bong-water-water-bong · Sep 23, 2026 · Sep 23, 2026 · Sep 23, 2026
diff --git a/.github/workflows/bump-hrx.yml b/.github/workflows/bump-hrx.yml
@@ -0,0 +1,130 @@
+# Copyright 2026 bong-water-water-bong
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# Keep third_party/llama.cpp and third_party/hrx-system on the pair AMD's
+# integration repo (ROCm/ggml-staging-automation) builds and tests together
+# (docs/hrx.md, "Staying current"). When AMD moves its pins, this workflow:
+#   1. syncs 1bit-MONSTER/llama.cpp: master to ggml-org, 1bit/hrx-vulkan to AMD's pin;
+#   2. opens a PR here moving both submodules.
+# GitHub-hosted CI builds that PR without HRX; run tests/hrx_lemonade_e2e.sh on
+# Strix Halo before merging it.
+#
+# Needs the secret HRX_BUMP_TOKEN: a fine-grained token with Contents
+# read/write on 1bit-MONSTER/llama.cpp and 1bit-MONSTER/engine, and Pull
+# requests read/write on 1bit-MONSTER/engine. (A PR opened with the default
+# GITHUB_TOKEN would not start CI, and that token cannot push to the fork.)
+name: bump-hrx
+
+on:
+  schedule:
+    - cron: "17 6 * * *"
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+concurrency:
+  group: bump-hrx
+  cancel-in-progress: false
+
+jobs:
+  bump:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Require the token
+        env:
+          HRX_BUMP_TOKEN: ${{ secrets.HRX_BUMP_TOKEN }}
+        run: |
+          if [ -z "$HRX_BUMP_TOKEN" ]; then
+            echo "::error::secret HRX_BUMP_TOKEN is not set (see the header of this workflow)"
+            exit 1
+          fi
+
+      - uses: actions/checkout@v4
+        with:
+          token: ${{ secrets.HRX_BUMP_TOKEN }}
+          persist-credentials: true
+
+      - name: Read AMD's tested pair and ours
+        id: pins
+        env:
+          GH_TOKEN: ${{ secrets.HRX_BUMP_TOKEN }}
+        run: |
+          set -euo pipefail
+          amd_llama=$(gh api repos/ROCm/ggml-staging-automation/contents/llama.cpp --jq .sha)
+          amd_hrx=$(gh api repos/ROCm/ggml-staging-automation/contents/hrx-system --jq .sha)
+          ours_llama=$(git ls-tree HEAD third_party/llama.cpp | awk '{print $3}')
+          ours_hrx=$(git ls-tree HEAD third_party/hrx-system | awk '{print $3}')
+          echo "AMD:  llama.cpp $amd_llama  hrx-system $amd_hrx"
+          echo "ours: llama.cpp $ours_llama  hrx-system $ours_hrx"
+          {
+            echo "amd_llama=$amd_llama"; echo "amd_hrx=$amd_hrx"
+            echo "ours_llama=$ours_llama"; echo "ours_hrx=$ours_hrx"
+            if [ "$amd_llama" = "$ours_llama" ] && [ "$amd_hrx" = "$ours_hrx" ]; then echo "changed=false"; else echo "changed=true"; fi
+          } >> "$GITHUB_OUTPUT"
+
+      - name: Sync the fork
+        if: steps.pins.outputs.changed == 'true'
+        env:
+          GH_TOKEN: ${{ secrets.HRX_BUMP_TOKEN }}
+          AMD_LLAMA: ${{ steps.pins.outputs.amd_llama }}
+        run: |
+          set -euo pipefail
+          # master mirrors ggml-org; 1bit/hrx-vulkan carries no commits of its own, so it
+          # moves to AMD's pin as long as that pin descends from it (no force needed).
