ComfyUI crashes due to invalid or unsupported data type Float8_e4m3fn.

[Benix-laj]

So I've just installed ComfyUI and ComfyUI manager, and downloaded resources for the Flux fp8 model. When I try to queue a prompt the miniconda terminal crashes due to invalid or unsupported data type Float8_e4m3fn so the process stops, and I get no image. I've got an AMD card (6800 XT) . I don't really understand any of the things I'm doing, I'm just trying to get the AI model running :/. It gave the same error without the manager too, so the installation of the manager shouldn't be a problem.

I installed ComfyUI based on this tutorial (miniconda, git, directml):
https://www.youtube.com/watch?v=8rB7RqKvU5U

Here is the Flux workflow I use:

Here's the list of links of the resources I downloaded for flux:
VAE: https://huggingface.co/black-forest-labs/FLUX.1-schnell/blob/main/ae.safetensors
Text encoder: https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/t5xxl_fp8_e4m3fn.safetensors
Clip: https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/clip_l.safetensors
Flux: https://huggingface.co/Comfy-Org/flux1-schnell/blob/main/flux1-schnell-fp8.safetensors

C:\Users\nyirc>call conda activate comfyui
[START] Security scan
[DONE] Security scan

ComfyUI-Manager: installing dependencies done.

** ComfyUI startup time: 2024-10-23 14:22:15.895454
** Platform: Windows
** Python version: 3.10.12 | packaged by Anaconda, Inc. | (main, Jul 5 2023, 19:01:18) [MSC v.1916 64 bit (AMD64)]
** Python executable: D:\Programs\Miniconda\envs\comfyui\python.exe
** ComfyUI Path: C:\Users\nyirc\ComfyUI
** Log path: C:\Users\nyirc\ComfyUI\comfyui.log

Prestartup times for custom nodes:
0.8 seconds: C:\Users\nyirc\ComfyUI\custom_nodes\ComfyUI-Manager

Using directml with device:
Total VRAM 1024 MB, total RAM 32544 MB
pytorch version: 2.4.1+cpu
Set vram state to: NORMAL_VRAM
Device: privateuseone
Using sub quadratic optimization for cross attention, if you have memory or speed issues try using: --use-split-cross-attention
[Prompt Server] web root: C:\Users\nyirc\ComfyUI\web
D:\Programs\Miniconda\envs\comfyui\lib\site-packages\kornia\feature\lightglue.py:44: FutureWarning: torch.cuda.amp.custom_fwd(args...) is deprecated. Please use torch.amp.custom_fwd(args..., device_type='cuda') instead.
@torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)

Loading: ComfyUI-Manager (V2.51.8)

ComfyUI Revision: 2781 [915fdb5] | Released on '2024-10-22'

Import times for custom nodes:
0.0 seconds: C:\Users\nyirc\ComfyUI\custom_nodes\websocket_image_save.py
0.2 seconds: C:\Users\nyirc\ComfyUI\custom_nodes\ComfyUI-Manager

Starting server

To see the GUI go to: http://127.0.0.0:0000
[ComfyUI-Manager] default cache updated: https://raw.githubusercontent.com/ltdrdata/ComfyUI-Manager/main/alter-list.json
[ComfyUI-Manager] default cache updated: https://raw.githubusercontent.com/ltdrdata/ComfyUI-Manager/main/model-list.json
[ComfyUI-Manager] default cache updated: https://raw.githubusercontent.com/ltdrdata/ComfyUI-Manager/main/github-stats.json
[ComfyUI-Manager] default cache updated: https://raw.githubusercontent.com/ltdrdata/ComfyUI-Manager/main/custom-node-list.json
[ComfyUI-Manager] default cache updated: https://raw.githubusercontent.com/ltdrdata/ComfyUI-Manager/main/extension-node-map.json
FETCH DATA from: C:\Users\nyirc\ComfyUI\custom_nodes\ComfyUI-Manager\extension-node-map.json [DONE]
got prompt
Using split attention in VAE
Using split attention in VAE
model weight dtype torch.float8_e4m3fn, manual cast: torch.float32
model_type FLOW
Requested to load FluxClipModel_
Loading 1 new model
loaded completely 0.0 9319.23095703125 True
clip missing: ['text_projection.weight']
Requested to load Flux
Loading 1 new model
[F1023 14:27:43.000000000 dml_util.cc:118] Invalid or unsupported data type Float8_e4m3fn.

[a-mecham]

Was there ever a fix for this.
I am having the same trouble using a RX 7600XT, using ComfyUI in Miniconda environment with Python 3.12.7. and dirtectml.

My Error when using Flux1 Dev is [F1202 12:21:25.000000000 dml_util.cc:118] Invalid or unsupported data type Float8_e4m3fn.

[iamone0201]

I am facing the same issue, and found the root cause of it in a YouTube comment:

The torch-directml package supports up to PyTorch 2.3.1, and unfortunately, PyTorch 2.3.1 is outdated and does not support Float8_e4m3fn.

If you install latest torch, it is not supported by AMD GPUs and comfyUI won't start.
So we will have to wait till torch-directml package gets updated to support new PyTorch version.