Dynamic VRAM unloading still doesn't seem to work on windows

[Myoko]

Custom Node Testing

• I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)

Expected Behavior

File "F:\Content\ComfyUI\execution.py", line 305, in _async_map_node_over_list
await process_inputs(input_dict, i)
File "F:\Content\ComfyUI\execution.py", line 293, in process_inputs
result = f(**inputs)
File "F:\Content\ComfyUI\nodes.py", line 969, in load_unet
model = comfy.sd.load_diffusion_model(unet_path, model_options=model_options)
File "F:\Content\ComfyUI\comfy\sd.py", line 1686, in load_diffusion_model
sd, metadata = comfy.utils.load_torch_file(unet_path, return_metadata=True)
~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "F:\Content\ComfyUI\comfy\utils.py", line 132, in load_torch_file
raise e
File "F:\Content\ComfyUI\comfy\utils.py", line 112, in load_torch_file
sd, metadata = load_safetensors(ckpt)
~~~~~~~~~~~~~~~~^^^^^^
File "F:\Content\ComfyUI\comfy\utils.py", line 100, in load_safetensors
sd[name] = data_area[start:end].view(_TYPES[info["dtype"]]).view(info["shape"])
~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: self.storage_offset() must be divisible by 4 to view Byte as Float (different element sizes), but got 595403

Prompt executed in 14.97 seconds

Actual Behavior

error

Steps to Reproduce

every times

Debug Logs

VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
Found quantization metadata version 1
Using MixedPrecisionOps for text encoder
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
Requested to load WanVAE
Model WanVAE prepared for dynamic VRAM loading. 484MB Staged. 0 patches attached.
FETCH ComfyRegistry Data: 70/123
Model WanVAE prepared for dynamic VRAM loading. 484MB Staged. 0 patches attached.
Requested to load QwenImageTEModel_
Model QwenImageTEModel_ prepared for dynamic VRAM loading. 7910MB Staged. 0 patches attached.
FETCH ComfyRegistry Data: 75/123
FETCH ComfyRegistry Data: 80/123
0 models unloaded.
Model WanVAE prepared for dynamic VRAM loading. 484MB Staged. 0 patches attached.
0 models unloaded.
Model WanVAE prepared for dynamic VRAM loading. 484MB Staged. 0 patches attached.
Model QwenImageTEModel_ prepared for dynamic VRAM loading. 7910MB Staged. 0 patches attached.
FETCH ComfyRegistry Data: 85/123
!!! Exception during processing !!! self.storage_offset() must be divisible by 4 to view Byte as Float (different element sizes), but got 595403
Traceback (most recent call last):
  File "F:\Content\ComfyUI\execution.py", line 527, 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 "F:\Content\ComfyUI\execution.py", line 331, 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 "F:\Content\ComfyUI\execution.py", line 305, in _async_map_node_over_list
    await process_inputs(input_dict, i)
  File "F:\Content\ComfyUI\execution.py", line 293, in process_inputs
    result = f(**inputs)
  File "F:\Content\ComfyUI\nodes.py", line 969, in load_unet
    model = comfy.sd.load_diffusion_model(unet_path, model_options=model_options)
  File "F:\Content\ComfyUI\comfy\sd.py", line 1686, in load_diffusion_model
    sd, metadata = comfy.utils.load_torch_file(unet_path, return_metadata=True)
                   ~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "F:\Content\ComfyUI\comfy\utils.py", line 132, in load_torch_file
    raise e
  File "F:\Content\ComfyUI\comfy\utils.py", line 112, in load_torch_file
    sd, metadata = load_safetensors(ckpt)
                   ~~~~~~~~~~~~~~~~^^^^^^
  File "F:\Content\ComfyUI\comfy\utils.py", line 100, in load_safetensors
    sd[name] = data_area[start:end].view(_TYPES[info["dtype"]]).view(info["shape"])
               ~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: self.storage_offset() must be divisible by 4 to view Byte as Float (different element sizes), but got 595403

Prompt executed in 14.97 seconds

Other

Trying to use dynamic VRAM test qwen edit 2511 but getting the same error every time.

[SaturMars]

SAME HERE!!!!

[tester4488]

windows rtx 4060
BSOD (blue screen of death)
last rows of log

[2026-02-03 05:05:26.141] Model VideoVAE prepared for dynamic VRAM loading. 4663MB Staged. 0 patches attached.
[2026-02-03 05:05:32.947] Prompt executed in 420.89 seconds
[2026-02-03 05:06:17.944] got prompt
[2026-02-03 05:06:18.298] SELECTED: input1
[2026-02-03 05:06:18.868] Model LTXAV prepared for dynamic VRAM loading. 13778MB Staged. 1370 patches attached.
[2026-02-03 05:13:26.942] 
100%|████████████████████████████████████████████████████████████████████████████████████| 6/6 [05:54<00:00, 59.26s/it]
100%|████████████████████████████████████████████████████████████████████████████████████| 6/6 [05:54<00:00, 59.13s/it]
[2026-02-03 05:13:27.063] SELECTED: input1
[2026-02-03 05:13:27.086] Processing VAE decode tile at row 0, col 0:
[2026-02-03 05:13:27.087]  Position: (0:5, 0:6)
[2026-02-03 05:13:27.087]  Size: 5x6
[2026-02-03 05:13:27.222] Model VideoVAE prepared for dynamic VRAM loading. 4663MB Staged. 0 patches attached.
[2026-02-03 05:13:34.337] Processing VAE decode tile at row 0, col 1:
[2026-02-03 05:13:34.338]  Position: (0:5, 3:9)
[2026-02-03 05:13:34.339]  Size: 5x6
[2026-02-03 05:13:34.693] Model VideoVAE prepared for dynamic VRAM loading. 4663MB Staged. 0 patches attached.
[2026-02-03 05:13:36.083] Processing VAE decode tile at row 0, col 2:
[2026-02-03 05:13:36.084]  Position: (0:5, 6:12)
[2026-02-03 05:13:36.084]  Size: 5x6
[2026-02-03 05:13:36.379] Model VideoVAE prepared for dynamic VRAM loading. 4663MB Staged. 0 patches attached.
[2026-02-03 05:13:37.520] Processing VAE decode tile at row 0, col 3:
[2026-02-03 05:13:37.520]  Position: (0:5, 9:18)
[2026-02-03 05:13:37.521]  Size: 5x9
[2026-02-03 05:13:37.813] Model VideoVAE prepared for dynamic VRAM loading. 4663MB Staged. 0 patches attached.

removed the --fast dynamic_vram flag, no problems

comfy ba5bf3f

[rattus128]

windows rtx 4060 BSOD (blue screen of death) last rows of log

[2026-02-03 05:05:26.141] Model VideoVAE prepared for dynamic VRAM loading. 4663MB Staged. 0 patches attached.
[2026-02-03 05:05:32.947] Prompt executed in 420.89 seconds
[2026-02-03 05:06:17.944] got prompt
[2026-02-03 05:06:18.298] SELECTED: input1
[2026-02-03 05:06:18.868] Model LTXAV prepared for dynamic VRAM loading. 13778MB Staged. 1370 patches attached.
[2026-02-03 05:13:26.942] 
100%|████████████████████████████████████████████████████████████████████████████████████| 6/6 [05:54<00:00, 59.26s/it]
100%|████████████████████████████████████████████████████████████████████████████████████| 6/6 [05:54<00:00, 59.13s/it]
[2026-02-03 05:13:27.063] SELECTED: input1
[2026-02-03 05:13:27.086] Processing VAE decode tile at row 0, col 0:
[2026-02-03 05:13:27.087]  Position: (0:5, 0:6)
[2026-02-03 05:13:27.087]  Size: 5x6
[2026-02-03 05:13:27.222] Model VideoVAE prepared for dynamic VRAM loading. 4663MB Staged. 0 patches attached.
[2026-02-03 05:13:34.337] Processing VAE decode tile at row 0, col 1:
[2026-02-03 05:13:34.338]  Position: (0:5, 3:9)
[2026-02-03 05:13:34.339]  Size: 5x6
[2026-02-03 05:13:34.693] Model VideoVAE prepared for dynamic VRAM loading. 4663MB Staged. 0 patches attached.
[2026-02-03 05:13:36.083] Processing VAE decode tile at row 0, col 2:
[2026-02-03 05:13:36.084]  Position: (0:5, 6:12)
[2026-02-03 05:13:36.084]  Size: 5x6
[2026-02-03 05:13:36.379] Model VideoVAE prepared for dynamic VRAM loading. 4663MB Staged. 0 patches attached.
[2026-02-03 05:13:37.520] Processing VAE decode tile at row 0, col 3:
[2026-02-03 05:13:37.520]  Position: (0:5, 9:18)
[2026-02-03 05:13:37.521]  Size: 5x9
[2026-02-03 05:13:37.813] Model VideoVAE prepared for dynamic VRAM loading. 4663MB Staged. 0 patches attached.

removed the --fast dynamic_vram flag, no problems

comfy ba5bf3f

