build: fix the build failures of Windows HIP release job

[lcy0321]

• Update the cache keys to include the HIP SDK version, preventing the use of outdated ROCm installation caches.
• Sync changes from release.yml to build.yml:
  • Update HIP SDK version to 25.Q3 and ROCm version to 6.4.2
  • Update the cache keys to reflect the new versions
• Remove Windows HIP release for gfx1151 since the current stable rocWMMA does not support gfx1151.

[slaren]

I tested this in my fork, but it doesn't compile with the new rocwmma version
https://github.com/slaren/llama.cpp/actions/runs/17710289321/job/50328038693

[lcy0321]

I tested this in my fork, but it doesn't compile with the new rocwmma version https://github.com/slaren/llama.cpp/actions/runs/17710289321/job/50328038693

I am facing the same error in my branch.

2025-09-14T11:18:22.0193699Z D:/a/llama.cpp/llama.cpp/rocwmma/library/include\rocwmma\internal./layout..\config.hpp:76:15: error: static assertion failed: Unsupported architecture
2025-09-14T11:18:22.0193902Z 76 | static_assert(0, "Unsupported architecture");

I am trying to deal with this error. If I can't fix it, I will try to revert the version of rocwmma

[slaren]

build: remove Windows HIP release for gfx1151
since the current stable rocWMMA does not support gfx1151.

Can we use the unstable version then? I would rather accept some instability than remove Strix Halo support. Support was already merged in May in ROCm/rocWMMA#538

[lcy0321]

Hi @slaren,

rocWMMA does not currently list gfx1151 among its supported architectures in the README.md, nor has it released a version with gfx1151 support. To enable the build for gfx1151, we need to locate a specific commit of rocWMMA, verify that we can build with it, and then modify the workflows to clone that commit instead of the current branch rocm-6.4.2.

The release workflow for this pull request has successfully passed on my fork:
https://github.com/lcy0321/llama.cpp/actions/runs/17711258621

I would prefer to merge this pull request first, and then you could create a separate PR for gfx1151 support. What do you think?

[CISC]

FYI release build just succeeded with rocm-6.2.4:
https://github.com/ggml-org/llama.cpp/actions/runs/17711218517/job/50330168262

[lcy0321]

FYI release build just succeeded with rocm-6.2.4: https://github.com/ggml-org/llama.cpp/actions/runs/17711218517/job/50330168262

Does this mean that we can build rocWMMA with GPU architectures that they haven't included in the code? That’s kind of surprising to me.