feat[notask]: build @qvac/fabric with the ROCm/HIP backend

[jpgaribotti]

🎯 What problem does this PR solve?

qvac-fabric[hip-backend] was requested per-consumer by @qvac/vla-ggml, which paired it with include-rocm: true on prebuilds-vla.yml and include-rocm-sdk: true on its cpp-lint call. That was correct while vla built its own ggml.

#3998 migrates vla-ggml onto the shared @qvac/fabric runtime and removes all three, but nothing on the fabric side picks the feature up:

• packages/fabric/vcpkg.json declares only vk-profiling; it never requests hip-backend.
• prebuilds-fabric.yml passes only artifact-name-prefix and include-vulkan-sdk: true; include-rocm defaults to false in reusable-prebuilds.yml.
• on-pr-fabric.yml's cpp-lint call passes no include-rocm-sdk.

Net effect once #3998 lands: include-rocm, include-rocm-sdk and .github/actions/setup-rocm have zero callers repo-wide, libqvac-ggml-hip.so is built by no pipeline and ships in no prebuild, and the HIP-preference branch at packages/vla-ggml/addon/src/utils/BackendSelection.cpp:132-148 becomes unreachable — hipDev is always null and pickBestGpuDevice() falls through to Vulkan. This gives back the ~18% gfx1151 warm-path win from #2781 / #2971 (479 → 384 ms).

Nothing turns red, because with no manifest requesting hip-backend the hip port is never resolved at all — the regression is a silent omission from the prebuild, not a build error. The vla-side CI that would have surfaced it is removed in the same change.

📝 How does it solve it?

Moves the request down to the shared runtime, so the backend ships once for every fabric consumer instead of per-addon.

• packages/fabric/vcpkg.json — request qvac-fabric[hip-backend]. version>= is already 10297.0.0, the same revision vla pinned for HIP. The port's hip dependency is platform: "linux & x64"-gated, and vla requested the feature unqualified across all 9 prebuild targets on main, so an unconditional request is proven safe on darwin/windows/android.
• packages/fabric/vcpkg-configuration.json — qvac-registry-vcpkg baseline c57eec31 → f04e2447, matching vla. Required, not cosmetic: qvac-fabric[hip-backend] depends on hip with no version constraint, so hip's version comes from the pinned baseline, and the hip port does not exist at c57eec31. Nothing else this package selects moves — qvac-fabric and qvac-lint-cpp are pinned above their baseline entries by version>=, opencl / vcpkg-cmake / vcpkg-cmake-config are identical in both baselines, and spirv-headers resolves from the separately pinned microsoft/vcpkg registry. CI confirms the resolution: qvac-fabric[core,gpu-backends,hip-backend,llama]:x64-linux@10297.0.0 and hip:x64-linux@7.13.0.
• prebuilds-fabric.yml — include-rocm: true, cross-compiling libqvac-ggml-hip.so (gfx1151) into the linux-x64 prebuild. No AMD GPU on the runner. Covers both callers of this workflow (on-pr-fabric.yml and on-merge-fabric.yml).
• on-pr-fabric.yml — include-rocm-sdk: true on the cpp-lint call. Easy to miss but mandatory: ggml-config.cmake resolves find_dependency(hip/hipblas/rocblas) at configure time, which is exactly why fec63dd21 added this for vla.
• packages/fabric/NOTICE — hip attribution, so the notice-drift check stays green.
• Version 0.8.0 → 0.9.0, CHANGELOG entry, and README backend/platform/build notes.

No build or packaging change is needed. packages/fabric/CMakeLists.txt:70-87 already walks GGML_AVAILABLE_BACKENDS and stages MODULE_LIBRARY/SHARED_LIBRARY targets via add_bare_module(qvac-fabric EXPORTS ...), and package.json files already publishes prebuilds/*/qvac__fabric/**. Enabling the feature is sufficient for the .so to be staged and published.

There is no cpp-tests-fabric, and test-fabric-stack-actions.yml only exercises the composite actions, so those are the only two CI touchpoints.

The hip port is deterministic, not build-time fail-safe

Worth stating explicitly because it is the opposite of what you might assume, and an earlier revision of this description got it wrong. The port hard-errors when no ROCm SDK is present rather than installing empty, quoting its portfile.cmake:

a host-dependent "skip" would let the vcpkg binary cache conflate a no-HIP build with a real HIP build (identical ABI).

Only the runtime path is fail-safe: the DL loader skips libqvac-ggml-hip.so on non-AMD hosts and falls back to Vulkan/CPU. Consequences:

• Building packages/fabric for linux-x64 now requires a ROCm/TheRock install, found via ROCM_PATH or /opt/rocm. Previously only vla-ggml developers needed one; this moves the requirement onto everyone who builds the shared runtime. Documented in the README build section. Other platforms are unaffected (linux & x64 gate).
• Any pipeline that builds fabric on linux-x64 without include-rocm / include-rocm-sdk will now fail rather than silently degrade. Both such call sites are updated here.