Ok we need to isolate more variables here as LTX template does work for me on my 3060 with 32GB RAM. Hopefully its not a 40 series card thing. Is that a custom node tiler? You crash when you tile changes up in size which might be a clue. If you redimension your video so its all even tile sizes does the crash go away?

[rattus128]

### Custom Node Testing

* [x]  I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)

Expected Behavior

File "F:\Content\ComfyUI\execution.py", line 305, in _async_map_node_over_list await process_inputs(input_dict, i) File "F:\Content\ComfyUI\execution.py", line 293, in process_inputs result = f(**inputs) File "F:\Content\ComfyUI\nodes.py", line 969, in load_unet model = comfy.sd.load_diffusion_model(unet_path, model_options=model_options) File "F:\Content\ComfyUI\comfy\sd.py", line 1686, in load_diffusion_model sd, metadata = comfy.utils.load_torch_file(unet_path, return_metadata=True) ~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "F:\Content\ComfyUI\comfy\utils.py", line 132, in load_torch_file raise e File "F:\Content\ComfyUI\comfy\utils.py", line 112, in load_torch_file sd, metadata = load_safetensors(ckpt) ~~~~~~~~~~~~~~~~^^^^^^ File "F:\Content\ComfyUI\comfy\utils.py", line 100, in load_safetensors sd[name] = data_area[start:end].view(_TYPES[info["dtype"]]).view(info["shape"]) ~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^ RuntimeError: self.storage_offset() must be divisible by 4 to view Byte as Float (different element sizes), but got 595403

Prompt executed in 14.97 seconds

Actual Behavior

error

Steps to Reproduce

every times

Debug Logs

VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
Found quantization metadata version 1
Using MixedPrecisionOps for text encoder
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
Requested to load WanVAE
Model WanVAE prepared for dynamic VRAM loading. 484MB Staged. 0 patches attached.
FETCH ComfyRegistry Data: 70/123
Model WanVAE prepared for dynamic VRAM loading. 484MB Staged. 0 patches attached.
Requested to load QwenImageTEModel_
Model QwenImageTEModel_ prepared for dynamic VRAM loading. 7910MB Staged. 0 patches attached.
FETCH ComfyRegistry Data: 75/123
FETCH ComfyRegistry Data: 80/123
0 models unloaded.
Model WanVAE prepared for dynamic VRAM loading. 484MB Staged. 0 patches attached.
0 models unloaded.
Model WanVAE prepared for dynamic VRAM loading. 484MB Staged. 0 patches attached.
Model QwenImageTEModel_ prepared for dynamic VRAM loading. 7910MB Staged. 0 patches attached.
FETCH ComfyRegistry Data: 85/123
!!! Exception during processing !!! self.storage_offset() must be divisible by 4 to view Byte as Float (different element sizes), but got 595403
Traceback (most recent call last):
File "F:\Content\ComfyUI\execution.py", line 527, 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 "F:\Content\ComfyUI\execution.py", line 331, 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 "F:\Content\ComfyUI\execution.py", line 305, in _async_map_node_over_list
await process_inputs(input_dict, i)
File "F:\Content\ComfyUI\execution.py", line 293, in process_inputs
result = f(**inputs)
File "F:\Content\ComfyUI\nodes.py", line 969, in load_unet
model = comfy.sd.load_diffusion_model(unet_path, model_options=model_options)
File "F:\Content\ComfyUI\comfy\sd.py", line 1686, in load_diffusion_model
sd, metadata = comfy.utils.load_torch_file(unet_path, return_metadata=True)
~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "F:\Content\ComfyUI\comfy\utils.py", line 132, in load_torch_file
raise e
File "F:\Content\ComfyUI\comfy\utils.py", line 112, in load_torch_file
sd, metadata = load_safetensors(ckpt)
~~~~~~~~~~~~~~~~^^^^^^
File "F:\Content\ComfyUI\comfy\utils.py", line 100, in load_safetensors
sd[name] = data_area[start:end].view(_TYPES[info["dtype"]]).view(info["shape"])
~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: self.storage_offset() must be divisible by 4 to view Byte as Float (different element sizes), but got 595403