+          gh api -X POST repos/1bit-MONSTER/llama.cpp/merge-upstream -f branch=master > /dev/null
+          git init -q fork && cd fork
+          git remote add fork "https://x-access-token:${GH_TOKEN}@github.com/1bit-MONSTER/llama.cpp.git"
+          git remote add amd https://github.com/AMD-Ecosystem/llama.cpp.git
+          git fetch -q --filter=blob:none amd "$AMD_LLAMA"
+          git fetch -q --filter=blob:none fork 1bit/hrx-vulkan
+          if ! git merge-base --is-ancestor FETCH_HEAD "$AMD_LLAMA"; then
+            echo "::error::1bit/hrx-vulkan has commits that AMD's pin does not contain; rebase them by hand"
+            exit 1
+          fi
+          git push -q fork "$AMD_LLAMA:refs/heads/1bit/hrx-vulkan"
+
+      - name: Open the bump PR
+        if: steps.pins.outputs.changed == 'true'
+        env:
+          GH_TOKEN: ${{ secrets.HRX_BUMP_TOKEN }}
+          AMD_LLAMA: ${{ steps.pins.outputs.amd_llama }}
+          AMD_HRX: ${{ steps.pins.outputs.amd_hrx }}
+          OURS_LLAMA: ${{ steps.pins.outputs.ours_llama }}
+          OURS_HRX: ${{ steps.pins.outputs.ours_hrx }}
+        run: |
+          set -euo pipefail
+          branch="bump-hrx/${AMD_LLAMA:0:12}-${AMD_HRX:0:12}"
+          if git ls-remote --exit-code origin "refs/heads/$branch" > /dev/null; then
+            echo "$branch already exists"; exit 0
+          fi
+          git config user.name "hrx-bump"
+          git config user.email "hrx-bump@users.noreply.github.com"
+          git switch -c "$branch"
+          git update-index --cacheinfo "160000,$AMD_LLAMA,third_party/llama.cpp"
+          git update-index --cacheinfo "160000,$AMD_HRX,third_party/hrx-system"
+          git commit -q -m "Bump HRX: llama.cpp ${AMD_LLAMA:0:12}, hrx-system ${AMD_HRX:0:12} (AMD's tested pair)"
+          git push -q origin "$branch"
+          gh pr create --base main --head "$branch" \
+            --title "Bump HRX: llama.cpp ${AMD_LLAMA:0:12}, hrx-system ${AMD_HRX:0:12}" \
+            --body "Moves both submodules to the pair ROCm/ggml-staging-automation now pins.
+
+          | | from | to |
+          |---|---|---|
+          | third_party/llama.cpp (1bit-MONSTER/llama.cpp 1bit/hrx-vulkan) | \`${OURS_LLAMA:0:12}\` | \`${AMD_LLAMA:0:12}\` |
+          | third_party/hrx-system (ROCm/hrx-system) | \`${OURS_HRX:0:12}\` | \`${AMD_HRX:0:12}\` |
+
+          CI here builds without HRX. Before merging, on Strix Halo:
+          \`cmake -B build -G Ninja -DONEBIT_HRX=ON && cmake --build build --target onebit && tests/hrx_lemonade_e2e.sh build/1bit build/hrx/llama/bin/llama-server\`"
diff --git a/.gitmodules b/.gitmodules
@@ -13,17 +13,15 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 #
-# HRX + Vulkan (docs/hrx.md). Only needed with -DONEBIT_HRX=ON.
-[submodule "third_party/hrx"]
-	path = third_party/hrx
-	url = https://github.com/ROCm/hrx.git
-	shallow = true
+# HRX + Vulkan (docs/hrx.md). Only needed with -DONEBIT_HRX=ON. The pair is AMD's
+# tested integration (ROCm/ggml-staging-automation); .github/workflows/bump-hrx.yml keeps it current.
 [submodule "third_party/hrx-system"]
 	path = third_party/hrx-system
 	url = https://github.com/ROCm/hrx-system.git
+	branch = main
 	shallow = true
 [submodule "third_party/llama.cpp"]
 	path = third_party/llama.cpp
-	url = https://github.com/bong-water-water-bong/llama.cpp.git
-	branch = 1bit-engine/hrx-vulkan
+	url = https://github.com/1bit-MONSTER/llama.cpp.git
+	branch = 1bit/hrx-vulkan
 	shallow = true
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -42,12 +42,12 @@ target_link_libraries(onebit PRIVATE lemonade-server-core)
 # ── Step 2: HRX + Vulkan in one llama.cpp build (docs/hrx.md) ─────────────
 # Off by default: it needs TheRock/ROCm and the three HRX submodules, and it
 # only runs on AMD GPUs. `1bit lemonade` then uses this llama-server for both
-# the llamacpp-hrx recipe (HRX20) and the llamacpp recipe's Vulkan backend.
-option(ONEBIT_HRX "Build HRX2 + Vulkan llama-server and wire it into Lemonade" OFF)
+# the llamacpp-hrx recipe (HRX0) and the llamacpp recipe's Vulkan backend.
+option(ONEBIT_HRX "Build HRX + Vulkan llama-server and wire it into Lemonade" OFF)
 if(ONEBIT_HRX)
     include(cmake/hrx.cmake)
     add_dependencies(onebit llama_hrx)
-    target_compile_definitions(onebit PRIVATE ONEBIT_HRX_SERVER="${ONEBIT_HRX_SERVER}")
+    target_compile_definitions(onebit PRIVATE ONEBIT_HRX_SERVER="${ONEBIT_HRX_SERVER}" ONEBIT_HRX_LIBHSA="${ONEBIT_HRX_LIBHSA}")
     add_test(NAME hrx_lemonade_e2e
              COMMAND ${CMAKE_SOURCE_DIR}/tests/hrx_lemonade_e2e.sh $<TARGET_FILE:onebit> ${ONEBIT_HRX_SERVER})
 endif()

diff --git a/app/main.cpp b/app/main.cpp
@@ -70,9 +70,9 @@ void usage(FILE* out) {
 }
 
 #ifdef ONEBIT_HRX_SERVER
-// Points Lemonade at this build's llama-server (HRX2 + Vulkan in one binary,
-// docs/hrx.md): the llamacpp-hrx recipe runs Q4NX on HRX20, and the llamacpp
-// recipe runs GGUF on Vulkan0, the faster device for standard quants. The
+// Points Lemonade at this build's llama-server (HRX + Vulkan in one binary,
+// docs/hrx.md): the llamacpp-hrx recipe runs on HRX0 (AMD's ggml-hrx), and the
+// llamacpp recipe runs GGUF on Vulkan0, the faster device for standard quants. The
 // config arrives merged with Lemonade's defaults, so only default values
 // ("builtin", "auto", unset) are replaced; anything else the user set wins.
 void use_engine_hrx_build(nlohmann::json& config) {
@@ -82,10 +82,14 @@ void use_engine_hrx_build(nlohmann::json& config) {
         if (!sec.contains(key) || sec[key] == "" || sec[key] == default_value) sec[key] = value;
     };
     set_default("hrx", "hrx_bin", "builtin", ONEBIT_HRX_SERVER);
-    set_default("hrx", "device", "", "HRX20");
+    set_default("hrx", "device", "", "HRX0");
     set_default("llamacpp", "vulkan_bin", "builtin", ONEBIT_HRX_SERVER);
     set_default("llamacpp", "backend", "auto", "vulkan");
     set_default("llamacpp", "device", "", "Vulkan0");
+    // HRX dlopens the HSA runtime; the distro's rejects gfx1151's PM4-emulation
+    // probe and HRX registers no device. llama-server inherits this environment,
+    // so point it at TheRock's unless the user already chose one.
+    setenv("IREE_HAL_AMDGPU_LIBHSA_PATH", ONEBIT_HRX_LIBHSA, /*overwrite=*/0);
 }
 #endif
 

diff --git a/cmake/hrx-config.cmake.in b/cmake/hrx-config.cmake.in
diff --git a/cmake/hrx.cmake b/cmake/hrx.cmake
@@ -13,114 +13,55 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-# HRX + Vulkan in one llama.cpp build (docs/hrx.md). Three pinned sources,
-# each cloned from its submodule into the build tree and patched there, so
-# the submodules stay clean:
+# HRX + Vulkan in one llama.cpp build (docs/hrx.md). Two pinned sources, the pair
+# AMD's integration repo (ROCm/ggml-staging-automation) builds and tests together:
 #
-#   hrx_runtime  ROCm/hrx 0bc22fb + patches/hrx/0001..0003: libhrx, libloomc,
-#                the Loom AMDGPU binding, assembled into ${ONEBIT_HRX_PREFIX}
-#   loom_link    ROCm/hrx-system 6743075f + patches/hrx/0002: the loom-link the
-#                fork's kernel catalog is linked with (the only version that
-#                still has --mode=selective; its artifacts are byte-identical
-#                to the working build's)
-#   llama_hrx    bong-water-water-bong/llama.cpp 1bit-engine/hrx-vulkan with
-#                GGML_HRX2 and GGML_VULKAN: one llama-server exposing HRX20
-#                and Vulkan0
+#   third_party/llama.cpp   1bit-MONSTER/llama.cpp 1bit/hrx-vulkan: AMD's
+#                           hrx-graph-develop-v2 (ggml-hrx) on ggml-org llama.cpp
+#   third_party/hrx-system  ROCm/hrx-system: libhrx, loomc and the Loom tools
+#
+# llama.cpp's ggml-hrx builds hrx-system itself when given HRX_SOURCE_DIR, with
+# the compilers of this build, so the only inputs are the two trees and TheRock.
