qwen-image-edit progress it/s possible wrong

[phil2sat]

I guess it adds time:

[INFO ] ggml_extend.hpp:1682 - qwenvl2.5 offload params (7165.44 MB, 338 tensors) to runtime backend (ROCm0), taking 1.10s
[DEBUG] ggml_extend.hpp:1582 - qwenvl2.5 compute buffer size: 7.24 MB(VRAM)
[DEBUG] conditioner.hpp:1572 - computing condition graph completed, taking 1396 ms
[INFO ] stable-diffusion.cpp:2441 - get_learned_condition completed, taking 2761 ms
[INFO ] stable-diffusion.cpp:2466 - sampling using Euler method
[INFO ] stable-diffusion.cpp:2560 - generating image: 1/1 - seed 42
[INFO ] ggml_extend.hpp:1682 - qwen_image offload params (8537.73 MB, 1933 tensors) to runtime backend (ROCm0), taking 1.23s
[DEBUG] ggml_extend.hpp:1582 - qwen_image compute buffer size: 6416.78 MB(VRAM)
|=================> | 7/20 - 925.00s/it

|====================> | 8/20 - 1057.73s/it

[leejet]

Could you please provide the full command line you ran (with all flags)? It will help me locate the issue more easily.

[wbruna]

I guess it adds time:

It's not clear what you consider 'wrong' here, but it looks like the displayed speed is showing the whole spent time, instead of the time per iteration?

@stduhpf made a similar remark on #834 , but I can't reproduce that behavior either. Which release are you running, and on which platform?

[stduhpf]

I'm building with MSVC (on Windows) on lastest master, an I can observe the problem both on Vulkan and ROCm builds. It's not just with Qwen image.

[phil2sat]

Could you please provide the full command line you ran (with all flags)? It will help me locate the issue more easily.

./sd -p "change hair to pink" -W 832 -H 1216 -t 8 --rng cuda -o /home/phil2sat/sd.cpp-webui/outputs/txt2img/194.png --diffusion-model /home/phil2sat/sd.cpp-webui/models/unet/qwen-rapid-aio-sfw-v5-Q3_K_S.gguf --vae /home/phil2sat/sd.cpp-webui/models/vae/qwen_image_vae.safetensors --qwen2vl /home/phil2sat/sd.cpp-webui/models/clip/Qwen2.5-VL-7B-Instruct-abliterated.Q8_0.gguf --offload-to-cpu --vae-tiling --color -v -r /tmp/gradio/f8b48cf061ccee7dd7913438af17994d99571d14dc3b09343800868292005dc2/ComfyUI_00113_.png --cfg-scale 1 --flow-shift 3 --sampling-method euler

linux rocm, 40a6a87

[wbruna]

Another data point: I don't see that issue running the master-326-40a6a87 Windows Vulkan binary release here under Wine. Tested SD1.5, SDXL, Qwen Edit and Flux.

[stduhpf]

Maybe it's related to the terminal? I'm using VSCode's built in terminal.

[stduhpf]

Hmm I have an update: Sometimes I see the correct value very briefly flashing before being overwritten by the accumulated value. Also sometimes the correct value ends up being displayed at the last step (but only the last one). Looks like a race condition or something of the sort. And I can confirm it also happens on Windows Terminal.

[phil2sat]

using normal terminal, so i guess its not vscode.
system archlinux with zsh