mirror of
https://github.com/Comfy-Org/desktop.git
synced 2026-11-04 23:47:43 +00:00
[GH-ISSUE #1476] Error: FP8 _scaled_mm failed, falling back to dequantization: Both scale_a and scale_b must be float (fp32) tensors. #425
Labels
No labels
area: desktop-app
Bounty
bug
bug-cop:answered
bug-cop:response-received
component: installer
component: preload
custom node
dependencies
Design
documentation
duplicate
enhancement
expected
frontend
good first issue
installation
Low-Priority
performance
Planned
platform: cross-platform
platform: windows
Potential Bug
Potential Bug
pull-request
question
refactor
Release
stale
User Support
V1
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
Comfy-Org/desktop#425
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @ironcloudy on GitHub (Dec 11, 2025).
Original GitHub issue: https://github.com/Comfy-Org/desktop/issues/1476
App Version
Desktop v0.6.0
ComfyUI v0.4.0
Expected Behavior
Using Model fluxmania_kreamania.safetensors from https://civitai.com/models/778691
was working with Flux workflow.
Actual Behavior
Getting Error:
SamplerCustomAdvanced
FP8 _scaled_mm failed, falling back to dequantization: Both scale_a and scale_b must be float (fp32) tensors.
This happens in ComfyUI v0.4.0.
Earlier version v.0.3.x worked.
Steps to Reproduce
Simple Flux workflow.
Debug Logs
Browser Logs
No response
Settings JSON
No response
Other
No response
┆Issue is synchronized with this Notion page by Unito
@dosubot[bot] commented on GitHub (Dec 11, 2025):
The error means the FP8 quantized model expects its scale tensors to be float32, but something in your setup is providing a different type. This usually happens if the model files or workflow nodes are mismatched, or if there was a change in dependencies (like PyTorch or ComfyUI) that made type checks stricter in v0.4.0.
For Flux FP8 models, you need to make sure you have the correct model files: the diffusion model (e.g.,
fluxmania_kreamania.safetensors), the VAE (ae.safetensorsfrom the official Flux releases), and both text encoders (clip_l.safetensorsandt5xxl_fp8_e4m3fn.safetensors). Place them in the right folders and double-check that your workflow nodes reference them correctly. Using the wrong VAE or CLIP model, or missing a required file, can cause this error or similar ones during sampling or generation [docs].If your setup worked in v0.3.x but not in v0.4.0, it's likely due to stricter type checks or changes in how models are loaded. Some users have also run into issues because of deprecated node inputs or workflow changes between versions—updating your workflow to match the latest node requirements can help [troubleshooting].
If you still see the error after verifying your models and workflow, try re-downloading the model files from trusted sources, resetting your Python environment, and making sure your PyTorch and CUDA versions match your hardware. Advanced workflows like Flux sometimes require features or fixes only available in the ComfyUI Nightly builds, not just the Desktop stable release [issue thread].
If you use custom nodes, try disabling them to rule out conflicts. Also, check your system memory if you experience hangs or "Reconnecting" errors during execution.
To reply, just mention @dosu.
Docs are dead. Just use Dosu.
@ironcloudy commented on GitHub (Dec 11, 2025):
The model is fp16.
I've verified vae, text encoders, and re-downloaded the model.
The workflow is a simple Flux.1 text to image workflow, nothing fancy.