Prompt executed in 14.97 seconds

Other

Trying to use dynamic VRAM test qwen edit 2511 but getting the same error every time.

Which version of qwen edit 2511? This error is sensitive to the exact safetensors file you have. I couldnt reproduce with fp8_mixed, are you using bf16?

If you are using a 3rd party repack or derivitive let me know where to get it.

[Myoko]

### Custom Node Testing ``` * [x] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help) ```

Expected Behavior

File "F:\Content\ComfyUI\execution.py", line 305, in _async_map_node_over_list await process_inputs(input_dict, i) File "F:\Content\ComfyUI\execution.py", line 293, in process_inputs result = f(**inputs) File "F:\Content\ComfyUI\nodes.py", line 969, in load_unet model = comfy.sd.load_diffusion_model(unet_path, model_options=model_options) File "F:\Content\ComfyUI\comfy\sd.py", line 1686, in load_diffusion_model sd, metadata = comfy.utils.load_torch_file(unet_path, return_metadata=True) ~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "F:\Content\ComfyUI\comfy\utils.py", line 132, in load_torch_file raise e File "F:\Content\ComfyUI\comfy\utils.py", line 112, in load_torch_file sd, metadata = load_safetensors(ckpt) ~~~~~~~~~~~~~~~~^^^^^^ File "F:\Content\ComfyUI\comfy\utils.py", line 100, in load_safetensors sd[name] = data_area[start:end].view(_TYPES[info["dtype"]]).view(info["shape"]) ~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^ RuntimeError: self.storage_offset() must be divisible by 4 to view Byte as Float (different element sizes), but got 595403
Prompt executed in 14.97 seconds

