error with --fast dynamic_vram

[tester4488]

with --fast dynamic_vram

Requested to load LTXAV
loaded partially; 3565.34 MB usable, 3302.78 MB loaded, 12060.69 MB offloaded, 262.55 MB buffer reserved, lowvram patches: 0
100%|████████████████████████████████████████████████████████████████████████████████████| 6/6 [02:48<00:00, 28.03s/it]
Unloaded partially: 845.81 MB freed, 2456.97 MB remains loaded, 262.72 MB buffer reserved, lowvram patches: 0
Requested to load VideoVAE
Model VideoVAE prepared for dynamic VRAM loading. 4663MB Staged. 0 patches attached.
!!! Exception during processing !!! The size of tensor a (3) must match the size of tensor b (28533760) at non-singleton dimension 4
Traceback (most recent call last):
  File "E:\SD\ComfyUI_windows_portable\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 "E:\SD\ComfyUI_windows_portable\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 "E:\SD\ComfyUI_windows_portable\ComfyUI\execution.py", line 305, in _async_map_node_over_list
    await process_inputs(input_dict, i)
  File "E:\SD\ComfyUI_windows_portable\ComfyUI\execution.py", line 293, in process_inputs
    result = f(**inputs)
  File "E:\SD\ComfyUI_windows_portable\ComfyUI\comfy_api\internal\__init__.py", line 149, in wrapped_func
    return method(locked_class, **inputs)
  File "E:\SD\ComfyUI_windows_portable\ComfyUI\comfy_api\latest\_io.py", line 1667, in EXECUTE_NORMALIZED
    to_return = cls.execute(*args, **kwargs)
  File "E:\SD\ComfyUI_windows_portable\ComfyUI\comfy_extras\nodes_lt.py", line 121, in execute
    t = vae.encode(encode_pixels)
  File "E:\SD\ComfyUI_windows_portable\ComfyUI\comfy\sd.py", line 998, in encode
    out = self.first_stage_model.encode(pixels_in).to(self.output_device).float()
          ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^
  File "E:\SD\ComfyUI_windows_portable\ComfyUI\comfy\ldm\lightricks\vae\causal_video_autoencoder.py", line 1167, in encode
    means, logvar = torch.chunk(self.encoder(x), 2, dim=1)
                                ~~~~~~~~~~~~^^^
  File "E:\env\v313\Lib\site-packages\torch\nn\modules\module.py", line 1775, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
  File "E:\env\v313\Lib\site-packages\torch\nn\modules\module.py", line 1786, in _call_impl
    return forward_call(*args, **kwargs)
  File "E:\SD\ComfyUI_windows_portable\ComfyUI\comfy\ldm\lightricks\vae\causal_video_autoencoder.py", line 285, in forward
    return self.forward_orig(*args, **kwargs)
           ~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
  File "E:\SD\ComfyUI_windows_portable\ComfyUI\comfy\ldm\lightricks\vae\causal_video_autoencoder.py", line 249, in forward_orig
    sample = self.conv_out(sample)
  File "E:\env\v313\Lib\site-packages\torch\nn\modules\module.py", line 1775, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
  File "E:\env\v313\Lib\site-packages\torch\nn\modules\module.py", line 1786, in _call_impl
    return forward_call(*args, **kwargs)
  File "E:\SD\ComfyUI_windows_portable\ComfyUI\comfy\ldm\lightricks\vae\causal_conv3d.py", line 72, in forward
    return self.conv(x) if x.shape[2] >= self.time_kernel_size else x[:, :, :0, :, :]
           ~~~~~~~~~^^^
  File "E:\env\v313\Lib\site-packages\torch\nn\modules\module.py", line 1775, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
  File "E:\env\v313\Lib\site-packages\torch\nn\modules\module.py", line 1786, in _call_impl
    return forward_call(*args, **kwargs)
  File "E:\SD\ComfyUI_windows_portable\ComfyUI\comfy\ops.py", line 414, in forward
    return self.forward_comfy_cast_weights(*args, **kwargs)
           ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
  File "E:\SD\ComfyUI_windows_portable\ComfyUI\comfy\ops.py", line 406, in forward_comfy_cast_weights
    weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
                                   ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "E:\SD\ComfyUI_windows_portable\ComfyUI\comfy\ops.py", line 207, in cast_bias_weight
    return cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype)
  File "E:\SD\ComfyUI_windows_portable\ComfyUI\comfy\ops.py", line 138, in cast_bias_weight_with_vbar
    post_cast.copy_(pre_cast)
    ~~~~~~~~~~~~~~~^^^^^^^^^^
RuntimeError: The size of tensor a (3) must match the size of tensor b (28533760) at non-singleton dimension 4

removing
--fast dynamic_vram
all work fine

Unloaded partially: 648.02 MB freed, 1188.03 MB remains loaded, 162.01 MB buffer reserved, lowvram patches: 0
Prompt executed in 00:13:17

[Myoko]

Error encountered when loading the model"self.storage_offset() must be divisible by 4 to view Byte as Float (different element sizes), but got 595403"

[rattus128]

Error encountered when loading the model"self.storage_offset() must be divisible by 4 to view Byte as Float (different element sizes), but got 595403"

can I get the workflow and full backtrace?

[rattus128]

@Myoko your bug is completely different to OPs. Can I get a link to your safetensors file that was loading when it crashed?

[rattus128]

reproduced on RTX3060 (NOTE the tile size varys based on VRAM to reproduce OPs exact conditions):

  File "/home/rattus/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/rattus/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/rattus/ComfyUI/comfy/ldm/lightricks/vae/causal_conv3d.py", line 72, in forward
    return self.conv(x) if x.shape[2] >= self.time_kernel_size else x[:, :, :0, :, :]
           ^^^^^^^^^^^^
  File "/home/rattus/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/rattus/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/rattus/ComfyUI/comfy/ops.py", line 414, in forward
    return self.forward_comfy_cast_weights(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/rattus/ComfyUI/comfy/ops.py", line 406, in forward_comfy_cast_weights
    weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
                                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/rattus/ComfyUI/comfy/ops.py", line 207, in cast_bias_weight
    return cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/rattus/ComfyUI/comfy/ops.py", line 138, in cast_bias_weight_with_vbar
    post_cast.copy_(pre_cast)
RuntimeError: The size of tensor a (3) must match the size of tensor b (1770496) at non-singleton dimension 4

Prompt executed in 1.51 seconds

[Myoko]

@Myoko 你的问题与 OP 的完全不同。我能获得那个崩溃时正在加载的 safetensors 文件的链接吗？

qwen edit 2511 https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/tree/main/split_files/diffusion_models

Model QwenImageTEModel_ prepared for dynamic VRAM loading. 7910MB Staged. 0 patches attached.
!!! 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

[tester4488]

reproduced on RTX3060 (NOTE the tile size varys based on VRAM to reproduce OPs exact conditions):

ok fixed with comfy 021ba20