From e238d2eaea8f6f831eacffa47754b7976cd49344 Mon Sep 17 00:00:00 2001
From: AI_Architect <admin@local.host>
Date: Tue, 23 Jun 2026 09:52:03 +0800
Subject: [PATCH] =?UTF-8?q?=E4=BF=AE=E5=A4=8DBUG?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 nodes_wan.py    | 29 +++++++++++++++++++++--------
 nodes_wan_t5.py |  3 ++-
 2 files changed, 23 insertions(+), 9 deletions(-)

diff --git a/nodes_wan.py b/nodes_wan.py
index a142aab..6b5a204 100644
--- a/nodes_wan.py
+++ b/nodes_wan.py
@@ -205,9 +205,13 @@ class XB_WanSampler:
         kw.update(extra)
         r = _wan_import("nodes_sampler").WanVideoSampler().process(**kw)
         if cleanup == "单次缓存清理":
-            torch.cuda.synchronize(); torch.cuda.empty_cache()
+            torch.cuda.synchronize()
+            torch.cuda.empty_cache()
         elif cleanup == "双次缓存清理":
-            torch.cuda.synchronize(); mm.soft_empty_cache(); torch.cuda.empty_cache(); gc.collect()
+            torch.cuda.synchronize()
+            mm.soft_empty_cache()
+            torch.cuda.empty_cache()
+            gc.collect()
         return r
 
 
@@ -262,8 +266,10 @@ class XB_WanVAELoader:
         mod = _wan_import("wanvideo.wan_video_vae")
         cls = mod.WanVideoVAE38 if sd["model.conv2.weight"].shape[0] == 48 else mod.WanVideoVAE
         vae = cls(dtype=dtype, pruning_rate=pruning_rate, cpu_cache=use_cpu_cache)
-        vae.load_state_dict(sd); del sd
-        vae.eval(); vae.to(dtype=dtype)
+        vae.load_state_dict(sd)
+        del sd
+        vae.eval()
+        vae.to(dtype=dtype)
         print(f"  [XB_WanVAELoader] {model_name} | {precision} | pruning={pruning_rate}")
         return (vae,)
 
@@ -293,12 +299,19 @@ class XB_WanDecode:
             vae=vae, samples=samples, enable_vae_tiling=enable_vae_tiling,
             tile_x=tile_x, tile_y=tile_y, tile_stride_x=tile_stride_x, tile_stride_y=tile_stride_y)
         if cleanup == "单次缓存清理":
-            torch.cuda.synchronize(); torch.cuda.empty_cache()
+            torch.cuda.synchronize()
+            torch.cuda.empty_cache()
         elif cleanup == "双次缓存清理":
-            torch.cuda.synchronize(); mm.soft_empty_cache(); torch.cuda.empty_cache(); gc.collect()
+            torch.cuda.synchronize()
+            mm.soft_empty_cache()
+            torch.cuda.empty_cache()
+            gc.collect()
         elif cleanup == "卸载显存模型":
-            mm.unload_all_models(); mm.soft_empty_cache()
-            torch.cuda.synchronize(); torch.cuda.empty_cache(); gc.collect()
+            mm.unload_all_models()
+            mm.soft_empty_cache()
+            torch.cuda.synchronize()
+            torch.cuda.empty_cache()
+            gc.collect()
         return r
 
 
diff --git a/nodes_wan_t5.py b/nodes_wan_t5.py
index d2d2b9b..c5f28aa 100644
--- a/nodes_wan_t5.py
+++ b/nodes_wan_t5.py
@@ -139,7 +139,8 @@ class XB_WanT5Loader:
         if quantization == "disabled":
             for v in sd.values():
                 if isinstance(v, torch.Tensor) and v.dtype == torch.float8_e4m3fn:
-                    quantization = "fp8_e4m3fn"; break
+                    quantization = "fp8_e4m3fn"
+                    break
         if "token_embedding.weight" not in sd and "shared.weight" not in sd:
             raise ValueError("Invalid T5 model, expected 'umt5-xxl' format")
         sd = _convert_t5_keys(sd)