Skip to content

PS5SD 0.9.16

Choose a tag to compare

@Darkatek7 Darkatek7 released this 28 Oct 10:23
· 31 commits to main since this release

PS5SD 0.9.16 switches the model without stopping the server, so the web UI and the API stay up, and lets clients wait for a model as A1111 and OpenAI clients expect.

A model switch without stopping the server

  • A switch loads the next model into the running server. Before, the server stopped and started again, and nothing answered on the port while the model loaded.
  • The web UI, the gallery, the library, the job list and the event stream keep answering during a switch.
  • A request that needs the model waits until it has loaded, then runs on it. A job queued meanwhile stays queued and runs on the new model.
  • Measured on the console, five rounds of FLUX.2 Klein 9B with a LoRA and SDXL against 0.9.15: the same times and memory, a lower CPU heap (0.81 instead of 0.94 GiB at the peak), and no request without an answer (0 of 893, against 19 of 862). CONSOLE.md has the table.

Waiting for a model

  • POST /sdapi/v1/options with sd_model_checkpoint answers once the model has loaded, as A1111 does: 8.3 s for SDXL on the console. A model that does not load answers 500 load_failed with the reason; after 10 minutes, 504 timeout.
  • The model of /v1/images/generations and /v1/images/edits picks the model: a model of the library is switched to and waited for. Any other name, such as the dall-e-3 or gpt-image-1 that OpenAI clients send, runs on the loaded model.

The switch in the event stream

  • GET /ps5sd/v1/events sends a model event with active, loaded, switching, standby and load_error whenever they change.
  • The web UI follows it, and shows a switch that another page or a client started.

See API.md for the details.

Install

Install from the homebrew store, or unpack PPSA99576.zip into /data/homebrew/ on the console and let ShadowMountPlus scan.

Next: see ROADMAP.md.