Metadata for the Linear UI is now sneakily provided via a `MetadataAccumulator` node, which the client populates / hooks up while building the graph.
Additionally, we provide the unexpanded graph with the metadata API response.
Both of these are embedded into the PNGs.
- Remove `metadata` from `ImageDTO`
- Split up the `images/` routes to accomodate this; metadata is only retrieved per-image
- `images/{image_name}` now gets the DTO
- `images/{image_name}/metadata` gets the new metadata
- `images/{image_name}/full` gets the full-sized image file
- Remove old metadata service
- Add `MetadataAccumulator` node, `CoreMetadataField`, hook up to `LatentsToImage` node
- Add `get_raw()` method to `ItemStorage`, retrieves the row from DB as a string, no pydantic parsing
- Update `images`related services to handle storing and retrieving the new metadata
- Add `get_metadata_graph_from_raw_session` which extracts the `graph` from `session` without needing to hydrate the session in pydantic, in preparation for providing it as metadata; also removes all references to the `MetadataAccumulator` node
Our model fields use `model_name`, but the API response uses `name`. Some places use `model_type` but the API response used `type`.
Changed the API response to provide `model_name` and `model_type`, which simplifies how we manage models on the client substantially.
- rewrite Dockerfile
- add a stage to build the UI
- add docker-compose.yml
- add docker-entrypoint.sh such that any command may be used at runtime
- docker-compose adds .env support - add a sample .env file
* fix the test of the config system
* Add torchmetrics==0.11.4 to installer
- Closes#3700
- Closes#3658
---------
Co-authored-by: Lincoln Stein <lstein@gmail.com>
Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com>
To be consistent with max_cache_size, the amount of memory to hold in
VRAM for model caching is now controlled by the max_vram_cache_size
configuration parameter.
