request failed
3 The download failed verification, by size or by hash