| 1 | # Echo 1.5 local server deployment profile for a consumer GPU. |
| 2 | # Model and pipeline behavior remain in the referenced inference configuration. |
| 3 | |
| 4 | inference: |
| 5 | config: configs/inference.consumer.fp4.yaml |
| 6 | checkpoint: checkpoints/echo15_fp4 |
| 7 | conditioning_cache_dir: data/conditioning_cache |
| 8 | |
| 9 | runtime: |
| 10 | gpu_ids: [0] |
| 11 | # Consumer GPUs must use explicit DiT layerwise offload. Auto/resident can |
| 12 | # trigger WDDM paging when model weights and activations approach VRAM size. |
| 13 | dit_residency: swap |
| 14 | gpu_headroom_fraction: 0.05 |
| 15 | ram_headroom_fraction: 0.10 |
| 16 | model_idle_seconds: 0 |
| 17 | disable_inference_workers: false |
| 18 | |
| 19 | server: |
| 20 | host: 0.0.0.0 |
| 21 | port: 8221 |
| 22 | workers: 1 |
| 23 | public_base_url: http://127.0.0.1:8221 |
| 24 | request_timeout_seconds: 30 |
| 25 | poll_seconds: 5 |
| 26 | queue_capacity: 1000 |
| 27 | callback_max_attempts: 0 |
| 28 | artifact_db_path: data/artifacts.sqlite3 |
| 29 | media_root: data/media |
| 30 | ffmpeg_path: ffmpeg |
| 31 | merge_timeout_seconds: 3600 |
| 32 | merge_max_input_bytes: 2147483648 |
| 33 |