+# IREE_ROCM_PATH is deliberately not passed: it switches hrx-system to "package"
+# mode, which requires TheRock's own aqlprofile-sdk headers (absent from our
+# /opt/rocm-therock); unset, hrx-system fetches its pinned HSA/AQL headers.
+# .github/workflows/bump-hrx.yml moves both pins when AMD moves its pair.
 include(ExternalProject)
-find_package(Git REQUIRED)
 
-set(ONEBIT_HRX_TOOLCHAIN "/opt/rocm-therock" CACHE PATH "ROCm/TheRock root whose amdclang builds the HRX runtime")
-set(ONEBIT_HRX_PREFIX "${CMAKE_BINARY_DIR}/hrx/prefix")
-set(_hrx_patches "${CMAKE_SOURCE_DIR}/patches/hrx")
+set(ONEBIT_HRX_TOOLCHAIN "/opt/rocm-therock" CACHE PATH "ROCm/TheRock root: its amdclang builds llama.cpp and HRX")
 
-foreach(_sub hrx hrx-system llama.cpp)
-    if(NOT EXISTS "${CMAKE_SOURCE_DIR}/third_party/${_sub}/.git")
+foreach(_sub hrx-system llama.cpp)
+    if(NOT EXISTS "${CMAKE_SOURCE_DIR}/third_party/${_sub}/CMakeLists.txt")
         message(FATAL_ERROR "ONEBIT_HRX needs third_party/${_sub}: git submodule update --init third_party/${_sub}")
     endif()
 endforeach()
 
-# Clone a submodule into the build tree at its pinned commit, then apply patches.
-function(_hrx_fetch_command out submodule)
-    set(_src "${CMAKE_SOURCE_DIR}/third_party/${submodule}")
-    execute_process(COMMAND ${GIT_EXECUTABLE} -C "${_src}" rev-parse HEAD
-                    OUTPUT_VARIABLE _rev OUTPUT_STRIP_TRAILING_WHITESPACE COMMAND_ERROR_IS_FATAL ANY)
-    set(_cmd ${CMAKE_COMMAND} -E rm -rf <SOURCE_DIR>
-        COMMAND ${GIT_EXECUTABLE} clone --quiet --shared --no-checkout "${_src}" <SOURCE_DIR>
-        COMMAND ${GIT_EXECUTABLE} -C <SOURCE_DIR> checkout --quiet ${_rev})
-    foreach(_patch ${ARGN})
-        list(APPEND _cmd COMMAND ${GIT_EXECUTABLE} -C <SOURCE_DIR> apply "${_hrx_patches}/${_patch}")
-    endforeach()
-    set(${out} ${_cmd} PARENT_SCOPE)
-endfunction()
-
-# The HRX build configuration of the working Strix Halo setup, verbatim.
