AMD Wants Your Logs To Help Optimize PyTorch & ComfyUI For Strix Halo, Radeon GPUs
([AMD] 6 Hours Ago
Collecting Data)
- Reference: 0001599834
- News link: https://www.phoronix.com/news/AMD-Logs-For-ComfyUI-PyTorch
- Source link:
If you are not satisfied with the current performance for PyTorch or ComfyUI / Stable Diffusion on your Strix Halo APU system or with other consumer RDNA3/RDNA4 Radeon consumer GPUs, AMD engineers are interested in your logs to help better optimize the performance going forward.
For any PyTorch program or ComfyUI / Stable Diffusion scenario especially where the performance is coming up short on the likes of the Strix Halo Radeon 8060S, Radeon RX 9000 series, or other RDNA3/RDNA4 GPUs, AMD is interested any performance logs willing to be shared to help them in tuning their libraries for better performance.
[1]This GitHub ticket is where the logs are being collected along with the Windows and Linux environment variables to set for collecting the relevant MIOpen and hipBLASLt logs.
It was further clarified there as well that they are interested in any RDNA3 or RDNA4 GPU target for optimizing:
"Yes. We're going to optimize all of the commonly used RDNA3/4 kernels and input on any of the architectures is welcome."
AMD's Anush Elangovan also commented on [2]X :
"We are working on performance uplifts for Strix Halo (can be any AMD GPU) and can use your help."
Again, see [3]this ticket for the details and to share any of the performance logs.
[1] https://github.com/ROCm/TheRock/issues/2591
[2] https://x.com/AnushElangovan/status/2001201855234261049
[3] https://github.com/ROCm/TheRock/issues/2591
For any PyTorch program or ComfyUI / Stable Diffusion scenario especially where the performance is coming up short on the likes of the Strix Halo Radeon 8060S, Radeon RX 9000 series, or other RDNA3/RDNA4 GPUs, AMD is interested any performance logs willing to be shared to help them in tuning their libraries for better performance.
[1]This GitHub ticket is where the logs are being collected along with the Windows and Linux environment variables to set for collecting the relevant MIOpen and hipBLASLt logs.
It was further clarified there as well that they are interested in any RDNA3 or RDNA4 GPU target for optimizing:
"Yes. We're going to optimize all of the commonly used RDNA3/4 kernels and input on any of the architectures is welcome."
AMD's Anush Elangovan also commented on [2]X :
"We are working on performance uplifts for Strix Halo (can be any AMD GPU) and can use your help."
Again, see [3]this ticket for the details and to share any of the performance logs.
[1] https://github.com/ROCm/TheRock/issues/2591
[2] https://x.com/AnushElangovan/status/2001201855234261049
[3] https://github.com/ROCm/TheRock/issues/2591