Merge upstream/master, keep local README.md

comfy/ldm/lumina/model.py

@@ -491,7 +491,8 @@ class NextDiT(nn.Module):
                 for layer_id in range(n_layers)
             ]
         )
-        self.norm_final = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        # This norm final is in the lumina 2.0 code but isn't actually used for anything.
+        # self.norm_final = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
         self.final_layer = FinalLayer(dim, patch_size, self.out_channels, z_image_modulation=z_image_modulation, operation_settings=operation_settings)
 
         if self.pad_tokens_multiple is not None:

comfy/model_management.py

@@ -1126,6 +1126,16 @@ if not args.disable_pinned_memory:
 
 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