File size: 1,602 Bytes
24196fc | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 | # ProMoE — Hub custom pipeline
Load checkpoints with **native Hugging Face diffusers** and this folder on the Hub (or via `custom_pipeline`):
```python
import torch
from diffusers import DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained(
"BiliSakura/ProMoE-diffusers",
trust_remote_code=True,
torch_dtype=torch.float16,
)
pipe.to("cuda")
```
## Hub layout
| Path | Purpose |
| --- | --- |
| `pipeline.py` | `ProMoEPipeline` |
| `transformer/` | backbone_diffmoe.py, backbone_dit.py, backbone_ecdit.py, backbone_promoe_ec.py, backbone_promoe_tc.py, backbone_tcdit.py, … |
| `scheduler/` | scheduling_flow_match_promoe.py |
## ImageNet class labels
Each variant keeps an English `id2label` map in `model_index.json` (DiT-style).
- `pipe.id2label` — id → English label (comma-separated synonyms)
- `pipe(class_labels=207, ...)` — class-conditional sampling with integer ids
Copy the full 1000-class `id2label` block from `BiliSakura/DiT-diffusers` when publishing a model repo.
## `model_index.json`
Copy entries from `model_index.json.example` into your model repo after `save_pretrained`.
Use `["_class_name"] = ["pipeline", "ProMoEPipeline"]` and custom module stems for each component.
- FlowMatch scheduler: `"scheduler": ["scheduling_flow_match_promoe", "ProMoEFlowMatchScheduler"]`
- VAE: `"vae": ["diffusers", "AutoencoderKL"]` with `stabilityai/sd-vae-ft-mse` weights or bundled safetensors
- ProMoE-TC presets: `ProMoE_TC_S`, `ProMoE_TC_B`, `ProMoE_TC_L`, `ProMoE_TC_XL` (see convert script)
Regenerate: `python scripts/build_community_pipelines.py`
|