VictorMylle/Kohya-ss-sd-scripts
1
0
Fork 1
mirror of https://github.com/kohya-ss/sd-scripts.git synced 2026-04-10 06:54:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: apply_noise_offset の dtype 不一致を修正

Browse Source 
torch.randn のデフォルト float32 により latents が暗黙的にアップキャストされる問題を修正。
float32/CPU で生成後に latents の dtype/device へ変換する安全なパターンを採用。

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
Kohya S
2026-03-29 19:06:33 +09:00
parent 9562588ce0
commit bfdbf04059
1 changed files with 3 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
library/leco_train_util.py
View File

@@ -365,7 +365,9 @@ def encode_prompt_sdxl(tokenize_strategy, text_encoding_strategy, text_encoders,
def apply_noise_offset(latents: torch.Tensor, noise_offset: Optional[float]) -> torch.Tensor:
if noise_offset is None:
return latents
return latents + noise_offset * torch.randn((latents.shape[0], latents.shape[1], 1, 1), device=latents.device)
noise = torch.randn((latents.shape[0], latents.shape[1], 1, 1), dtype=torch.float32, device="cpu")
noise = noise.to(dtype=latents.dtype, device=latents.device)
return latents + noise_offset * noise
def get_initial_latents(scheduler, batch_size: int, height: int, width: int, n_prompts: int = 1) -> torch.Tensor: