Skip to content

[Bug upstream?] Crash (SCALE op fail) when running Flux at very high resolution on ROCm #948

@stduhpf

Description

@stduhpf

.\buildhip\bin\sd.exe --diffusion-model ..\ComfyUI\models\unet\\flux-mini-q8_0.gguf --vae ..\ComfyUI\models\vae\flux\ae.f16.gguf -p "a lovely cat"--cfg-scale 1 --sampling-method euler --steps 24 --t5xxl ..\ComfyUI\models\clip\t5\t5xxl_q8_0.gguf --clip_l ..\ComfyUI\models\clip\clip_l\clip_l.safetensors --preview proj -H 1856 -W 1856 --vae-tiling --vae-tile-size 64 -v

[...]
[DEBUG] stable-diffusion.cpp\ggml_extend.hpp:1600 - flux compute buffer size: 19805.33 MB(VRAM)
[ERROR] stable-diffusion.cpp\ggml_extend.hpp:75   - ggml_cuda_compute_forward: SCALE failed
[ERROR] stable-diffusion.cpp\ggml_extend.hpp:75   - ROCm error: invalid configuration argument
[ERROR] stable-diffusion.cpp\ggml_extend.hpp:75   -   current device: 0, in function ggml_cuda_compute_forward at [...]/stable-diffusion.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:2540
[ERROR] stable-diffusion.cpp\ggml_extend.hpp:75   -   err
[...]/stable-diffusion.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:88: ROCm error

On Windows 10, Rocm 6.2, Driver 25.Q3.1, RX 6800 (gfx1030)

Highest resolution I was able to generate without running into this issue is 1856x1792.

Of course enabling Flash attention fixes that (or maybe it happens at absurldy high resolutions)

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions