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

.github/workflows/on-pr-fabric.yml

@@ -189,6 +189,9 @@ jobs:
       sha: ${{ github.event.pull_request.base.sha || github.sha }}
       pr_head_sha: ${{ github.event.pull_request.head.sha }}
       workdir: packages/fabric
+      # fabric pulls qvac-fabric[hip-backend]; ggml-config find_dependency(hip)
+      # needs ROCm on the lint runner at configure time.
+      include-rocm-sdk: true
 
   detect-fabric-stack:
     if: needs.authorize.outputs.allowed == 'true'

.github/workflows/prebuilds-fabric.yml

@@ -44,4 +44,10 @@ jobs:
       repository: ${{ inputs.repository }}
       artifact-name-prefix: fabric-
       include-vulkan-sdk: true
+      # Cross-compile the ROCm/HIP backend (gfx1151) into the linux-x64 prebuild
+      # so every @qvac/fabric consumer gets libqvac-ggml-hip.so from the shared
+      # runtime. No AMD GPU needed on the runner, but the SDK is mandatory: the
+      # `hip` port is deterministic and hard-fails without ROCm rather than
+      # installing empty, so dropping this input breaks the linux-x64 build.
+      include-rocm: true
     secrets: inherit

packages/fabric/CHANGELOG.md

@@ -1,5 +1,42 @@
 # Changelog
 
+## [0.9.0] - 2026-08-25
+
+### Added
+
+- ROCm/HIP compute backend (`libqvac-ggml-hip.so`, gfx1151 / Strix Halo) in the
+  linux-x64 prebuild, via the `qvac-fabric[hip-backend]` feature. It ships as a
+  `GGML_BACKEND_DL` module under `prebuilds/linux-x64/qvac__fabric/` alongside
+  Vulkan, so every consumer of the shared runtime can select it — previously the
+  feature was requested per-consumer by `@qvac/vla-ggml`, which no longer builds
+  its own ggml. At **runtime** this is fail-safe: the DL loader skips the module
+  on non-AMD hosts and falls back to Vulkan/CPU. At **build time** it is not —
+  the `hip` port is deterministic and hard-fails when no ROCm SDK is found,
+  because a host-dependent skip would let the vcpkg binary cache conflate a
+  no-HIP build with a real HIP build under an identical ABI hash. Building
+  `packages/fabric` for linux-x64 therefore requires a ROCm/TheRock install,
+  located via `ROCM_PATH` or `/opt/rocm`. Other platforms are unaffected — the
+  `hip` dependency is gated on `linux & x64`.
+
+### Changed
+
+- `prebuilds-fabric.yml` sets `include-rocm: true` so the linux-x64 prebuild
+  cross-compiles the HIP backend, and `on-pr-fabric.yml`'s `cpp-lint` sets
+  `include-rocm-sdk: true` because `ggml-config.cmake` resolves
+  `find_dependency(hip/hipblas/rocblas)` at configure time. No AMD GPU is
+  required on either runner, but the SDK itself is mandatory on both: without it
+  `cpp-lint` fails at configure time resolving the port's `$ENV{ROCM_PATH}` shim
+  to an empty prefix (`/lib/cmake/hip/hip-config.cmake`), which a warm vcpkg
+  binary cache can disguise as a successful `hip` install.
+- `qvac-registry-vcpkg` baseline `c57eec31` -> `f04e2447`, matching
+  `@qvac/vla-ggml`. Required because `qvac-fabric[hip-backend]` depends on `hip`
+  with no version constraint, so its version comes from the pinned baseline, and
+  the `hip` port does not exist at `c57eec31`. No other version selected by this
+  package changes: `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.
+
 ## [0.8.0] - 2026-08-24
 
 ### Changed

packages/fabric/NOTICE

@@ -23,5 +23,7 @@ C++ Dependencies
 
 --- mit (MIT License) ---
 
+  hip (system ROCm/HIP SDK shim, linux-x64 only)
+    https://rocm.docs.amd.com
   qvac-fabric (forked llama.cpp + ggml)
     https://github.com/tetherto/qvac-fabric-llm.cpp

packages/fabric/README.md

@@ -25,6 +25,8 @@ consumer guide.
   ship under `prebuilds/<platform>/qvac__fabric/` and are loaded at runtime via
   `ggml_backend_load_all_from_path()`. On **macOS, Windows, and iOS** the backends
   are linked statically inside `qvac__fabric.bare` and self-register on load.
+  On **linux-x64** this includes the ROCm/HIP backend (`libqvac-ggml-hip.so`,
+  gfx1151) alongside Vulkan; the DL loader skips it on non-AMD hosts.
 
 ## Architecture
 
@@ -68,11 +70,17 @@ npm install
 npm run build   # bare-make generate && bare-make build && bare-make install
 ```
 
+On **linux-x64** a ROCm/TheRock SDK is required, discovered via `ROCM_PATH` or
+`/opt/rocm`. The `hip` port is deterministic — it hard-fails rather than
+installing empty, so that the vcpkg binary cache cannot conflate a no-HIP build
+with a real one under the same ABI hash. Other platforms need nothing extra; the
+`hip` dependency is gated on `linux & x64`.
+
 ## Supported platforms
 
 | Platform | Triplet | Backends |
 |----------|---------|----------|
-| Linux | `x64-linux`, `arm64-linux` | shared `.so` under `prebuilds/<platform>/qvac__fabric/` |
+| Linux | `x64-linux`, `arm64-linux` | shared `.so` under `prebuilds/<platform>/qvac__fabric/` (x64 also ships ROCm/HIP) |
 | macOS | `arm64-osx` | static (CPU, Metal) inside `.bare` |
 | Windows | (default MSVC) | static inside `.bare` |
 | Android | `arm64-android` | shared `.so` under `prebuilds/<platform>/qvac__fabric/` |

packages/fabric/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@qvac/fabric",
-  "version": "0.8.0",
+  "version": "0.9.0",
   "description": "Shared bare addon hosting the qvac-fabric (forked llama.cpp + ggml) runtime for QVAC inference addons",
   "addon": true,
   "engines": {

packages/fabric/vcpkg-configuration.json

@@ -1,7 +1,7 @@
 {
   "default-registry": {
     "kind": "git",
-    "baseline": "c57eec31bdc37ce3a98537b312f5a366b554db6a",
+    "baseline": "f04e244701f462c4c8fead7b50bcaffa52c052ac",
     "repository": "https://github.com/tetherto/qvac-registry-vcpkg.git"
   },
   "registries": [

packages/fabric/vcpkg.json

@@ -6,7 +6,10 @@
     },
     {
       "name": "qvac-fabric",
-      "version>=": "10297.0.0"
+      "version>=": "10297.0.0",
+      "features": [
+        "hip-backend"
+      ]
     },
     {
       "name": "qvac-lint-cpp",