Comfyui crashing in second re-run. First run completes fine. OOM

[sudeep333]

Earlier everything was working fine. I updated and it have started creating error.
I have 100GB RAM, 3060-12GB. Using Wan2.2
First run completes just fine and then i re-run the workflow and it crashes with the error below.

I have recently updated CUDA.

I made a fresh install and it is doing the same.

Pin error.
Pin error.
Pin error.
Pin error.
Pin error.
Pin error.
Pin error.
Pin error.
loaded partially; 5043.19 MB usable, 4320.84 MB loaded, 22932.13 MB offloaded, 840.08 MB buffer reserved, lowvram patches: 368
!!! Exception during processing !!! CUDA error: out of memory
Search for cudaErrorMemoryAllocation' in https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html for more information. CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1 Compile with TORCH_USE_CUDA_DSA` to enable device-side assertions.

Traceback (most recent call last):
File "E:\ComfyUI-5\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\execution.py", line 516, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI-5\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\execution.py", line 330, in get_output_data
return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI-5\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\execution.py", line 304, in _async_map_node_over_list
await process_inputs(input_dict, i)
File "E:\ComfyUI-5\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\execution.py", line 292, in process_inputs
result = f(**inputs)
^^^^^^^^^^^
File "E:\ComfyUI-5\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\nodes.py", line 1572, in sample
return common_ksampler(model, noise_seed, steps, cfg, sampler_name, scheduler, positive, negative, latent_image, denoise=denoise, disable_noise=disable_noise, start_step=start_at_step, last_step=end_at_step, force_full_denoise=force_full_denoise)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI-5\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\nodes.py", line 1505, in common_ksampler
samples = comfy.sample.sample(model, noise, steps, cfg, sampler_name, scheduler, positive, negative, latent_image,
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI-5\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\comfy\sample.py", line 60, in sample
samples = sampler.sample(noise, positive, negative, cfg=cfg, latent_image=latent_image, start_step=start_step, last_step=last_step, force_full_denoise=force_full_denoise, denoise_mask=noise_mask, sigmas=sigmas, callback=callback, disable_pbar=disable_pbar, seed=seed)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI-5\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\custom_nodes\ComfyUI-TiledDiffusion\utils.py", line 51, in KSampler_sample
return orig_fn(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI-5\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\comfy\samplers.py", line 1178, in sample
return sample(self.model, noise, positive, negative, cfg, self.device, sampler, sigmas, self.model_options, latent_image=latent_image, denoise_mask=denoise_mask, callback=callback, disable_pbar=disable_pbar, seed=seed)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI-5\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\comfy\samplers.py", line 1068, in sample
return cfg_guider.sample(noise, latent_image, sampler, sigmas, denoise_mask, callback, disable_pbar, seed)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI-5\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\comfy\samplers.py", line 1050, in sample
output = executor.execute(noise, latent_image, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI-5\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\comfy\patcher_extension.py", line 112, in execute
return self.original(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI-5\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\comfy\samplers.py", line 987, in outer_sample
noise = noise.to(device)
^^^^^^^^^^^^^^^^
torch.AcceleratorError: CUDA error: out of memory
Search for cudaErrorMemoryAllocation' in https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html for more information. CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1 Compile with TORCH_USE_CUDA_DSA` to enable device-side assertions.

[sudeep333]

I have disabled this node and everything worked fine. I think there is some conflict. Disable "Patch Sage Attention KJ" and then try.

Edit: Some workflows are still giving same error.

[rattus128]

The pin error into an async out of memory is suspicious. Can I get a task manager screenshot on the GPU page for your 3060 as it goes into the crash?

[rattus128]

BTW, with the pin errors, --disable-pinned-memory will probably work for you. Unfortuately, doing FP16 WAN on a 3060 you probably really need pinning for performance. If this does make it go away, let me know. Im interested if the default scaled fp8 models make your problems go away too.

[monnky]

i am also using RTX 3060 12GB , after new comfyui update 0.7.0, all content generations taking lot of time than before, and too many OOM crash than previous versions. i suspect memory leak problem.

updates should increase the speed , but in comfyUI , day by day its going !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

@comfyanonymous

---

13.01.2026 @comfyanonymous @sudeep333 @tester4488 @rattus128 @kijai

update 9.1 proudly i can say, this is really an amazing update, my 12 minute process went down to only 5 minutes,

again thank you very much @comfyanonymous Comfy-UI team and all other supporting developers.

[tester4488]

same, also swap file use is too high (windows), i roll back comfyui to commit 6ca3d5c for now and all work fine,
my .bat (conda python 3.13 rtx 4060):
python ComfyUI\main.py --fp8_e4m3fn-unet --fp8_e4m3fn-text-enc --auto-launch --disable-api-nodes --bf16-vae --disable-async-offload --enable-manager --enable-manager-legacy-ui --preview-method auto

[wood-123456]

老弟試試看調低功耗跟調整驅動版本

[sudeep333]

I made a fresh install today and only installed very basic nodes just to test everything. Comfyui is breaking again in the second run. This is what the error says.

!!! Exception during processing !!! CUDA error: out of memory
Search for cudaErrorMemoryAllocation' in https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html for more information. CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1 Compile with TORCH_USE_CUDA_DSA` to enable device-side assertions.

