"""Shared component loaders — load UMT5-XXL text encoder, AutoencoderKLWan VAE, and CLIP-ViT-H/14 image encoder ONCE at module load. Inject into every pipeline via `from_pretrained(..., text_encoder=, vae=, image_encoder=)`. Saves ~15 GB of duplicated weights vs loading per-pipeline. See RESEARCH.md §8.1. On ZeroGPU we read from the stitched Wan 2.2 T2V dir (mounted weights + bundled configs) so from_pretrained reads big binaries through the read-only volume mount (zero disk cost) and small JSONs from the bundled copy (working around the HF Volume small-file truncation bug). The Wan 2.2 T2V repo ships with the same UMT5-XXL + AutoencoderKLWan that every Wan pipeline shares, so loading them from this single stitched dir is correct. """ from __future__ import annotations import functools from typing import TYPE_CHECKING from utils.backend import detect if TYPE_CHECKING: import torch # noqa: F401 def _wan22_t2v_path() -> str: """Source for the shared UMT5 + VAE: the Wan 2.2 T2V CDN mirror on ZeroGPU (fast download — the volume mount is too slow), upstream as local fallback.""" if detect().is_zerogpu: return "techfreakworm/wan2.2-t2v-a14b" return "Wan-AI/Wan2.1-T2V-14B-Diffusers" @functools.lru_cache(maxsize=1) def text_encoder(): """UMT5-XXL — shared across every Wan pipeline (2.1 + 2.2).""" from transformers import UMT5EncoderModel backend = detect() return UMT5EncoderModel.from_pretrained( _wan22_t2v_path(), subfolder="text_encoder", torch_dtype=backend.dtype, ) @functools.lru_cache(maxsize=1) def vae(): """AutoencoderKLWan — same class handles Wan 2.1 (8×8×4) and Wan 2.2 (16×16×4) VAEs. The Wan 2.2 5B model uses a different VAE config but the same class. For TI2V-5B you'd load a separate VAE via `from_pretrained` against that repo's vae subfolder. """ from diffusers import AutoencoderKLWan backend = detect() instance = AutoencoderKLWan.from_pretrained( _wan22_t2v_path(), subfolder="vae", torch_dtype=backend.vae_dtype, ) # Memory savers — required at higher resolutions on every backend. instance.enable_tiling() instance.enable_slicing() return instance @functools.lru_cache(maxsize=1) def image_encoder(): """CLIP-ViT-H/14 — used by I2V, FLF2V, Animate.""" import torch from transformers import CLIPVisionModel return CLIPVisionModel.from_pretrained( "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers", subfolder="image_encoder", torch_dtype=torch.float32, # must stay fp32 per diffusers docs )