bong-water-water-bong · Sep 25, 2026 · Sep 25, 2026 · Sep 25, 2026
diff --git a/.github/workflows/bump-hrx.yml b/.github/workflows/bump-hrx.yml
@@ -15,20 +15,17 @@
 #
 # 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"). The engine pins 1bit/hrx-vulkan-patched: AMD's
-# commit plus our own commits on top. 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. rebases our commits onto AMD's pin and moves 1bit/hrx-vulkan-patched there,
-#      tagging the previous tip so the commit the engine pins stays reachable;
-#   3. opens a PR here moving both submodules.
-# A rebase conflict fails the run: rebase 1bit/hrx-vulkan-patched by hand.
-# GitHub-hosted CI builds that PR without HRX; run tests/serve_e2e.sh on
-# Strix Halo before merging it.
+# (docs/hrx.md): AMD-Ecosystem/llama.cpp (ggml-hrx) and ROCm/hrx-system, unchanged.
+# When AMD moves its pair, this opens a PR here moving both submodules.
+# GitHub-hosted CI builds that PR without HRX; run tests/serve_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.)
+# The private trees -DONEBIT_GPU_PRIVATE builds (1bit-MONSTER/gpu-kernels) are not
+# moved from here; that repository keeps its own pins.
+#
+# Uses the secret HRX_BUMP_TOKEN (Contents and Pull requests read/write on
+# 1bit-MONSTER/engine): a PR opened with the default GITHUB_TOKEN would not
+# start CI.
 name: bump-hrx
 
 on:
@@ -59,9 +56,8 @@ jobs:
       - uses: actions/checkout@v4
         with:
           token: ${{ secrets.HRX_BUMP_TOKEN }}
-          persist-credentials: true
 
-      - name: Read AMD's tested pair and ours
+      - name: Compare with AMD's tested pair
         id: pins
         env:
           GH_TOKEN: ${{ secrets.HRX_BUMP_TOKEN }}
@@ -71,63 +67,12 @@ jobs:
           amd_hrx=$(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 $3}')
-          # the AMD commit our patches sit on
-          ours_base=$(git ls-remote https://github.com/1bit-MONSTER/llama.cpp.git refs/heads/1bit/hrx-vulkan | awk '{print $1}')
           echo "AMD:  llama.cpp $amd_llama  hrx-system $amd_hrx"
-          echo "ours: llama.cpp $ours_llama (on AMD $ours_base)  hrx-system $ours_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=$ours_hrx"; echo "ours_base=$ours_base"
-            if [ "$amd_llama" = "$ours_base" ] && [ "$amd_hrx" = "$ours_hrx" ]; then echo "changed=false"; else echo "changed=true"; fi
-          } >> "$GITHUB_OUTPUT"
-
-      - name: Sync the fork and rebase our commits
-        id: fork
-        if: steps.pins.outputs.changed == 'true'
-        env:
-          GH_TOKEN: ${{ secrets.HRX_BUMP_TOKEN }}
-          AMD_LLAMA: ${{ steps.pins.outputs.amd_llama }}
-          OURS_BASE: ${{ steps.pins.outputs.ours_base }}
-        run: |
-          set -euo pipefail
-          # master mirrors ggml-org; 1bit/hrx-vulkan carries no commits of its own, so it
-          # moves to AMD's pin as long as that pin descends from it (no force needed).
-          gh api -X POST repos/1bit-MONSTER/llama.cpp/merge-upstream -f branch=master > /dev/null
-          git init -q fork && cd fork
-          git remote add fork "https://x-access-token:${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 "$AMD_LLAMA:refs/heads/1bit/hrx-vulkan"
-
-          # our commits: OURS_BASE..1bit/hrx-vulkan-patched, replayed onto AMD's pin, in a clone
-          # whose only (partial) remote is the fork, which now holds AMD's pin too
-          cd .. && git init -q patched && cd patched
-          git remote add fork "https://x-access-token:${GH_TOKEN}@github.com/1bit-MONSTER/llama.cpp.git"
-          git fetch -q --filter=blob:none fork 1bit/hrx-vulkan
-          git fetch -q --filter=blob:none fork 1bit/hrx-vulkan-patched
-          patched=$(git rev-parse FETCH_HEAD)
-          rebased=$patched
-          if [ "$OURS_BASE" != "$AMD_LLAMA" ]; then
-            git checkout -q --detach "$patched"
-            if ! git -c user.name=hrx-bump -c user.email=hrx-bump@users.noreply.github.com rebase -q --onto "$AMD_LLAMA" "$OURS_BASE"; then
-              git rebase --abort || true
-              echo "::error::our commits on 1bit/hrx-vulkan-patched do not rebase onto AMD's ${AMD_LLAMA:0:12}; rebase them by hand"
-              exit 1
-            fi
-            rebased=$(git rev-parse HEAD)
-            # keep the commit the engine pins now reachable before moving the branch
-            git tag "patched-${patched:0:12}" "$patched"
-            git push -q fork "refs/tags/patched-${patched:0:12}"
-            git push -q --force-with-lease="refs/heads/1bit/hrx-vulkan-patched:$patched" fork "$rebased:refs/heads/1bit/hrx-vulkan-patched"
-          fi
-          {
-            echo "rebased=$rebased"
-            echo "patches<<EOF"; git log --format='- `%h` %s' "$AMD_LLAMA..$rebased"; echo "EOF"
+            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: Open the bump PR
@@ -136,8 +81,6 @@ jobs:
           GH_TOKEN: ${{ secrets.HRX_BUMP_TOKEN }}
           AMD_LLAMA: ${{ steps.pins.outputs.amd_llama }}
           AMD_HRX: ${{ steps.pins.outputs.amd_hrx }}
-          REBASED: ${{ steps.fork.outputs.rebased }}
-          PATCHES: ${{ steps.fork.outputs.patches }}
           OURS_LLAMA: ${{ steps.pins.outputs.ours_llama }}
           OURS_HRX: ${{ steps.pins.outputs.ours_hrx }}
         run: |
@@ -149,7 +92,7 @@ jobs:
           git config user.name "hrx-bump"
           git config user.email "hrx-bump@users.noreply.github.com"
           git switch -c "$branch"
-          git update-index --cacheinfo "160000,$REBASED,third_party/llama.cpp"
+          git update-index --cacheinfo "160000,$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"
@@ -159,11 +102,8 @@ jobs:
 
