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.

Make maximum_batch_area take into account python2.0 attention function.

comfyanonymous6fc4917634d4
diff --git a/comfy/model_management.py b/comfy/model_management.py
index b0640d674..39df8d9a7 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -275,8 +275,17 @@ def xformers_enabled_vae():
return XFORMERS_ENABLED_VAE
def pytorch_attention_enabled():
+ global ENABLE_PYTORCH_ATTENTION
return ENABLE_PYTORCH_ATTENTION
+def pytorch_attention_flash_attention():
+ global ENABLE_PYTORCH_ATTENTION
+ if ENABLE_PYTORCH_ATTENTION:
+ #TODO: more reliable way of checking for flash attention?
+ if torch.version.cuda: #pytorch flash attention only works on Nvidia
+ return True
+ return False
+
def get_free_memory(dev=None, torch_free_too=False):
global xpu_available
global directml_enabled
@@ -312,9 +321,9 @@ def maximum_batch_area():
return 0
memory_free = get_free_memory() / (1024 * 1024)
- if xformers_enabled():
+ if xformers_enabled() or pytorch_attention_flash_attention():
#TODO: this needs to be tweaked
- area = 50 * memory_free
+ area = 20 * memory_free
else:
#TODO: this formula is because AMD sucks and has memory management issues which might be fixed in the future
area = ((memory_free - 1024) * 0.9) / (0.6)