Got unsupported ScalarType Float8_e4m3fn

[itswhateverman]

I assume the following is indicating we can't start with an fp8 model and need fp16? Could the node upcast first to a useable format? I know it may not be ideal but it may still have utility.

---

Error Details

• Node Type: Mixed Cast Flux Loader
• Exception Type: TypeError

• Exception Message: Got unsupported ScalarType Float8_e4m3fn

  Stack Trace

  
  File "X:\comfy2\ComfyUI_windows_portable\ComfyUI\execution.py", line 317, in execute
  output_data, output_ui, has_subgraph = get_output_data(obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
                                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  
  File "X:\comfy2\ComfyUI_windows_portable\ComfyUI\execution.py", line 192, in get_output_data
  return_values = _map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  
  File "X:\comfy2\ComfyUI_windows_portable\ComfyUI\execution.py", line 169, in _map_node_over_list
  process_inputs(input_dict, i)
  
  File "X:\comfy2\ComfyUI_windows_portable\ComfyUI\execution.py", line 158, in process_inputs
  results.append(getattr(obj, func)(**inputs))
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  
  File "X:\comfy2\ComfyUI_windows_portable\ComfyUI\custom_nodes\cg-mixed-casting\mixed_gguf_node.py", line 178, in func
  sd = mixed_gguf_sd_loader(path, metadata=config)
       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  
  File "X:\comfy2\ComfyUI_windows_portable\ComfyUI\custom_nodes\cg-mixed-casting\mixed_gguf_node.py", line 130, in mixed_gguf_sd_loader
  qt = quants.quantize(tnsr.numpy(), qtype=qtype)
                       ^^^^^^^^^^^^

[itswhateverman]

the above replies are malware spam fyi

edit: they've been removed

[chrisgoringe]

Thanks for the report.

Try an update (git pull in the cg-mixed-casting directory) and see if it works now.

[itswhateverman]

After update, now errors on BFloat16:

I'm trying to load custom models such as jibMixFlux_v10. (fp8) I have this issue with the other custom models I have tried. I don't experience this using fp16 models which properly quantize.

Mixed Cast Flux Loader

Got unsupported ScalarType BFloat16

ComfyUI Error Report

Error Details

• Node Type: Mixed Cast Flux Loader
• Exception Type: TypeError

• Exception Message: Got unsupported ScalarType BFloat16

  Stack Trace

  
  File "X:\comfy2\ComfyUI_windows_portable\ComfyUI\execution.py", line 317, in execute
  output_data, output_ui, has_subgraph = get_output_data(obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
                                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  
  File "X:\comfy2\ComfyUI_windows_portable\ComfyUI\execution.py", line 192, in get_output_data
  return_values = _map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  
  File "X:\comfy2\ComfyUI_windows_portable\ComfyUI\execution.py", line 169, in _map_node_over_list
  process_inputs(input_dict, i)
  
  File "X:\comfy2\ComfyUI_windows_portable\ComfyUI\execution.py", line 158, in process_inputs
  results.append(getattr(obj, func)(**inputs))
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  
  File "X:\comfy2\ComfyUI_windows_portable\ComfyUI\custom_nodes\cg-mixed-casting\mixed_gguf_node.py", line 181, in func
  sd = mixed_gguf_sd_loader(path, metadata=config)
       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  
  File "X:\comfy2\ComfyUI_windows_portable\ComfyUI\custom_nodes\cg-mixed-casting\mixed_gguf_node.py", line 133, in mixed_gguf_sd_loader
  qt = quants.quantize(tnsr.to(torch.bfloat16).numpy(), qtype=qtype)
                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

[chrisgoringe]

If you go to the file mixed_gguf_node.py and find line 133 (the one shown in the error), change bfloat16 to float and it should work.

That change will be in the next push...

[chrisgoringe]

Should be fixed in the new version.

[itswhateverman]

yep thanks