chaiNNer
chaiNNer copied to clipboard
joeyballentine
0.11.5 on macos Big Sur 11.6
In this version, the installation went smoothly and I downloaded the ncnn packages. but this time the model I am using on windows and I am sure it works (4x-UltraSharp-opt-fp16.bin ) gives a black screen. Anyone have experience with this before?
By the way, the reason for using ncnn is that the others are not using the gpu. pytorch is using only the cpu.
thank you
It seems like ncnn on macos is still broken unfortunately. It has something to do with vulkan, and I'm not sure what I can do to fix it. But I'll try
It seems like ncnn on macos is still broken unfortunately. It has something to do with vulkan, and I'm not sure what I can do to fix it. But I'll try
thanks a lot Joey
i try macbook pro m1 (Macos Monterey 12.5.1 ) work perfectly and very fast. but Onnx cannot be installed in this either.
Onnx does not have an arm64 (m1) version, therefore it is unable to be installed.
Also this is the first I've heard of the ncnn m1 version I just released working. I was last told it produced black outputs as well. So this is good news I guess :)
oh, Now I noticed that if it is more than 512x512 image size it gives a black screen.
Thats probably due to the auto tiling then. If you set a target tile size of 512 does it always work?
Hey @dem10com, I think I fixed this issue with 0.11.6 and the ncnn update that came with it. Could you give that a try?
@joeyballentine i try, install and update ncnn (2022.8.29). But i am getting this error.
An error occurred in a NCNN Upscale Image node:
An unexpected error occurred during NCNN processing.
Input values (partial):
• Image: RGB image 2048x2048
• Tile Size Target: 0
I change the tile size target but nothing change.
Hey @dem10com, I think I fixed this issue with 0.11.6 and the ncnn update that came with it. Could you give that a try?
Can you set the target tile size to 0 (auto) and try again? I believe that error means you ran out of vram
@joeyballentine I tried with "0" at first and it didn't work. and try again nothing change. by the way my gpu is radeon RX 5700 XT 8gb.
Oh that's weird. Well in that case could you try a lower v tile size target like 512
@joeyballentine here is the results;
but if i exit the program and start it again. It works for a 512x512 image. same way when i restart i get same error on 2048x2048 or 1024x1024 master file. most likely vram error as you said
Interesting. I wonder if it's not properly clearing out the vram after crashing when this happens. Can you tell by your vram in task manager or something that chaiNNer still has the vram allocated when this happens?
If it was, shouldn't I get a good result for the 2048x2048 master file when I turn on the computer for the first time?
When I check after getting the error, I get the result like this. 5 pyton 3.9 procces running in background. I just realized that even after exiting the program these continue to run in the background
Alphabetic Sort
using first time chainner.
i reset my computer. Run chainner. Look at my vram on statics. Almost empty
After i run and get error and quit application. memory never clear.
After waiting for a while, the memory is still not reset, and when you close the Pyton 3.9 on my background proccess vram is resetting.
i think the problem is python 3.9 is constantly running in the background even though we quit chainner.
Oh wow that's not good. chaiNNer is supposed to quit its backend when it closes. Maybe there's something specific on Mac causing that to not work properly. Does that happen no matter what, or only if you get an ncnn error?
I only use ncnn on mac, so I can say yes. I tried cpu mode and pytorch, there is no such problem.
Ok, this looks like one of three things:
- I either need to ignore vkQueueSubmit errors
- I need to manually flush vram on vkQueueSubmit error
- I need to go back and update ncnn to not raise en exception on vkQueueSubmit error
I'll investigate a bit, but idk how much I can accomplish without being able to test it myself
@dem10com could you try installing this? Install the .whl file via command line by doing /Application Support/chaiNNer/python/python/bin/python3.9 -m pip install /whatever/the/path/to/the/wheel/is.whl
Please let me know if the python process continues to get stuck using this build, or if it successfully quits now. Thanks.
ncnn-vulkan-python-vkQueuefailure-test-1-macos_cp39-macosx_arm64.zip
@joeyballentine i try but, i am not using ARM. My macos system is intel based hardware. ( hackintosh )
ERROR: ncnn_vulkan-2022.8.30-cp39-cp39-macosx_11_0_arm64.whl is not a supported wheel on this platform.
I install, Run chainner and dfter exiting the program, it still continues to run in the background. memory usage is almost %100. But the high resolution file (2048x2048) I tried before gave black result instead of the error I got.
