Add a command line argument to enable backend:cudaMallocAsync

comfy/cli_args.py

@@ -40,6 +40,7 @@ def __call__(self, parser, namespace, values, option_string=None):
 parser.add_argument("--output-directory", type=str, default=None, help="Set the ComfyUI output directory.")
 parser.add_argument("--auto-launch", action="store_true", help="Automatically launch ComfyUI in the default browser.")
 parser.add_argument("--cuda-device", type=int, default=None, metavar="DEVICE_ID", help="Set the id of the cuda device this instance will use.")
+parser.add_argument("--cuda-malloc", action="store_true", help="Enable cudaMallocAsync.")
 parser.add_argument("--dont-upcast-attention", action="store_true", help="Disable upcasting of attention. Can boost speed but increase the chances of black images.")
 
 fp_group = parser.add_mutually_exclusive_group()

comfy/model_management.py

@@ -204,7 +204,7 @@ def is_nvidia():
 def get_torch_device_name(device):
     if hasattr(device, 'type'):
         if device.type == "cuda":
-            return "{} {}".format(device, torch.cuda.get_device_name(device))
+            return "{} {} : {}".format(device, torch.cuda.get_device_name(device), torch.cuda.get_allocator_backend())
         else:
             return "{}".format(device.type)
     else:

main.py

@@ -51,7 +51,6 @@ def execute_script(script_path):
 import gc
 
 from comfy.cli_args import args
-import comfy.utils
 
 if os.name == "nt":
     import logging
@@ -62,7 +61,16 @@ def execute_script(script_path):
         os.environ['CUDA_VISIBLE_DEVICES'] = str(args.cuda_device)
         print("Set cuda device to:", args.cuda_device)
 
+    if args.cuda_malloc:
+        env_var = os.environ.get('PYTORCH_CUDA_ALLOC_CONF', None)
+        if env_var is None:
+            env_var = "backend:cudaMallocAsync"
+        else:
+            env_var += ",backend:cudaMallocAsync"
+
+        os.environ['PYTORCH_CUDA_ALLOC_CONF'] = env_var
 
+import comfy.utils
 import yaml
 
 import execution