Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

Add a command line argument to enable backend:cudaMallocAsync

comfyanonymous1679abd86d94
diff --git a/comfy/cli_args.py b/comfy/cli_args.py
index bef1868b9..9a388b03b 100644
--- a/comfy/cli_args.py
+++ b/comfy/cli_args.py
@@ -40,6 +40,7 @@ parser.add_argument("--extra-model-paths-config", type=str, default=None, metava
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()
diff --git a/comfy/model_management.py b/comfy/model_management.py
index 92c8ac842..69542cc37 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -204,7 +204,7 @@ print(f"Set vram state to: {vram_state.name}")
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:
diff --git a/main.py b/main.py
index 802e4bfe4..a22545573 100644
--- a/main.py
+++ b/main.py
@@ -51,7 +51,6 @@ import threading
import gc
from comfy.cli_args import args
-import comfy.utils
if os.name == "nt":
import logging
@@ -62,7 +61,16 @@ if __name__ == "__main__":
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