AMD ROCm or ZLUDA?

[andres-gr]

Anyone had any success story using either ROCm or ZLUDA with AMD?

Can't seem to get any decent performance out of DirectML on AMD.

[Elise96nl]

Tried everything to get zluda working. you have to disable cudnnn, that's possible with torch, but not with onnx runtime. So no zluda for this app :(

This is my directml benchmarks, with different torch, biggest problem is onnx directml doesnt handle more then 1 thread. so this will be it

nfo python: 3.10.6 • torch: 2.3.1+cu118 • gradio: 4.44.0 onnxruntime-directml 1.20.1 torch-directml 0.2.5.dev240914
Processing ******** -_repair_03-18-35.mp4 took 111.07 secs, 4.44 frames/s

info python: 3.10.6 • torch: 2.4.1+cpu • gradio: 4.44.0 onnxruntime-directml 1.20.1 torch-directml 0.2.5.dev240914
Processing ******** -_repair_03-18-35.mp4 took 112.20 secs, 4.39 frames/s

info python: 3.10.6 • torch: 2.3.1+cu118 • gradio: 4.44.0 onnxruntime-directml 1.20.1
Processing ******** -_repair_03-18-35.mp4 took 112.40 secs, 4.39 frames/s

info python: 3.10.6 • torch: 2.1.2+cu118 • gradio: 4.44.0 onnxruntime-directml 1.15.1
Processing ******** -_repair_03-18-35.mp4 took 120.54 secs, 4.09 frames/s

everything newer got me memory leaks, and crashes.