Disable GGML_HIP_ROCWMMA_FATTN

[slojosic-amd]

This PR should give opportunity for testing performances with bigger context sizes after disabling GGML_HIP_ROCWMMA_FATTN
It should help to fix issues like this one: #36

[Goldenkoron]

Please merge this, rocm has been close to unusable for awhile now with strix halo and I want to use newer models that are not compatible with the older working versions.

[slojosic-amd]

@Goldenkoron did you confirm with https://github.com/lemonade-sdk/llamacpp-rocm/actions/runs/20528450015/artifacts/4972554015 that disabling GGML_HIP_ROCWMMA_FATTN gives you better perf numbers for #36

[Goldenkoron]

@Goldenkoron did you confirm with https://github.com/lemonade-sdk/llamacpp-rocm/actions/runs/20528450015/artifacts/4972554015 that disabling GGML_HIP_ROCWMMA_FATTN gives you better perf numbers for #36

Sorry I didn't see a windows release was sent. I'll test this later today when I'm off work and report back.

[danielholanda]

Thanks @slojosic-amd and @Goldenkoron. Please provide some numbers that shows that this indeed solves the problem and I will be happy to merge this.

[danielholanda]

Confirmed as discussed in #36