The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Widen OOM_EXCEPTION to AcceleratorError form (#12835)
rattus535c16ce6e3d
diff --git a/comfy/ldm/modules/attention.py b/comfy/ldm/modules/attention.pyindex 10d051325..b193fe5e8 100644--- a/comfy/ldm/modules/attention.py+++ b/comfy/ldm/modules/attention.py@@ -372,7 +372,8 @@ def attention_split(q, k, v, heads, mask=None, attn_precision=None, skip_reshaper1[:, i:end] = einsum('b i j, b j d -> b i d', s2, v)del s2break- except model_management.OOM_EXCEPTION as e:+ except Exception as e:+ model_management.raise_non_oom(e)if first_op_done == False:model_management.soft_empty_cache(True)if cleared_cache == False:diff --git a/comfy/ldm/modules/diffusionmodules/model.py b/comfy/ldm/modules/diffusionmodules/model.pyindex 805592aa5..fcbaa074f 100644--- a/comfy/ldm/modules/diffusionmodules/model.py+++ b/comfy/ldm/modules/diffusionmodules/model.py@@ -258,7 +258,8 @@ def slice_attention(q, k, v):r1[:, :, i:end] = torch.bmm(v, s2)del s2break- except model_management.OOM_EXCEPTION as e:+ except Exception as e:+ model_management.raise_non_oom(e)model_management.soft_empty_cache(True)steps *= 2if steps > 128:@@ -314,7 +315,8 @@ def pytorch_attention(q, k, v):try:out = comfy.ops.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=False)out = out.transpose(2, 3).reshape(orig_shape)- except model_management.OOM_EXCEPTION:+ except Exception as e:+ model_management.raise_non_oom(e)logging.warning("scaled_dot_product_attention OOMed: switched to slice attention")oom_fallback = Trueif oom_fallback:diff --git a/comfy/ldm/modules/sub_quadratic_attention.py b/comfy/ldm/modules/sub_quadratic_attention.pyindex fab145f1c..f982afc2b 100644--- a/comfy/ldm/modules/sub_quadratic_attention.py+++ b/comfy/ldm/modules/sub_quadratic_attention.py@@ -169,7 +169,8 @@ def _get_attention_scores_no_kv_chunking(try:attn_probs = attn_scores.softmax(dim=-1)del attn_scores- except model_management.OOM_EXCEPTION:+ except Exception as e:+ model_management.raise_non_oom(e)logging.warning("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 # noqa: F821 attn_scores is not definedtorch.exp(attn_scores, out=attn_scores)diff --git a/comfy/model_management.py b/comfy/model_management.pyindex 07bc8ad67..81550c790 100644--- a/comfy/model_management.py+++ b/comfy/model_management.py@@ -270,6 +270,18 @@ try:except:OOM_EXCEPTION = Exception+def is_oom(e):+ if isinstance(e, OOM_EXCEPTION):+ return True+ if isinstance(e, torch.AcceleratorError) and getattr(e, 'error_code', None) == 2:+ discard_cuda_async_error()+ return True+ return False++def raise_non_oom(e):+ if not is_oom(e):+ raise e+XFORMERS_VERSION = ""XFORMERS_ENABLED_VAE = Trueif args.disable_xformers:diff --git a/comfy/sd.py b/comfy/sd.pyindex 888ef1e77..adcd67767 100644--- a/comfy/sd.py+++ b/comfy/sd.py@@ -954,7 +954,8 @@ class VAE:if pixel_samples is None:pixel_samples = torch.empty((samples_in.shape[0],) + tuple(out.shape[1:]), device=self.output_device)pixel_samples[x:x+batch_number] = out- except model_management.OOM_EXCEPTION:+ except Exception as e:+ model_management.raise_non_oom(e)logging.warning("Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.")#NOTE: We don't know what tensors were allocated to stack variables at the time of the#exception and the exception itself refs them all until we get out of this except block.@@ -1029,7 +1030,8 @@ class VAE:samples = torch.empty((pixel_samples.shape[0],) + tuple(out.shape[1:]), device=self.output_device)samples[x:x + batch_number] = out- except model_management.OOM_EXCEPTION:+ except Exception as e:+ model_management.raise_non_oom(e)logging.warning("Warning: Ran out of memory when regular VAE encoding, retrying with tiled VAE encoding.")#NOTE: We don't know what tensors were allocated to stack variables at the time of the#exception and the exception itself refs them all until we get out of this except block.diff --git a/comfy_extras/nodes_upscale_model.py b/comfy_extras/nodes_upscale_model.pyindex 97b9e948d..db4f9d231 100644--- a/comfy_extras/nodes_upscale_model.py+++ b/comfy_extras/nodes_upscale_model.py@@ -86,7 +86,8 @@ class ImageUpscaleWithModel(io.ComfyNode):pbar = comfy.utils.ProgressBar(steps)s = comfy.utils.tiled_scale(in_img, lambda a: upscale_model(a), tile_x=tile, tile_y=tile, overlap=overlap, upscale_amount=upscale_model.scale, pbar=pbar)oom = False- except model_management.OOM_EXCEPTION as e:+ except Exception as e:+ model_management.raise_non_oom(e)tile //= 2if tile < 128:raise ediff --git a/execution.py b/execution.pyindex 7ccdbf93e..a7791efed 100644--- a/execution.py+++ b/execution.py@@ -612,7 +612,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed,logging.error(traceback.format_exc())tips = ""- if isinstance(ex, comfy.model_management.OOM_EXCEPTION):+ if comfy.model_management.is_oom(ex):tips = "This error means you ran out of memory on your GPU.\n\nTIPS: If the workflow worked before you might have accidentally set the batch_size to a large number."logging.info("Memory summary: {}".format(comfy.model_management.debug_memory_summary()))logging.error("Got an OOM, unloading all loaded models.")