Test / managers /vae_manager.py
EuuIia's picture
Create managers/vae_manager.py
8815ceb verified
raw
history blame
2.29 kB
# vae_manager.py — versão simples (beta 1.0)
# Responsável por decodificar latentes (B,C,T,H,W) → pixels (B,C,T,H',W') em [0,1].
import torch
import contextlib
class _SimpleVAEManager:
def __init__(self, pipeline=None, device=None, autocast_dtype=torch.float32):
"""
pipeline: objeto do LTX que expõe decode_latents(...) ou .vae.decode(...)
device: "cuda" ou "cpu" onde a decodificação deve ocorrer
autocast_dtype: dtype de autocast quando em CUDA (bf16/fp16/fp32)
"""
self.pipeline = pipeline
self.device = device or ("cuda" if torch.cuda.is_available() else "cpu")
self.autocast_dtype = autocast_dtype
def attach_pipeline(self, pipeline, device=None, autocast_dtype=None):
self.pipeline = pipeline
if device is not None:
self.device = device
if autocast_dtype is not None:
self.autocast_dtype = autocast_dtype
@torch.no_grad()
def decode(self, latents_5d: torch.Tensor) -> torch.Tensor:
"""
Decodifica todo o bloco 5D de uma vez, replicando o fluxo simples do deformes4D.
Retorna tensor de pixels 5D em [0,1] com shape (B,C,T,H',W').
"""
if self.pipeline is None:
raise RuntimeError("VAE Manager sem pipeline. Chame attach_pipeline primeiro.")
# Garante device correto
latents_5d = latents_5d.to(self.device, non_blocking=True)
ctx = torch.autocast(device_type="cuda", dtype=self.autocast_dtype) if self.device == "cuda" else contextlib.nullcontext()
with ctx:
if hasattr(self.pipeline, "decode_latents"):
pixels_5d = self.pipeline.decode_latents(latents_5d)
elif hasattr(self.pipeline, "vae") and hasattr(self.pipeline.vae, "decode"):
pixels_5d = self.pipeline.vae.decode(latents_5d)
else:
raise RuntimeError("Pipeline não expõe decode_latents nem vae.decode.")
# Normaliza para [0,1] se vier em [-1,1]
if pixels_5d.min() < 0:
pixels_5d = (pixels_5d.clamp(-1, 1) + 1.0) / 2.0
else:
pixels_5d = pixels_5d.clamp(0, 1)
return pixels_5d
# Singleton global de uso simples
vae_manager_singleton = _SimpleVAEManager()