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.

Disable pinned memory automatically on integrated GPUs (#16746)

rattus4d981c9b09db
diff --git a/comfy/model_management.py b/comfy/model_management.py
index 4ea7cbae7..13b690bdb 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -1624,7 +1624,16 @@ def get_disk_swap_total():
logging.warning("Could not get amount of swap memory on system.")
return total
-if not args.disable_pinned_memory:
+def is_integrated_gpu():
+ device = get_torch_device()
+ return device.type == "cuda" and bool(torch.cuda.get_device_properties(device).is_integrated)
+
+DISABLE_PINNED_MEMORY = args.disable_pinned_memory
+if not DISABLE_PINNED_MEMORY and is_integrated_gpu():
+ # Integrated GPU VRAM is carved out of system RAM, so pinning host memory only takes RAM from the GPU.
+ DISABLE_PINNED_MEMORY = True
+
+if not DISABLE_PINNED_MEMORY:
if is_nvidia() or is_amd():
ram = get_total_memory(torch.device("cpu"))
if WINDOWS:
diff --git a/comfy/pinned_memory.py b/comfy/pinned_memory.py
index b3057c620..76374d2d5 100644
--- a/comfy/pinned_memory.py
+++ b/comfy/pinned_memory.py
@@ -7,8 +7,6 @@ import comfy_aimdo.host_buffer
import comfy_aimdo.torch
import torch
-from comfy.cli_args import args
-
def _add_to_bucket(module, module_pin, buckets, size, priority):
bucket = buckets.setdefault(size, [])
entry = [-priority, 0, module]
@@ -51,7 +49,7 @@ def get_pin(module, subset="weights"):
pins = module.__dict__.get("_pins")
module_pin = None if pins is None else pins.get(subset)
pin = None if module_pin is None else module_pin.get("pin")
- if pin is None or module_pin["registered"] or args.disable_pinned_memory:
+ if pin is None or module_pin["registered"] or comfy.model_management.DISABLE_PINNED_MEMORY:
return pin
_, _, stack_split, pinned_size, *_ = module._pin_state[subset]
@@ -70,7 +68,7 @@ def get_pin(module, subset="weights"):
def pin_memory(module, subset="weights", size=None):
pin_state = module._pin_state
- if args.disable_pinned_memory:
+ if comfy.model_management.DISABLE_PINNED_MEMORY:
return
pin = get_pin(module, subset)