Merge 6f4d889053 into 7ac999bf30

Adds a new command-line argument `--total-ram` to limit the amount of
system RAM that ComfyUI considers available, allowing users to simulate
lower memory environments. This enables more predictable behavior when
testing or running on systems with limited resources.

Rationale:

I run Comfy inside a Docker container. Using `mem_limit` doesn't hide
total system RAM from Comfy, so OOM can occur easily. Cache pressure
limits cause frequent out-of-memory errors. Adding this flag allows
precise control over visible memory.

Signed-off-by: blob42 <contact@blob42.xyz>

comfy/cli_args.py

@@ -90,6 +90,7 @@ parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE"
 parser.add_argument("--oneapi-device-selector", type=str, default=None, metavar="SELECTOR_STRING", help="Sets the oneAPI device(s) this instance will use.")
 parser.add_argument("--disable-ipex-optimize", action="store_true", help="Disables ipex.optimize default when loading models with Intel's Extension for Pytorch.")
 parser.add_argument("--supports-fp8-compute", action="store_true", help="ComfyUI will act like if the device supports fp8 compute.")
+parser.add_argument("--total-ram", type=float, default=0, help="Maximum system RAM visible to comfy in GB (default 0: all)")
 
 class LatentPreviewMethod(enum.Enum):
     NoPreviews = "none"

comfy/clip_vision.py

@@ -66,6 +66,7 @@ class ClipVisionModel():
         outputs = Output()
         outputs["last_hidden_state"] = out[0].to(comfy.model_management.intermediate_device())
         outputs["image_embeds"] = out[2].to(comfy.model_management.intermediate_device())
+        outputs["image_sizes"] = [pixel_values.shape[1:]] * pixel_values.shape[0]
         if self.return_all_hidden_states:
             all_hs = out[1].to(comfy.model_management.intermediate_device())
             outputs["penultimate_hidden_states"] = all_hs[:, -2]

comfy/model_management.py

@@ -192,8 +192,12 @@ def get_total_memory(dev=None, torch_total_too=False):
     if dev is None:
         dev = get_torch_device()
 
-    if hasattr(dev, 'type') and (dev.type == 'cpu' or dev.type == 'mps'):
-        mem_total = psutil.virtual_memory().total
+    if hasattr(dev, "type") and (dev.type == "cpu" or dev.type == "mps"):
+        mem_total = 0
+        if args.total_ram != 0:
+            mem_total = args.total_ram * 1024 * 1024
+        else:
+            mem_total = psutil.virtual_memory().total
         mem_total_torch = mem_total
     else:
         if directml_enabled:
@@ -236,8 +240,15 @@ def mac_version():
         return None
 
 total_vram = get_total_memory(get_torch_device()) / (1024 * 1024)
-total_ram = psutil.virtual_memory().total / (1024 * 1024)
-logging.info("Total VRAM {:0.0f} MB, total RAM {:0.0f} MB".format(total_vram, total_ram))
+
+total_ram = 0
+if args.total_ram != 0:
+    total_ram = args.total_ram * (1024)  # arg in GB
+else:
+    total_ram = psutil.virtual_memory().total / (1024 * 1024)
+logging.info(
+    "Total VRAM {:0.0f} MB, total RAM {:0.0f} MB".format(total_vram, total_ram)
+)
 
 try:
     logging.info("pytorch version: {}".format(torch_version))

requirements.txt

@@ -1,5 +1,5 @@
 comfyui-frontend-package==1.36.14
-comfyui-workflow-templates==0.8.10
+comfyui-workflow-templates==0.8.11
 comfyui-embedded-docs==0.4.0
 torch
 torchsde