wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-01-15 08:40:50 +08:00
Code Issues Actions 11 Packages Projects Releases Wiki Activity

Merge 1e8a9cef85 into 519c941165

Browse Source
This commit is contained in:
Angel Tsvetkov 2025-12-03 11:34:01 -05:00 committed by GitHub
commit 4ce5a1d16b
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 48 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

49
comfy/ldm/modules/diffusionmodules/model.py
View File

@ -9,6 +9,11 @@ from comfy import model_management
import comfy.ops
ops = comfy.ops.disable_weight_init
import inspect
import os
if 'vae_attention_counter' not in globals():
vae_attention_counter = {}
if model_management.xformers_enabled_vae():
import xformers
import xformers.ops
@ -336,7 +341,49 @@ def vae_attention():
logging.info("Using xformers attention in VAE")
return xformers_attention
elif model_management.pytorch_attention_enabled_vae():
logging.info("Using pytorch attention in VAE")
#Common causes for duplicate VAE loading:
#1.Different precision/dtype loading - ComfyUI might load the same VAE twice with different precisions (fp16/fp32) or for different devices (CPU/GPU)
#2.Encoder and Decoder initialization - VAEs have separate encoder and decoder components that might initialize attention separately
#3.Model switching or reinitialization - If you have workflows or settings that switch between models during startup
#4.Checkpoint with embedded VAE + separate VAE - You might have a checkpoint that includes a VAE, plus a separate VAE file
# Global counter for VAE attention initialization
def get_detailed_vae_info():
frame = inspect.currentframe()
try:
# Get model info
model_name = "Unknown model"
for f in inspect.getouterframes(frame):
local_vars = f.frame.f_locals
for var_name in ['model_path', 'vae_path', 'checkpoint_path', 'config_path']:
if var_name in local_vars and local_vars[var_name]:
model_name = os.path.basename(str(local_vars[var_name]))
break
if 'self' in local_vars:
obj = local_vars['self']
if hasattr(obj, 'model_path') and obj.model_path:
model_name = os.path.basename(obj.model_path)
break
if model_name != "Unknown model":
break
# Use the global counter
if model_name not in vae_attention_counter:
vae_attention_counter[model_name] = 0
vae_attention_counter[model_name] += 1
count = vae_attention_counter[model_name]
# Determine component based on call count
components = ["encoder", "decoder"]
component = components[min(count - 1, len(components) - 1)]
return model_name, component, count
finally:
del frame
model_name, component, count = get_detailed_vae_info()
logging.info(f"Using pytorch attention in VAE {component} ({count}/2) - Model: {model_name}")
return pytorch_attention
else:
logging.info("Using split attention in VAE")