Actual Behavior

error

Steps to Reproduce

every times

Debug Logs

VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
Found quantization metadata version 1
Using MixedPrecisionOps for text encoder
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
Requested to load WanVAE
Model WanVAE prepared for dynamic VRAM loading. 484MB Staged. 0 patches attached.
FETCH ComfyRegistry Data: 70/123
Model WanVAE prepared for dynamic VRAM loading. 484MB Staged. 0 patches attached.
Requested to load QwenImageTEModel_
Model QwenImageTEModel_ prepared for dynamic VRAM loading. 7910MB Staged. 0 patches attached.
FETCH ComfyRegistry Data: 75/123
FETCH ComfyRegistry Data: 80/123
0 models unloaded.
Model WanVAE prepared for dynamic VRAM loading. 484MB Staged. 0 patches attached.
0 models unloaded.
Model WanVAE prepared for dynamic VRAM loading. 484MB Staged. 0 patches attached.
Model QwenImageTEModel_ prepared for dynamic VRAM loading. 7910MB Staged. 0 patches attached.
FETCH ComfyRegistry Data: 85/123
!!! Exception during processing !!! self.storage_offset() must be divisible by 4 to view Byte as Float (different element sizes), but got 595403
Traceback (most recent call last):
File "F:\Content\ComfyUI\execution.py", line 527, 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 "F:\Content\ComfyUI\execution.py", line 331, 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 "F:\Content\ComfyUI\execution.py", line 305, in _async_map_node_over_list
await process_inputs(input_dict, i)
File "F:\Content\ComfyUI\execution.py", line 293, in process_inputs
result = f(**inputs)
File "F:\Content\ComfyUI\nodes.py", line 969, in load_unet
model = comfy.sd.load_diffusion_model(unet_path, model_options=model_options)
File "F:\Content\ComfyUI\comfy\sd.py", line 1686, in load_diffusion_model
sd, metadata = comfy.utils.load_torch_file(unet_path, return_metadata=True)

