RFD 1044 details: why a planner, the nine actions, two runtimes
Why a planner earns its place here
The domain carries a_load and a_unload as real actions. The peak memory is then a planning result, and not an accident.
make_depth does not read the layer output. Marigold runs on the original image, thus the planner may order it before or after make_layers. A hand-written script fixes one order forever.
The nine, and their guards
| Action | Needs |
|---|---|
| a_inpaint | lama, the image |
| a_encode_prompt | layerdiff_te1, layerdiff_te2 |
| a_diffuse_layers | the embeds, the inpaint |
| a_decode_rgb | the latents |
| a_decode_alpha | the latents |
| a_depth_encode | marigold_te, marigold_unet |
| a_depth_decode | the depth latents |
| a_write_psd | rgb, alpha, and depth |
a_decode_rgb and a_decode_alpha read the same latents. That is the one fact a reader of the old script always missed.
Two runtimes, one domain
RFD 1030 records a Replicate path at 9.82 GB in bf16, and a local ggml path at about 2.7 GB in Q4_K_M. The domain does not change between them. Only the action bodies change.