Traceback (most recent call last):
File "E:\ComfyUI-6\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\execution.py", line 516, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI-6\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\execution.py", line 330, in get_output_data
return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI-6\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\execution.py", line 304, in async_map_node_over_list
await process_inputs(input_dict, i)
File "E:\ComfyUI-6\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\execution.py", line 292, in process_inputs
result = f(**inputs)
File "E:\ComfyUI-6\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy_api\internal_init.py", line 149, in wrapped_func
return method(locked_class, **inputs)
File "E:\ComfyUI-6\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy_api\latest_io.py", line 1519, in EXECUTE_NORMALIZED
to_return = cls.execute(*args, **kwargs)
File "E:\ComfyUI-6\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy_extras\nodes_wan.py", line 47, in execute
concat_latent_image = vae.encode(image[:, :, :, :3])
File "E:\ComfyUI-6\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\sd.py", line 870, in encode
model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI-6\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\model_management.py", line 674, in load_models_gpu
free_memory(total_memory_required[device] * 1.1 + extra_mem, device)
~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI-6\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\model_management.py", line 606, in free_memory
if current_loaded_models[i].model_unload(memory_to_free):
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^
File "E:\ComfyUI-6\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\model_management.py", line 529, in model_unload
freed = self.model.partially_unload(self.model.offload_device, memory_to_free)
File "E:\ComfyUI-6\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\model_patcher.py", line 917, in partially_unload
m.to(device_to)
~~~~^^^^^^^^^^^
File "E:\ComfyUI-6\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1371, in to
return self._apply(convert)
~~~~~~~~~~~^^^^^^^^^
File "E:\ComfyUI-6\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 957, in _apply
param_applied = fn(param)
File "E:\ComfyUI-6\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1357, in convert
return t.to(
~~~~^
device,
^^^^^^^
dtype if t.is_floating_point() or t.is_complex() else None,
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
non_blocking,
^^^^^^^^^^^^^
)
^
torch.AcceleratorError: CUDA error: out of memory
Search for cudaErrorMemoryAllocation' in https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html for more information. CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1 Compile with TORCH_USE_CUDA_DSA` to enable device-side assertions.

Prompt executed in 8.27 seconds
Exception in thread Thread-13 (prompt_worker):
Traceback (most recent call last):
File "threading.py", line 1043, in _bootstrap_inner
File "threading.py", line 994, in run
File "E:\ComfyUI-6\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\main.py", line 270, in prompt_worker
comfy.model_management.soft_empty_cache()
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "E:\ComfyUI-6\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\model_management.py", line 1539, in soft_empty_cache
torch.cuda.empty_cache()
~~~~~~~~~~~~~~~~~~~~~~^^
File "E:\ComfyUI-6\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\cuda\memory.py", line 224, in empty_cache
torch._C._cuda_emptyCache()
~~~~~~~~~~~~~~~~~~~~~~~~~^^
torch.AcceleratorError: CUDA error: out of memory
Search for cudaErrorMemoryAllocation' in https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html for more information. CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1 Compile with TORCH_USE_CUDA_DSA` to enable device-side assertions.

[sudeep333]

After the crash. on second run.

[sudeep333]

Ok one more finding. I changed the default mode to fp8_e5m2 and it have started working fine. Can someone please explain it ?

Will i get degraded output beause of this change ?

[rattus128]

Ok one more finding. I changed the default mode to fp8_e5m2 and it have started working fine. Can someone please explain it ?

You are changing the precision of the loaded model to an fp8 scheme from fp16

Will i get degraded output bceause of this change ?

You might. Theres a much better way to do this though. Rather than start with the fp16 models you use here and cast them to fp8, redownload WAN as the scaled FP8 model. Its actually the default in the workflow template and its half the size of the ones you are using.

Looking at your screenshot you have large RAM usage before your crash, and then it refuses your pin for some reason.

The reason that change works, it it cuts the size of the loaded model in half and the amount you need to offload to RAM is cut in half accordingly. Everything gets better from a resource management point of view.

But downloading the scaled_fp8 models and using instead to make it even better, while using a much better and tested fp8 scheme. You should see a noticeable speed increase too.

You are offloading WAN fp16 on a 3060 which does work (from personal experience) but the offload volume is huge and stressing something on your system.

[rattus128]

Ok one more finding. I changed the default mode to fp8_e5m2 and it have started working fine. Can someone please explain it ?

You are changing the precision of the loaded model to an fp8 scheme from fp16

Will i get degraded output bceause of this change ?

You might. Theres a much better way to do this though. Rather than start with the fp16 models you use here and cast them to fp8, redownload WAN as the scaled FP8 model. Its actually the default in the workflow template and its half the size of the ones you are using.

Looking at your screenshot you have large RAM usage before your crash, and then it refuses your pin for some reason.

The reason that change works, it it cuts the size of the loaded model in half and the amount you need to offload to RAM is cut in half accordingly. Everything gets better from a resource management point of view.

But downloading the scaled_fp8 models and using instead to make it even better, while using a much better and tested fp8 scheme. You should see a noticeable speed increase too.

You are offloading WAN fp16 on a 3060 which does work (from personal experience) but the offload volume is huge and stressing something on your system.

Ok I see you have a 3rd party WAN variant. Its light2xv though. You could try scaled fp8 with light2xv as lora (which is the template).

Does --disable-pinned-memory work for you.

Also try --disable-smart-memory.

[sudeep333]

@rattus128 : Thank you for the valuable review and advice. I will do what you have suggested and will post the result. Thank you. Have a nice day.

[psychananaz]

Also try --disable-smart-memory.

This has been my only solution since I got this very same issue a few weeks back.
I would recommend anyone tries that first.