bad efficiency when synthesizing on AMD GPU

[DiPiVR]

Hello, I tried to synthesize on my AMD 7900 XTX, however the performance results are... questionable. (audio file is fine though)

I tried both CPU and GPU mode. It ran almost twice as long on GPU, with 100% utilization. Do you have any ideas on why it is and what can I adjust to improve?

The test sentence: "This is a moderately long sentence, also I like trains."
CPU - 5800X3D: Processing time: 5.311s, Real-time factor: 1.118, utilization: 50%
GPU - 7900XTX: Processing time: 8.369s, Real-time factor: 1.843, utilization: 100%

Details of my setup: Linux Mint 22 (Ubuntu 22.04 base), AMD ROCm 7.2, Python 3.12.3, coqui-tts installed via pip in venv, torch 2.11.0+rocm7.2 installed with official instructions, server ran with tts-server --model_name tts_models/multilingual/multi-dataset/xtts_v2 --device cuda.

[eginhard]

I don't have a ROCm device unfortunately, so can't do any testing myself. But if you or anyone can pinpoint any specific issues I'm happy to help fixing them.

[DiPiVR]

I dug deeper and found a workaround here

TLDR: use env variable MIOPEN_FIND_MODE=2

Same setup, updates: ROCm 7.2.4, pytorch 2.12.1. The results are way better:

CPU - 5800X3D: Processing time: 5.038s, Real-time factor: 1.148
GPU - 7900XTX: Processing time: 1.467s, Real-time factor: 0.368