amoe-lora 0.1.0: working-state framework — attach/toggle/detach invariant-tested (toggle law + bit-exact detach), checkpoint v1 + legacy import verified against shipped campaign artifacts, reference-grade train/align with guards, DDP-aware, honesty diagnostics first-class
906715b verified | # examples/attach_caption.py — the production quickstart (see the | |
| # caption repo for the full card): load Qwen3.5-0.8B, attach the | |
| # shipped caption anchor, caption an image. Requires [hf] extras + GPU. | |
| from huggingface_hub import hf_hub_download | |
| import torch, amoe | |
| from transformers import AutoModelForImageTextToText, AutoProcessor | |
| ck = hf_hub_download("AbstractPhil/qwen3.5-0.8b-relay-caption", "v35e4_caption_full_s0.pt") | |
| model = AutoModelForImageTextToText.from_pretrained("Qwen/Qwen3.5-0.8B", torch_dtype=torch.float32).cuda() | |
| proc = AutoProcessor.from_pretrained("Qwen/Qwen3.5-0.8B") | |
| h = amoe.attach(model, ck, strict=False) # legacy ckpt lacks meta | |
| print("attached:", h.names, "— detach() restores bit-exact") | |