-set(_hrx_flags
-    -DCMAKE_C_COMPILER=${ONEBIT_HRX_TOOLCHAIN}/bin/amdclang
-    -DCMAKE_CXX_COMPILER=${ONEBIT_HRX_TOOLCHAIN}/bin/amdclang++
-    -DCMAKE_ASM_COMPILER=${ONEBIT_HRX_TOOLCHAIN}/bin/amdclang
-    -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release
-    -DHRX_ENABLE_IPO=OFF -DHRX_INSTALL_TESTS=OFF
-    -DIREE_ALLOCATOR_SYSTEM=libc -DIREE_BUILD_BENCHMARKS=OFF -DIREE_BUILD_DOCS=OFF
-    -DIREE_BUILD_PYTHON_BINDINGS=OFF -DIREE_BUILD_SAMPLES=OFF -DIREE_BUILD_TESTS=OFF
-    -DIREE_DEPENDENCY_MODE=pinned -DIREE_DEV_MODE=OFF -DIREE_ENABLE_ASSERTIONS=OFF
-    -DIREE_ENABLE_LIBBACKTRACE=OFF -DIREE_ENABLE_LLD=OFF -DIREE_ENABLE_POSITION_INDEPENDENT_CODE=ON
-    -DIREE_ENABLE_RUNTIME_TRACING=OFF -DIREE_ENABLE_WERROR_FLAG=OFF
-    -DIREE_HAL_AMDGPU_DEVICE_BINARY_BUILD_MODE=prebuilt -DIREE_HAL_AMDGPU_LIBHSA_STATIC=OFF
-    "-DIREE_HAL_AMDGPU_TARGETS=gfx9-generic|gfx90a|gfx9-4-generic|gfx10-1-generic|gfx10-3-generic|gfx11-generic|gfx12-generic"
-    -DIREE_HAL_DRIVER_AMDGPU=ON -DIREE_HAL_DRIVER_DEFAULTS=OFF -DIREE_HAL_DRIVER_TASK=ON
-    -DIREE_HAL_DRIVER_LOCAL_TASK=ON -DIREE_HAL_DRIVER_VULKAN=OFF -DIREE_HAL_DRIVER_WEBGPU=OFF
-    -DIREE_HAL_EXECUTABLE_LOADER_DEFAULTS=ON -DIREE_VISIBILITY_HIDDEN=ON
-    -DLIBHRX_BUILD_CTS=OFF -DLIBHRX_BUILD_CUDA_BINDING=OFF -DLIBHRX_BUILD_HIP_BINDING=ON
-    -DLIBHRX_BUILD_PASSTHROUGH=ON
-    -DLOOM_BUILD=ON -DLOOM_EMIT_AMDGPU=OFF -DLOOM_EMIT_LLVMIR=OFF -DLOOM_EMIT_SPIRV=OFF -DLOOM_EMIT_WASM=OFF
-    -DLOOM_EXECUTE_DEFAULTS=ON -DLOOM_EXECUTE_IREE_HAL=ON -DLOOM_IMPORT_MLIR=OFF -DLOOM_IMPORT_TILELANG=OFF
-    -DLOOM_TARGET_AMDGPU=ON -DLOOM_TARGET_AMDGPU_TARGETS=loom_defaults -DLOOM_TARGET_ARCH_AMDGPU=OFF
-    -DLOOM_TARGET_ARCH_LLVMIR=OFF -DLOOM_TARGET_ARCH_SPIRV=OFF -DLOOM_TARGET_ARCH_WASM=OFF
-    -DLOOM_TARGET_ARCH_X86=OFF -DLOOM_TARGET_DEFAULTS=ON -DLOOM_TARGET_LLVMIR=ON -DLOOM_TARGET_SPIRV=ON
-    -DLOOM_TARGET_WASM=OFF -DLOOM_TARGET_X86=ON)
-
-_hrx_fetch_command(_fetch_runtime hrx
-    0001-libhrx-skip-install-export.patch
-    0002-loomc-drop-export-restriction.patch
-    0003-amdgpu-tolerate-missing-pm4-emulation-probe.patch)
-ExternalProject_Add(hrx_runtime
-    PREFIX ${CMAKE_BINARY_DIR}/hrx/runtime
-    DOWNLOAD_COMMAND ${_fetch_runtime}
-    LIST_SEPARATOR |
-    CMAKE_GENERATOR Ninja
-    CMAKE_ARGS ${_hrx_flags} -DLIBHRX_BUILD=ON
-    # Only what llama.cpp's ggml-hrx2 links; the full tree has unrelated targets.