           | | from | to |
           |---|---|---|
-          | third_party/llama.cpp (1bit-MONSTER/llama.cpp 1bit/hrx-vulkan-patched) | \`${OURS_LLAMA:0:12}\` | \`${REBASED:0:12}\` (on AMD \`${AMD_LLAMA:0:12}\`) |
+          | third_party/llama.cpp (AMD-Ecosystem/llama.cpp) | \`${OURS_LLAMA:0:12}\` | \`${AMD_LLAMA:0:12}\` |
           | third_party/hrx-system (ROCm/hrx-system) | \`${OURS_HRX:0:12}\` | \`${AMD_HRX:0:12}\` |
 
-          Our commits, rebased onto AMD's pin:
-          ${PATCHES}
-
-          CI here builds without HRX. Before merging, on Strix Halo (and \`test-backend-ops -b HRX0\`):
+          CI here builds without HRX. Before merging, on Strix Halo:
           \`cmake -B build -G Ninja -DONEBIT_HRX=ON && cmake --build build --target onebit && for d in vulkan hrx; do tests/serve_e2e.sh build/1bit <Qwen3-0.6B Q4_K_M .gguf> \$d; done\`"
diff --git a/.github/workflows/bump-rocmfpx.yml b/.github/workflows/bump-rocmfpx.yml
@@ -13,18 +13,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 #
-# Keep third_party/llama.cpp-vulkan on upstream llama.cpp's latest release
+# Keeps third_party/llama.cpp-rocmfpx (the lean option, docs/lean.md) on upstream
+# charlie12345/ROCmFPX's main: once a week, when main has moved, it opens a PR here
+# moving the submodule. ROCmFPX publishes no releases, and its main moves many times a
+# day, so weekly. GitHub-hosted CI has no GPU: rerun docs/lean.md's table on Strix Halo
+# before merging.
 #
-# Keeps third_party/llama.cpp-rocmfpx (the lean option, docs/lean.md) current. The engine
-# pins 1bit/vulkan-rocmi4 on our fork 1bit-MONSTER/ROCmFPX: charlie12345/ROCmFPX's main plus
-# our commits (ROCmI4 on Vulkan). Once a week, when upstream main has moved past our base, it
-#   1. syncs the fork's main with upstream,
-#   2. rebases our commits onto upstream main and moves 1bit/vulkan-rocmi4 there (the old
-#      head is kept as a tag, rocmi4-<sha>),
-#   3. opens a PR here moving the submodule to the rebased head.
-# A rebase conflict fails the run: rebase 1bit/vulkan-rocmi4 by hand.
-# ROCmFPX publishes no releases, and its main moves many times a day, so weekly.
-# GitHub-hosted CI has no GPU: rerun docs/lean.md's table on Strix Halo before merging.
+# The private ROCmFPX tree -DONEBIT_GPU_PRIVATE builds (1bit-MONSTER/gpu-kernels) is not
+# moved from here; that repository keeps its own pin.
 #
 # Uses the secret HRX_BUMP_TOKEN (Contents and Pull requests read/write on
 # 1bit-MONSTER/engine): a PR opened with the default GITHUB_TOKEN would not start CI.
@@ -59,38 +55,18 @@ jobs:
         with:
           token: ${{ secrets.HRX_BUMP_TOKEN }}
 
