| 1 | # Echo 1.5 full DMD BF16 single-GPU inference configuration. |
| 2 | # Relative paths are resolved from the repository root. |
| 3 | |
| 4 | paths: |
| 5 | checkpoint: checkpoints/echo15_full_dmd |
| 6 | gemma_path: checkpoints/gemma-3-12b |
| 7 | requests_dir: examples/the_last_visa/requests |
| 8 | requests_glob: "*.json" |
| 9 | output_root: inference_result |
| 10 | # Set this to use complete precomputed text/image/audio conditioning. |
| 11 | conditioning_cache_dir: |
| 12 | |
| 13 | video: |
| 14 | num_frames: 241 |
| 15 | height: 736 |
| 16 | width: 1280 |
| 17 | fps: 25 |
| 18 | seed: 42 |
| 19 | |
| 20 | # Tiled decode is the safe public default across supported devices. Use |
| 21 | # `untiled` explicitly only after measuring sufficient device-memory headroom. |
| 22 | video_vae: |
| 23 | decode_mode: tiled |
| 24 | tile_size_frames: 64 |
| 25 | tile_overlap_frames: 24 |
| 26 | tile_size_pixels: 512 |
| 27 | tile_overlap_pixels: 64 |
| 28 | |
| 29 | # The production BF16 DMD schedule. Each step predicts x0 and then re-noises |
| 30 | # with fresh Gaussian noise at the next sigma. |
| 31 | denoising: |
| 32 | steps: [1000, 994, 988, 981, 975, 909, 725, 422, 0] |
| 33 | sigmas: [1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875, 0.0] |
| 34 | |
| 35 | memory: |
| 36 | max_size: 7 |
| 37 | downscale_factor: 1 |
| 38 | enable_audio: true |
| 39 | position_mode: slot_center |
| 40 | position_offset: 500.0 |
| 41 | position_slot_stride: 50.0 |
| 42 | |
| 43 | # Production-aligned speech/music/effects separation. Run |
| 44 | # `python scripts/setup_msst.py` before inference. |
| 45 | voice_filter: |
| 46 | enabled: true |
| 47 | backend: msst_speech |
| 48 | min_output_rms: 0.004 |
| 49 | msst_dir: third_party/MSST-WebUI |
| 50 | msst_model_path: checkpoints/msst/model_bandit_plus_dnr_sdr_11.47.chpt |
| 51 | msst_config_path: third_party/MSST-WebUI/configs_backup/multi_stem_models/model_bandit_plus_dnr_sdr_11.47.chpt.yaml |
| 52 | msst_model_type: bandit |
| 53 | msst_sample_rate: 44100 |
| 54 | msst_device: auto |
| 55 | msst_local_rank_env: LOCAL_RANK |
| 56 | |
| 57 | inference: |
| 58 | device: cuda |
| 59 | dtype: bfloat16 |
| 60 | # Match the production encoder guard and avoid oversized text conditioning. |
| 61 | prompt_max_chars: 1500 |
| 62 | # Increase for throughput when memory allows; use 1 for the lowest text-stage peak. |
| 63 | text_batch_size: 1 |
| 64 | image_batch_size: 1 |
| 65 | audio_batch_size: 1 |
| 66 | # Experimental H3-style DiT weight streaming. Disabled keeps the release |
| 67 | # pipeline unchanged. Immutable block weights remain on CPU when enabled. |
| 68 | dit_layerwise_offload: |
| 69 | enabled: false |
| 70 | resident_blocks: 0 |
| 71 | prefetch_blocks: 1 |
| 72 | pin_memory: true |
| 73 |