298977887/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2025-01-11 02:15:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

Relax minimum ratio of weights loaded in memory on nvidia.

Browse Source 
This should make it possible to do higher res images/longer videos by
further offloading weights to CPU memory.

Please report an issue if this slows down things on your system.
This commit is contained in:
comfyanonymous 2024-12-22 03:06:37 -05:00
parent 601ff9e3db
commit 57f330caf9
1 changed files with 7 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
comfy/model_management.py
View File

@ -188,6 +188,11 @@ def is_nvidia():
return True return True
return False return False
MIN_WEIGHT_MEMORY_RATIO = 0.4
if is_nvidia():
MIN_WEIGHT_MEMORY_RATIO = 0.2
ENABLE_PYTORCH_ATTENTION = False ENABLE_PYTORCH_ATTENTION = False
if args.use_pytorch_cross_attention: if args.use_pytorch_cross_attention:
ENABLE_PYTORCH_ATTENTION = True ENABLE_PYTORCH_ATTENTION = True
@ -509,7 +514,8 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
model_size = loaded_model.model_memory_required(torch_dev) model_size = loaded_model.model_memory_required(torch_dev)
loaded_memory = loaded_model.model_loaded_memory() loaded_memory = loaded_model.model_loaded_memory()
current_free_mem = get_free_memory(torch_dev) + loaded_memory current_free_mem = get_free_memory(torch_dev) + loaded_memory
lowvram_model_memory = max(64 * 1024 * 1024, (current_free_mem - minimum_memory_required), min(current_free_mem * 0.4, current_free_mem - minimum_inference_memory()))
lowvram_model_memory = max(64 * 1024 * 1024, (current_free_mem - minimum_memory_required), min(current_free_mem * MIN_WEIGHT_MEMORY_RATIO, current_free_mem - minimum_inference_memory()))
lowvram_model_memory = max(0.1, lowvram_model_memory - loaded_memory) lowvram_model_memory = max(0.1, lowvram_model_memory - loaded_memory)
if model_size <= lowvram_model_memory: #only switch to lowvram if really necessary if model_size <= lowvram_model_memory: #only switch to lowvram if really necessary
lowvram_model_memory = 0 lowvram_model_memory = 0