Enable ROCWMMA on all builds

[danielholanda]

Description

ROCWMMA often shows better prompt processing across models.

This PR compiles ROCM with ROCWMMA enabled.

Results

gpt-oss-20b

As shown below, ROCWMMA enabled (first test) gives a large speed boost for prompt processing (pp512): 1423.31 t/s vs 969.03 t/s without it. Roughly 47% faster.

With ROCWMMA

C:\Users\daniel-halo\Downloads\llama-windows-rocm-gfx1151-x64-fa>llama-bench.exe -m "C:\Users\daniel-halo\.cache\huggingface\hub\models--unsloth--gpt-oss-20b-GGUF\snapshots\4953a383fb4ca25caeb9bc0e366c441cbb219afe\gpt-oss-20b-Q4_K_M.gguf" -ngl 99 -fa 1
HIP Library Path: C:\Users\daniel-halo\Downloads\llama-windows-rocm-gfx1151-x64-fa\amdhip64_7.dll
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 ROCm devices:
  Device 0: AMD Radeon(TM) 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
| gpt-oss ?B Q4_K - Medium       |  10.81 GiB |    20.91 B | ROCm       |  99 |  1 |           pp512 |       1423.31 ± 7.65 |
| gpt-oss ?B Q4_K - Medium       |  10.81 GiB |    20.91 B | ROCm       |  99 |  1 |           tg128 |         71.35 ± 0.28 |

Without ROCWMMA

C:\Users\daniel-halo\Downloads\llama-b1028-windows-rocm-gfx1151-x64-main>llama-bench.exe -m "C:\Users\daniel-halo\.cache\huggingface\hub\models--unsloth--gpt-oss-20b-GGUF\snapshots\4953a383fb4ca25caeb9bc0e366c441cbb219afe\gpt-oss-20b-Q4_K_M.gguf" -ngl 99 -fa 1            
HIP Library Path: C:\Users\daniel-halo\Downloads\llama-b1028-windows-rocm-gfx1151-x64-main\amdhip64_7.dll
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 ROCm devices:
  Device 0: AMD Radeon(TM) 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
| gpt-oss ?B Q4_K - Medium       |  10.81 GiB |    20.91 B | ROCm       |  99 |  1 |           pp512 |       969.03 ± 12.72 |
| gpt-oss ?B Q4_K - Medium       |  10.81 GiB |    20.91 B | ROCm       |  99 |  1 |           tg128 |         69.89 ± 0.27 |

build: c24f4e26 (6149)

Target Devices

Compilation works across all target families as shown here:
https://github.com/lemonade-sdk/llamacpp-rocm/actions/runs/16947880973

Open Questions

• Should we really use the latest version of https://github.com/ROCm/rocWMMA.git at all times or snap to a specific stable one?
• Can we heavily simplify the ROCWMMA patch, given that HIP: disable sync warp shuffel operators from clr amd_warp_sync_functions.h ggml-org/llama.cpp#15273 merged?

Closes #7

[kpoineal]

@danielholanda Is this going to be merged at some point? My testing shows a fairly significant performance boost.

[danielholanda]

Yes, I'm just now actively working on #9 and adding tests to ensure things work properly, so we can confidently merge this.

[fassn]

@danielholanda Is this going to be merged at some point? My testing shows a fairly significant performance boost.

To support this comment, here are my results.

[danielholanda]

@fassn Thanks for your comments. Planning on merging this today if this passes: https://github.com/lemonade-sdk/llamacpp-rocm/actions/runs/17588392336

[danielholanda]

All tests passing. Merging!