ROCm 7.2.0 RDNA2 Support Contribution

[Maxritz]

I've been working on extending rocm_sdk_builder to support ROCm 7.2.0 for RDNA2 GPUs (RX 6700 XT and gfx1030-1032 family).

What I've built:

Complete patch set for all 16 ROCm 7.2 libraries
FP8→INT8 conversion layer (enables FLUX.1/SD3 on RDNA2)
WMMA software emulation using RDNA2's native dot instructions
VGPR auto-segmentation for 256 VGPR hardware limit
Comprehensive documentation (build guide, troubleshooting, optimization guides)
The work is ready in my fork: https://github.com/Maxritz/rocm_sdk_builder/tree/rocm-7.2-rdna2-support

I'd like to contribute this back to rocm_sdk_builder. Would you be open to reviewing a pull request? I'm currently testing the build on my RX 6700 XT system and can provide benchmarks/validation results.

This would enable millions of RDNA2 GPU owners to use ROCm 7.2 and modern FP8 models on consumer hardware.

if anyone wants to try it , please do and let me know what you get. I still have to try and compile the whole thing

[patientx]

would they work with windows and what about pytorch packages ?

[Maxritz]

linux.. am doing these on nobara
Will try to do a windows build too.. i mean you can try too.. if you load this to antigravity! or vscode i think

[patientx]

this is for 7.2 though so no windows then, we only have 7.1 and 7.1.2 atm ?

[Maxritz]

see if my repo can be used.. it has the files.. it may compile! not sure