ci: add Linux HIP (ROCm) backend build to release workflow

[dev-miro26]

Overview

Add HIP/ROCm backend builds (Linux and Windows) to the Menlo CI release pipeline so that Jan can distribute pre-built llama-server binaries for AMD GPU users.

What changed in menlo-build.yml:

• build-hip-linux — new standalone job that builds llama-server with -DGGML_HIP=ON inside the rocm/dev-ubuntu-22.04:6.2 container, targeting gfx906/gfx908/gfx90a/gfx1030/gfx1100/gfx1101/gfx1102. Uploads llama-{version}-bin-linux-hip-x64.tar.gz to the GitHub release.
• build-hip-windows — new standalone job that installs AMD HIP SDK (26.Q1) on a windows-2022 runner, builds llama-server with HIP clang via the Unix Makefiles generator, bundles required runtime DLLs (libhipblas.dll, libhipblaslt.dll, rocblas.dll + libraries), code-signs, and uploads llama-{version}-bin-win-hip-x64.tar.gz.
• create-checksum-file — updated to depend on both HIP jobs and conditionally include their SHA-512 checksums in checksum.yml only when the respective build succeeds.
• Both HIP jobs use continue-on-error: true so a HIP build failure does not block the rest of the release (CUDA, Vulkan, CPU artifacts are always published).

Note: macOS is not included because AMD does not support ROCm/HIP on macOS.

Additional information

• The Linux HIP job follows the same pattern as the existing CUDA matrix builds (checkout → replace Makefile → install tools → build → version file → package → checksum → upload release asset).
• The Windows HIP job follows the upstream llama.cpp windows-hip job approach from release.yml: it uses Unix Makefiles + HIP clang instead of MSVC/Ninja, since HIP compilation on Windows requires the HIP SDK's clang compiler. ROCm installation is cached via actions/cache@v5.
• This enables Jan to auto-detect AMD GPU + ROCm on a user's system and download the appropriate llama.cpp backend without requiring users to select it manually.

GPU targets

Architecture	Example GPUs
gfx906	Radeon VII, MI50
gfx908	MI100
gfx90a	MI210, MI250, MI250X
gfx1030	RX 6800, RX 6900 XT
gfx1100	RX 7900 XTX, RX 7900 XT
gfx1101	RX 7800 XT, RX 7700 XT
gfx1102	RX 7600

[geeksville]

I'm just a bystander with a Strix Halo based laptop. Any chance this change could be extended to also support gfx1151? (rdna 3.5). I could test it on my Asus Flow Z13 128GB model if it would help.

[dev-miro26]

Architecture	Generation	Example GPUs	Status
gfx908	CDNA 1	MI100	existing
gfx90a	CDNA 2	MI210, MI250, MI250X	existing
gfx942	CDNA 3	MI300X, MI300A	new
gfx1030	RDNA 2	RX 6800, RX 6900 XT	existing
gfx1100	RDNA 3	RX 7900 XTX, RX 7900 XT	existing
gfx1101	RDNA 3	RX 7800 XT, RX 7700 XT	existing
gfx1102	RDNA 3	RX 7600	existing
gfx1150	RDNA 3.5	Strix Point APUs	new
gfx1151	RDNA 3.5	Strix Halo (e.g. Asus Flow Z13)	new
gfx1200	RDNA 4	RX 9070 XT	new
gfx1201	RDNA 4	RX 9070	new

[dev-miro26]

@geeksville
please check again.
Thanks

[dev-miro26]

Hi, @louis-jan
Could you please review this PR?

[louis-jan]

cc @qnixsynapse

[louis-jan]

Also @Minh141120

[louis-jan]

I think this wont work because it needs correct runner with AMD gpu. Cmiiw @Minh141120

[qnixsynapse]

Offline compilation (i.e without AMD GPU might be possible. But yes, @Minh141120 might know better.

[Minh141120]

Hi @dev-miro26, really appreciate the work on this 🙌.

To help us validate the changes (especially across both Linux and Windows), would you mind sharing a short video showing a successful build and Jan running on both platforms?

This will help us validate the changes more confidently.

[dev-miro26]

@Minh141120
I am sorry. As I said to @louis-jan before, I don't have AMD device. So I am finding who has AMD device.
@Vanalite : Could you test this please?

[Minh141120]

Really appreciate the effort here, especially given the hardware limitation.

Since this introduces a new backend across both Windows and Linux, we’ll need validation on real AMD hardware before merging to ensure it works reliably for users.

Without proper test results, we won’t be able to proceed with the merge at this time. We’ll keep the PR open and revisit it once test results are available from verified AMD hardware.

[dev-miro26]

Hi, @louis-jan
How are you?
I think Vanalite can test this PR.
Right?