# # For licensing see accompanying LICENSE file. # Copyright (C) 2026 Apple Inc. All Rights Reserved. # """Checkpoint loading for Wan backbones, FastFrames generators and their add-on networks.""" from __future__ import annotations import json from pathlib import Path import torch from .wan.causal_model import CausalWanModel from .wan.model import WAN_CONFIGS, WanModel __all__ = ["load_state_dict", "build_wan", "load_wan_base"] _PREFIXES = ("model._fsdp_wrapped_module.", "_fsdp_wrapped_module.", "model.") def load_state_dict(path, key=None): """Tensor state from ``.safetensors`` (a file, or the ``.safetensors.index.json`` of a sharded file) or a torch pickle (optionally nested under ``key``).""" path = Path(path) if path.name.endswith(".safetensors.index.json"): from safetensors.torch import load_file state = {} for name in sorted(set(json.loads(path.read_text())["weight_map"].values())): state.update(load_file(str(path.parent / name))) elif path.suffix == ".safetensors": from safetensors.torch import load_file state = load_file(str(path)) else: state = torch.load(path, map_location="cpu", mmap=True, weights_only=True) if key is not None: state = state[key] return {_strip(k): v for k, v in state.items()} def _strip(name): for prefix in _PREFIXES: if name.startswith(prefix): return name[len(prefix):] return name def build_wan(name, state_dict, *, causal=True, dtype=torch.bfloat16, device="cuda"): """Instantiate a (causal) Wan2.1 transformer without materialising random weights.""" cls = CausalWanModel if causal else WanModel with torch.device("meta"): model = cls(WAN_CONFIGS[name]) model.load_state_dict(state_dict, strict=True, assign=True) return model.to(device=device, dtype=dtype).eval().requires_grad_(False) def load_wan_base(wan_dir, *, causal=False, dtype=torch.bfloat16, device="cuda"): """Load the released Wan2.1 weights (``diffusion_pytorch_model*.safetensors``).""" from safetensors.torch import load_file wan_dir = Path(wan_dir) index = wan_dir / "diffusion_pytorch_model.safetensors.index.json" files = (sorted(set(json.loads(index.read_text())["weight_map"].values())) if index.exists() else ["diffusion_pytorch_model.safetensors"]) state = {} for name in files: state.update(load_file(str(wan_dir / name))) config = json.loads((wan_dir / "config.json").read_text()) name = "Wan2.1-T2V-14B" if config["dim"] == 5120 else "Wan2.1-T2V-1.3B" return build_wan(name, state, causal=causal, dtype=dtype, device=device)