The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Fix/tweak pinned memory accounting (#13221)
rattus8d723d2caa95
diff --git a/comfy/model_management.py b/comfy/model_management.pyindex ce079cf2f..0eebf1ded 100644--- a/comfy/model_management.py+++ b/comfy/model_management.py@@ -1326,9 +1326,9 @@ MAX_PINNED_MEMORY = -1if not args.disable_pinned_memory:if is_nvidia() or is_amd():if WINDOWS:- MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.45 # Windows limit is apparently 50%+ MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.40 # Windows limit is apparently 50%else:- MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.95+ MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.90logging.info("Enabled pinned memory {}".format(MAX_PINNED_MEMORY // (1024 * 1024)))PINNING_ALLOWED_TYPES = set(["Tensor", "Parameter", "QuantizedTensor"])@@ -1403,8 +1403,6 @@ def unpin_memory(tensor):if torch.cuda.cudart().cudaHostUnregister(ptr) == 0:TOTAL_PINNED_MEMORY -= PINNED_MEMORY.pop(ptr)- if len(PINNED_MEMORY) == 0:- TOTAL_PINNED_MEMORY = 0return Trueelse:logging.warning("Unpin error.")