ci: add riscv64 to release binaries

[gounthar]

Overview

Add riscv64 to the Ubuntu CPU release matrix using cross-compilation from a standard ubuntu-24.04 runner with gcc-14-riscv64-linux-gnu.

Changes:

• Add riscv64 / ubuntu-24.04 matrix entry with cross-compile flags
• Add a dedicated "Install cross-compiler (riscv64)" step (installs gcc-14-riscv64-linux-gnu)
• Exclude riscv64 from the "Toolchain workaround (GCC 14)" step (cross-compiler is already GCC 14)
• Pass cross-compilation flags via matrix.cmake_cross into the cmake configure step
• Add riscv64 line to release notes template

Closes #20988

Why cross-compilation rather than a native riscv64 runner?

We benchmarked both approaches on the same codebase. On RISE Project ubuntu-24.04-riscv runners, a full Release build takes roughly 49 minutes (build step alone). Cross-compiling from ubuntu-24.04 x86_64 with a gcc-14 RISC-V cross-compiler takes roughly 5 minutes total — about a 10x difference. Using a standard GitHub-hosted runner also avoids any dependency on runner availability or a third-party GitHub App installation.

What the binary supports

• GGML_CPU_ALL_VARIANTS=ON — multiple ISA code paths compiled in; runtime detection selects the best available (RVV, ZVFH, etc.)
• LLAMA_OPENSSL=OFF — avoids the libssl-dev:riscv64 multiarch setup needed for cross-compilation; the binary is fully functional without it
• LLAMA_FATAL_WARNINGS=ON — kept consistent with other ubuntu-cpu builds; gcc-14 cross-compiler handles _Float16 without warnings on riscv64

Additional information

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: NO

[CISC]

The existing riscv64 CI (build-riscv.yml) already validates that llama.cpp builds and passes tests on this architecture.

Nevertheless, can you provide a link to successful release on your fork?

[CISC]

The ubuntu-24.04-riscv runner label is provided by the RISE Project (https://riseproject.dev) and is available free of charge to open-source GitHub repositories. To enable it, a RISE admin needs to install the RISE GitHub App on the ggml-org organization. I can help coordinate that if needed.

We already have a RISC-V runner from Cloud-V, should be able to use that?

[CISC]

Relevant info: #16682 (comment)

[gounthar]

Good to know the Cloud-V runners are back up. That simplifies things since they are already configured on ggml-org. 👍

One question: what GCC version do the Cloud-V RISCV64 runners have? The release build needs GCC 14+ because GCC 13 cannot parse the full ISA string from /proc/cpuinfo and rejects _Float16 in pedantic mode (at least on the rise runner). I hit both issues when testing on RISE runners (which run GCC 13.3.0 on Ubuntu 24.04).

If the Cloud-V runners have GCC 14, the release build should work with the same flags as x64/s390x. If not, we would need to either disable GGML_NATIVE and LLAMA_FATAL_WARNINGS, or install GCC 14 as a setup step.

I will update the PR to use runs-on: RISCV64 once I know the GCC version.

[CISC]

I will update the PR to use runs-on: RISCV64 once I know the GCC version.

14.2.0, but read the linked comment above.

[taronaeo]

As requested by @/CISC, any chance you had a successful run in your fork? Can you share a link?

[gounthar]

The release workflow only triggers on tags and workflow_dispatch, not on PRs, so CI did not run on this PR. I cannot trigger it on my fork either because the workflow needs the RISCV64 Cloud-V runners which are only configured on ggml-org.

I am building locally on native riscv64 hardware (BananaPi F3, GCC 14.2.0, rv64gc+RVV) and will share the results shortly. The existing CI (build-riscv.yml) also validates riscv64 builds on Cloud-V runners, so the release build should work with the same infrastructure.

Would a maintainer be able to trigger a manual workflow_dispatch run to validate the release matrix change?

[gounthar]

Native riscv64 build succeeded on a BananaPi F3 (SpacemiT K1, GCC 14.2.0, rv64gc+RVV):

• llama-cli (4.6M) + llama-server (6.5M) produced
• Build time: 56 minutes with -j8
• ./llama-cli --version outputs: version: 1 (b0f0dd3) built with GNU 14.2.0 for Linux riscv64
• Full build log: https://gist.github.com/gounthar/3d836c6114ce8e9064d59872985787f2

The Cloud-V RISCV64 runners also have GCC 14.2.0, so the release build should work with the same toolchain.

[gounthar]

Rebased from upstream master and switched the approach: the original version used RISE Project ubuntu-24.04-riscv runners (which required installing a third-party GitHub App on ggml-org), and an earlier draft used Cloud-V RISCV64 runners. This version uses cross-compilation from a standard ubuntu-24.04 runner instead.

The practical reason: we measured build times on both. Native riscv64 (RISE runner) takes ~49 minutes for the build step. Cross-compiling with gcc-14-riscv64-linux-gnu from x86_64 takes ~5 minutes total. The end result is the same set of binaries, tested against the same runtime ISA detection (GGML_CPU_ALL_VARIANTS=ON).

The only dependency added is gcc-14-riscv64-linux-gnu and g++-14-riscv64-linux-gnu, which are in the standard Ubuntu 24.04 package repos — no new runners, apps, or service accounts needed.

[CISC]

Please provide a link to a successful release on your fork (disable s390x to allow it to complete).

[CISC]

Either make it link against the right libssl, or build with boringssl, we don't want a release without.

[CISC]

Suggested change

	${{ matrix.cmake_cross || '' }} \
	${{ matrix.cmake_cross }} \

Doesn't it already evaluate to an empty string if not defined?