-    BUILD_COMMAND ${CMAKE_COMMAND} --build <BINARY_DIR> --target
-        libhrx_src_libhrx_hrx loomc_shared loom_binding_c_target_amdgpu_amdgpu.objects
-    INSTALL_COMMAND ${CMAKE_COMMAND} -DSRC=<SOURCE_DIR> -DBIN=<BINARY_DIR> -DPREFIX=${ONEBIT_HRX_PREFIX}
-        -DAR=${CMAKE_AR} -DCONFIG_IN=${CMAKE_SOURCE_DIR}/cmake/hrx-config.cmake.in
-        -P ${CMAKE_SOURCE_DIR}/cmake/hrx_assemble_prefix.cmake
-    USES_TERMINAL_BUILD ON)
-
-_hrx_fetch_command(_fetch_system hrx-system 0002-loomc-drop-export-restriction.patch)
-ExternalProject_Add(loom_link
-    PREFIX ${CMAKE_BINARY_DIR}/hrx/loom-link
-    DOWNLOAD_COMMAND ${_fetch_system}
-    LIST_SEPARATOR |
-    CMAKE_GENERATOR Ninja
-    CMAKE_ARGS ${_hrx_flags} -DLIBHRX_BUILD=OFF
-    BUILD_COMMAND ${CMAKE_COMMAND} --build <BINARY_DIR> --target loom_tools_loom-link_loom-link
-    INSTALL_COMMAND ${CMAKE_COMMAND} -E copy <BINARY_DIR>/loom/src/loom/tools/loom-link/loom-link
-        ${ONEBIT_HRX_PREFIX}/bin/loom-link
-    USES_TERMINAL_BUILD ON)
+# HRX dlopens the HSA runtime. The distro's libhsa-runtime64 rejects the
+# HSA_AMD_AGENT_INFO_PM4_EMULATION probe on gfx1151, so HRX registers no device;
+# TheRock's answers it. `1bit lemonade` passes this path to llama-server as
+# IREE_HAL_AMDGPU_LIBHSA_PATH (docs/hrx.md).
+file(GLOB_RECURSE _hrx_libhsa "${ONEBIT_HRX_TOOLCHAIN}/*/libhsa-runtime64.so.1")
+list(SORT _hrx_libhsa)
+list(GET _hrx_libhsa 0 _hrx_libhsa_first)
+if(NOT _hrx_libhsa_first)
+    message(FATAL_ERROR "ONEBIT_HRX: no libhsa-runtime64.so.1 under ONEBIT_HRX_TOOLCHAIN=${ONEBIT_HRX_TOOLCHAIN}")
+endif()
+set(ONEBIT_HRX_LIBHSA "${_hrx_libhsa_first}" CACHE FILEPATH "TheRock HSA runtime HRX loads (IREE_HAL_AMDGPU_LIBHSA_PATH)")
+message(STATUS "ONEBIT_HRX: HSA runtime ${ONEBIT_HRX_LIBHSA}")
 
 set(ONEBIT_HRX_SERVER "${CMAKE_BINARY_DIR}/hrx/llama/bin/llama-server")
 ExternalProject_Add(llama_hrx
     SOURCE_DIR ${CMAKE_SOURCE_DIR}/third_party/llama.cpp
     BINARY_DIR ${CMAKE_BINARY_DIR}/hrx/llama
     DOWNLOAD_COMMAND ""
     CMAKE_GENERATOR Ninja
-    CMAKE_ARGS -DCMAKE_BUILD_TYPE=Release -Dhrx_DIR=${ONEBIT_HRX_PREFIX}/cmake/hrx
-        -DGGML_HRX2=ON -DGGML_VULKAN=ON -DGGML_CUDA=OFF -DGGML_HIP=OFF -DGGML_NATIVE=ON -DGGML_CPU=ON
+    CMAKE_ARGS -DCMAKE_BUILD_TYPE=Release
+        -DCMAKE_C_COMPILER=${ONEBIT_HRX_TOOLCHAIN}/bin/amdclang
+        -DCMAKE_CXX_COMPILER=${ONEBIT_HRX_TOOLCHAIN}/bin/amdclang++
+        -DHRX_SOURCE_DIR=${CMAKE_SOURCE_DIR}/third_party/hrx-system
+        -DGGML_HRX=ON -DGGML_VULKAN=ON -DGGML_CUDA=OFF -DGGML_HIP=OFF -DGGML_NATIVE=ON -DGGML_CPU=ON
         -DLLAMA_BUILD_SERVER=ON -DLLAMA_CURL=OFF
-        -DGGML_HRX2_LOOM_LINK_EXECUTABLE=${ONEBIT_HRX_PREFIX}/bin/loom-link
     BUILD_COMMAND ${CMAKE_COMMAND} --build <BINARY_DIR> --target llama-server llama-bench
     INSTALL_COMMAND ""
     BUILD_ALWAYS ON
-    DEPENDS hrx_runtime loom_link
     USES_TERMINAL_BUILD ON)