[pull] master from comfyanonymous:master

comfy/model_management.py

@@ -1126,6 +1126,16 @@ def cast_to_device(tensor, device, dtype, copy=False):
 
 PINNING_ALLOWED_TYPES = set(["Parameter", "QuantizedTensor"])
 
+def discard_cuda_async_error():
+    try:
+        a = torch.tensor([1], dtype=torch.uint8, device=get_torch_device())
+        b = torch.tensor([1], dtype=torch.uint8, device=get_torch_device())
+        _ = a + b
+        torch.cuda.synchronize()
+    except torch.AcceleratorError:
+        #Dump it! We already know about it from the synchronous return
+        pass
+
 def pin_memory(tensor):
     global TOTAL_PINNED_MEMORY
     if MAX_PINNED_MEMORY <= 0:
@@ -1158,6 +1168,9 @@ def pin_memory(tensor):
         PINNED_MEMORY[ptr] = size
         TOTAL_PINNED_MEMORY += size
         return True
+    else:
+        logging.warning("Pin error.")
+        discard_cuda_async_error()
 
     return False
 
@@ -1186,6 +1199,9 @@ def unpin_memory(tensor):
         if len(PINNED_MEMORY) == 0:
             TOTAL_PINNED_MEMORY = 0
         return True
+    else:
+        logging.warning("Unpin error.")
+        discard_cuda_async_error()
 
     return False
 

comfy_extras/nodes_dataset.py

@@ -667,16 +667,19 @@ class ResizeImagesByLongerEdgeNode(ImageProcessingNode):
 
     @classmethod
     def _process(cls, image, longer_edge):
-        img = tensor_to_pil(image)
-        w, h = img.size
-        if w > h:
-            new_w = longer_edge
-            new_h = int(h * (longer_edge / w))
-        else:
-            new_h = longer_edge
-            new_w = int(w * (longer_edge / h))
-        img = img.resize((new_w, new_h), Image.Resampling.LANCZOS)
-        return pil_to_tensor(img)
+        resized_images = []
+        for image_i in image:
+            img = tensor_to_pil(image_i)
+            w, h = img.size
+            if w > h:
+                new_w = longer_edge
+                new_h = int(h * (longer_edge / w))
+            else:
+                new_h = longer_edge
+                new_w = int(w * (longer_edge / h))
+            img = img.resize((new_w, new_h), Image.Resampling.LANCZOS)
+            resized_images.append(pil_to_tensor(img))
+        return torch.cat(resized_images, dim=0)
 
 
 class CenterCropImagesNode(ImageProcessingNode):