mm: guard against double pin and unpin explicitly (comfyanonymous#10672)

As commented, if you let cuda be the one to detect double pin/unpinning it actually creates an asyc GPU error.

Author: rattus128

comfy/model_management.py

@@ -1103,6 +1103,12 @@ def pin_memory(tensor):
  if not is_device_cpu(tensor.device):
  return False
 
+ if tensor.is_pinned():
+ #NOTE: Cuda does detect when a tensor is already pinned and would
+ #error below, but there are proven cases where this also queues an error
+ #on the GPU async. So dont trust the CUDA API and guard here
+ return False
+
  size = tensor.numel() * tensor.element_size()
  if (TOTAL_PINNED_MEMORY + size) > MAX_PINNED_MEMORY:
  return False
@@ -1123,6 +1129,12 @@ def unpin_memory(tensor):
  if not is_device_cpu(tensor.device):
  return False
 
+ if not tensor.is_pinned():
+ #NOTE: Cuda does detect when a tensor is already pinned and would
+ #error below, but there are proven cases where this also queues an error
+ #on the GPU async. So dont trust the CUDA API and guard here
+ return False
+
  ptr = tensor.data_ptr()
  if torch.cuda.cudart().cudaHostUnregister(ptr) == 0:
  TOTAL_PINNED_MEMORY -= PINNED_MEMORY.pop(ptr)