Try again with vae tiled decoding if regular fails because of OOM.

comfy/ldm/modules/attention.py

@@ -20,11 +20,6 @@ if model_management.xformers_enabled():
 import os
 _ATTN_PRECISION = os.environ.get("ATTN_PRECISION", "fp32")
 
-try:
-    OOM_EXCEPTION = torch.cuda.OutOfMemoryError
-except:
-    OOM_EXCEPTION = Exception
-
 def exists(val):
     return val is not None
 
@@ -312,7 +307,7 @@ class CrossAttentionDoggettx(nn.Module):
                     r1[:, i:end] = einsum('b i j, b j d -> b i d', s2, v)
                     del s2
                 break
-            except OOM_EXCEPTION as e:
+            except model_management.OOM_EXCEPTION as e:
                 if first_op_done == False:
                     torch.cuda.empty_cache()
                     torch.cuda.ipc_collect()

comfy/ldm/modules/diffusionmodules/model.py

@@ -13,11 +13,6 @@ if model_management.xformers_enabled():
     import xformers
     import xformers.ops
 
-try:
-    OOM_EXCEPTION = torch.cuda.OutOfMemoryError
-except:
-    OOM_EXCEPTION = Exception
-
 def get_timestep_embedding(timesteps, embedding_dim):
     """
     This matches the implementation in Denoising Diffusion Probabilistic Models:
@@ -221,7 +216,7 @@ class AttnBlock(nn.Module):
                     r1[:, :, i:end] = torch.bmm(v, s2)
                     del s2
                 break
-            except OOM_EXCEPTION as e:
+            except model_management.OOM_EXCEPTION as e:
                 steps *= 2
                 if steps > 128:
                     raise e

comfy/ldm/modules/sub_quadratic_attention.py

@@ -24,10 +24,7 @@ except ImportError:
 from torch import Tensor
 from typing import List
 
-try:
-    OOM_EXCEPTION = torch.cuda.OutOfMemoryError
-except:
-    OOM_EXCEPTION = Exception
+import model_management
 
 def dynamic_slice(
     x: Tensor,
@@ -161,7 +158,7 @@ def _get_attention_scores_no_kv_chunking(
     try:
         attn_probs = attn_scores.softmax(dim=-1)
         del attn_scores
-    except OOM_EXCEPTION:
+    except model_management.OOM_EXCEPTION:
         print("ran out of memory while running softmax in  _get_attention_scores_no_kv_chunking, trying slower in place softmax instead")
         attn_scores -= attn_scores.max(dim=-1, keepdim=True).values
         torch.exp(attn_scores, out=attn_scores)