🧪 How was it tested?

Full green CI on this exact commit via workflow_dispatch on the branch — run 32878043355, success in 23m 7s, all 9 prebuild targets plus cpp-lint, sanity-checks, fabric-npm-consumer-smoke and merge-guard.

Downloaded the resulting fabric-linux-x64 artifact and inspected it:

• linux-x64/qvac__fabric/libqvac-ggml-hip.so is present — 67.0 MiB uncompressed, 11.6 MiB gzipped — beside libqvac-ggml-vulkan.so and the 14 CPU variants.
• Its only embedded code object is hipv4-amdgcn-amd-amdhsa--gfx1151, confirming the intended target and that it is not fattened with extra architectures.
• cpp-lint logs show ggml::ggml-hip among the linkable targets.

So the packaging cost is concrete: +11.6 MiB to the published tarball for every @qvac/fabric consumer, on linux-x64 only.

Still needs hardware: runtime confirmation on a Strix Halo (gfx1151) host that BackendSelection logs preferring HIP/ROCm GPU. Not reproducible in CI — the runners compile HIP but have no AMD GPU.

Not verified byte-for-byte: that the other 8 targets are unchanged. Satisfied by construction instead — the hip dependency is platform: "linux & x64"-gated, so the feature contributes nothing off linux-x64, and all 8 built green.

actionlint was not run locally: it is not installed on this machine and the repo forbids curl | bash installers. The policy-tests job in security-baseline.yml lints both edited workflows.

⚠️ Why the automatic PR checks are red

on-pr-fabric.yml is triggered by pull_request_target, which loads the workflow YAML from the base branch. Until this PR merges, the auto checks run main's copies of prebuilds-fabric.yml and on-pr-fabric.yml — which do not pass include-rocm / include-rocm-sdk — against this branch's source tree, which does request hip-backend. Mismatched by construction.

The two runs are a clean controlled experiment, same head SHA aaaeca1:

run	trigger	result
32875616325	pull_request_target	failure
32878043355	workflow_dispatch	success

The failing run's logs print the resolved inputs as include-rocm: false and include-rocm-sdk: false while checking out aaaeca1, and the only jobs that fail are exactly the two this PR edits (cpp-lint, prebuild / linux-x64) plus the merge-guard aggregate.

They fail in two different-looking ways, same root cause:

• prebuild / linux-x64 — vcpkg binary cache miss, so the portfile ran and hard-errored: hip port: no ROCm SDK found.
• cpp-lint — cache hit (hip:x64-linux@7.13.0 restored in 197 ms, portfile never executed), so it failed later at consumer configure: CMake Error at build/_vcpkg/x64-linux/share/hip/hip-config.cmake:1 (include): /lib/cmake/hip/hip-config.cmake. The leading / is the port's relocatable include("$ENV{ROCM_PATH}/…") shim expanding an unset ROCM_PATH to empty. Noting it because a warm binary cache can disguise a missing SDK as a successful hip install.

Both resolve on merge, and for any PR branched off a main that contains this change.

🔌 API Changes

None. @qvac/fabric has no JS API; this only changes what ships inside the linux-x64 prebuild.

Notes for reviewers

This is a cost/policy decision, not just a checkbox. hip-backend was opt-in per-consumer precisely so that only vla paid for it. With the runtime shared, that granularity is gone: every fabric consumer's linux-x64 prebuild carries the HIP module (+11.6 MiB compressed), every fabric linux-x64 build and lint needs ROCm on the runner, and — because the port is deterministic — so does every local linux-x64 build.

Two alternatives were considered and rejected:

1. Gate it behind a manifest feature + CMake option, mirroring vk-profiling, with prebuilds-fabric.yml passing -D HIP_BACKEND=ON. Keeps local builds working without ROCm. Rejected to stay consistent with how vla-ggml already declares it, and to avoid a build configuration that CI exercises but developers do not.
2. Drop HIP explicitly — remove the dead branch in BackendSelection.cpp, fix the packages/vla-ggml/CHANGELOG.md claim that HIP "is provided by @qvac/fabric when present in its prebuilds", and delete the now-callerless include-rocm / include-rocm-sdk inputs and the setup-rocm action. This is the right change if the size or build cost is judged unacceptable.

Merge ordering: independent of #3998 and safe to merge in either order. Merging this first means HIP is never absent from a published prebuild.

Raised from review of #3998.

[github-actions]

Review Status

Current Status: ✅ APPROVED
Approvals so far: Team Lead: 1, Member: 1