File "F:\Content\ComfyUI\comfy\utils.py", line 132, in load_torch_file
raise e
File "F:\Content\ComfyUI\comfy\utils.py", line 112, in load_torch_file
sd, metadata = load_safetensors(ckpt)
~~~~~~~~~~~~~~~~^^^^^^
File "F:\Content\ComfyUI\comfy\utils.py", line 100, in load_safetensors
sd[name] = data_area[start:end].view(_TYPES[info["dtype"]]).view(info["shape"])
~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: self.storage_offset() must be divisible by 4 to view Byte as Float (different element sizes), but got 595403
Prompt executed in 14.97 seconds
### Other
Trying to use dynamic VRAM test qwen edit 2511 but getting the same error every time.

Which version of qwen edit 2511? This error is sensitive to the exact safetensors file you have. I couldnt reproduce with fp8_mixed, are you using bf16?

If you are using a 3rd party repack or derivitive let me know where to get it.

Comfy's official 2511 mixed version
https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/blob/main/split_files/diffusion_models/qwen_image_edit_2511_fp8mixed.safetensors

[MrSeri0us]

Hi, I'm getting a similar error with WAN 2.2 using this checkpoint:

https://civitai.com/models/1981116?modelVersionId=2316905

self.storage_offset() must be divisible by 4 to view Byte as Float (different element sizes), but got 316294

no issue using wan2.2_i2v_high_noise_14B_fp16.safetensors

[tester4488]

Ok we need to isolate more variables here as LTX template does work for me on my 3060 with 32GB RAM. Hopefully its not a 40 series card thing. Is that a custom node tiler? You crash when you tile changes up in size which might be a clue. If you redimension your video so its all even tile sizes does the crash go away?

i'm using the ltxv vae tiled decode 4 4 3
checkpoint ltx-2-19b-dev-fp4.safetensors
python 3.13 torch 2.9.1 cuda 13 , 4060 8GB 32GB ram

bat:

call activate E:\env\v313

cd E:\SD\ComfyUI_windows_portable
python ComfyUI\main.py --fp8_e4m3fn-unet --fp8_e4m3fn-text-enc  --auto-launch --disable-api-nodes --bf16-vae  --enable-manager --enable-manager-legacy-ui  --preview-method auto  --front-end-version Comfy-Org/ComfyUI_frontend@1.36.14 

[jovan2009]

I don't use dynamic_vram beside trying it out because for me it is a mixed bag. In recent weeks I run mostly LTX 2 distilled workflows. I noticed on the plus side that it starts to render the first 8 steps quicker than the default (meaning without dynamic_vram enabled). But when it gets to the next stage (the 3 ksampler steps after upsampling the latent) I get a bad case of GPU running at "100%" but at room temperature (the classic thing that happens when you spill over the hardware VRAM in computer RAM) and it takes forever to finish (actually I usually don't let it finish, I just Ctrl+C and start the server again without dynamic_vram enabled). If it doesn't happen during ksampler stage 2 it happens during VAE decoding. So, for now dynamic_vram is not better than the default, unfortunately, it doesn't crash (usually) but it doesn't help me very much with my workflows (my experience with it is limited, I don't use it beside trying it out).

I am on Windows 11, RTX 5600 Ti 16 GB, 64 GB RAM, 128 GB pagefile, Pytorch 2.11.0 nightly, Python 3.14.2

[rattus128]

I don't use dynamic_vram beside trying it out because for me it is a mixed bag. In recent weeks I run mostly LTX 2 distilled workflows. I noticed on the plus side that it starts to render the first 8 steps quicker than the default (meaning without dynamic_vram enabled). But when it gets to the next stage (the 3 ksampler steps after upsampling the latent) I get a bad case of GPU running at "100%" but at room temperature (the classic thing that happens when you spill over the hardware VRAM in computer RAM) and it takes forever to finish (actually I usually don't let it finish, I just Ctrl+C and start the server again without dynamic_vram enabled). If it doesn't happen during ksampler stage 2 it happens during VAE decoding. So, for now dynamic_vram is not better than the default, unfortunately, it doesn't crash (usually) but it doesn't help me very much with my workflows (my experience with it is limited, I don't use it beside trying it out).

I am on Windows 11, RTX 5600 Ti 16 GB, 64 GB RAM, 128 GB pagefile, Pytorch 2.11.0 nightly, Python 3.14.2

The feature is early days and its behind an optional switch for this reason. Im fixing bugs everyday. We can work with your use case if you give us the full information as stopping spilled memory is a critical feature of this and if its causing you to spill when it didnt before that is something I can look at.

Can you try with --disable-pinned-memory and screenshot task manager on the GPU page and Memory when you thinks its spilling?

Are use using the TCC driver or WSL by any chance?

[rattus128]

@tester4488 I see a race condition that was introduced but subsequently fixed in other work that may explain your BSOD. Its not a sure thing but maybe the latest is in better shape.

[jovan2009]

The feature is early days and its behind an optional switch for this reason. Im fixing bugs everyday. We can work with your use case if you give us the full information as stopping spilled memory is a critical feature of this and if its causing you to spill when it didnt before that is something I can look at.

As I said, I didn't use dynamic_vram very much beside trying now and then, days apart, to check what happens on my system. My experience was that is makes spilling (which happens anyway, especially with large models like LTX 2) much detrimental in terms of slowdown and this with the very first run of the workflow, at second K-Sampler stage. By comparison I don't get that with the default settings unless I set an unusual large length or resolution and it happens mostly during VAE decoding, not during K-Sampler.

Can you try with --disable-pinned-memory and screenshot task manager on the GPU page and Memory when you thinks its spilling?

I will certainly make sure to test that. About that, I even remember strange thing happened at some point recent weeks, I got a repeating message in console regarding pined memory but I can't remember what exactly. I was something like

...
...
Pinned bla bla bla
Pinned bla bla bla
Pinned bla bla bla
Pinned bla bla bla
...
...

over and over again, in a loop, leaving ComfyUI server in limbo, not crashing but not doing anything else either. I had to Ctrl+C to exit the server.

Are use using the TCC driver or WSL by any chance?

No, TCC is something I just found out what it is by googling it in order to give you an answer. I generally use the latest available game ready nvidia driver at the moment (either whql or bug fixing release). And I never used WSL.

[ryamldess]

It seems like this could be related to #12335, the immediately preceding issue? I've noticed that this issue seems to be resolved by either explicitly enabling CUDA Malloc or disabling smart memory in Server-Config or via CLI flags if running from VS. HTH until there is an official patch/fix. VRAM appears to be explicitly/intentionally not unloaded with dynamic memory. I think the free_memory() flow probably just needs to be cleaned up a bit, there are multiple intersecting conditionals, and it should probably just free up VRAM every time, imho. @isxzkas-tech has a patch in that other issue, but idk what the status is with it.