-      - name: Rebase our commits onto ROCmFPX main
+      - name: Compare with ROCmFPX main
         id: pins
         env:
           GH_TOKEN: ${{ secrets.HRX_BUMP_TOKEN }}
         run: |
           set -euo pipefail
           upstream=$(gh api repos/charlie12345/ROCmFPX/commits/main --jq .sha)
           ours=$(git ls-tree HEAD third_party/llama.cpp-rocmfpx | awk '{print $3}')
-          git init -q /tmp/fork && cd /tmp/fork
-          git remote add up https://github.com/charlie12345/ROCmFPX.git
-          git remote add fork "https://x-access-token:${GH_TOKEN}@github.com/1bit-MONSTER/ROCmFPX.git"
-          git fetch -q --filter=blob:none up main
-          git fetch -q --filter=blob:none fork 1bit/vulkan-rocmi4
-          head=$(git rev-parse FETCH_HEAD)
-          base=$(git merge-base "$head" "$upstream")
-          echo "upstream ${upstream}, our base ${base}, our head ${head}, engine pin ${ours}"
-          if [ "$base" = "$upstream" ]; then
-            { echo "changed=false"; } >> "$GITHUB_OUTPUT"; exit 0
-          fi
-          gh api -X POST repos/1bit-MONSTER/ROCmFPX/merge-upstream -f branch=main > /dev/null
-          git checkout -q --detach "$head"
-          if ! git -c user.name=rocmfpx-bump -c user.email=rocmfpx-bump@users.noreply.github.com rebase -q --onto "$upstream" "$base"; then
-            git rebase --abort || true
-            echo "::error::our commits on 1bit/vulkan-rocmi4 do not rebase onto ROCmFPX ${upstream:0:12}; rebase them by hand"
-            exit 1
-          fi
-          rebased=$(git rev-parse HEAD)
-          git tag "rocmi4-${head:0:12}" "$head"
-          git push -q fork "refs/tags/rocmi4-${head:0:12}"
-          git push -q --force-with-lease="refs/heads/1bit/vulkan-rocmi4:$head" fork "$rebased:refs/heads/1bit/vulkan-rocmi4"
+          echo "upstream ${upstream}, ours ${ours}"
           {
-            echo "changed=true"; echo "upstream=$rebased"; echo "ours=$ours"; echo "base=$upstream"
+            echo "upstream=$upstream"; echo "ours=$ours"
+            if [ "$upstream" = "$ours" ]; then echo "changed=false"; else echo "changed=true"; fi
           } >> "$GITHUB_OUTPUT"
 
       - name: Open the bump PR
@@ -99,22 +75,21 @@ jobs:
           GH_TOKEN: ${{ secrets.HRX_BUMP_TOKEN }}
           UPSTREAM: ${{ steps.pins.outputs.upstream }}
           OURS: ${{ steps.pins.outputs.ours }}
-          BASE: ${{ steps.pins.outputs.base }}
         run: |
           set -euo pipefail
           branch="bump-rocmfpx/${UPSTREAM:0:12}"
           if git ls-remote --exit-code origin "refs/heads/$branch" > /dev/null; then
             echo "$branch already exists"; exit 0
           fi
-          count=$(gh api "repos/charlie12345/ROCmFPX/compare/${OURS}...${BASE}" --jq .ahead_by 2>/dev/null || echo "?")
+          count=$(gh api "repos/charlie12345/ROCmFPX/compare/${OURS}...${UPSTREAM}" --jq .ahead_by 2>/dev/null || echo "?")
           git config user.name "rocmfpx-bump"
           git config user.email "rocmfpx-bump@users.noreply.github.com"
           git switch -c "$branch"
           git update-index --cacheinfo "160000,$UPSTREAM,third_party/llama.cpp-rocmfpx"
-          git commit -q -m "Bump ROCmFPX (lean): 1bit/vulkan-rocmi4 ${UPSTREAM:0:12} on charlie12345/ROCmFPX ${BASE:0:12}"
+          git commit -q -m "Bump ROCmFPX (lean): charlie12345/ROCmFPX ${UPSTREAM:0:12}"
           git push -q origin "$branch"
           gh pr create --base main --head "$branch" \
             --title "Bump ROCmFPX (lean): ${UPSTREAM:0:12}" \
-            --body "Moves third_party/llama.cpp-rocmfpx from \`${OURS:0:12}\` to \`${UPSTREAM:0:12}\`: our commits (ROCmI4 on Vulkan) rebased onto ROCmFPX main \`${BASE:0:12}\` (${count} upstream commits).
+            --body "Moves third_party/llama.cpp-rocmfpx from \`${OURS:0:12}\` to ROCmFPX main \`${UPSTREAM:0:12}\` (${count} commits).
 
           CI here has no GPU. Before merging, on Strix Halo: build with \`-DONEBIT_LEAN=ON -DONEBIT_LEAN_ROCM=ON\`, run \`ctest -R serve_e2e_lean\`, and rerun the table in docs/lean.md."
diff --git a/.gitmodules b/.gitmodules
@@ -14,16 +14,17 @@
 # limitations under the License.
 #
 # 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.
+# tested integration (ROCm/ggml-staging-automation), unchanged; .github/workflows/bump-hrx.yml
+# keeps it current. -DONEBIT_GPU_PRIVATE builds other trees instead (docs/hrx.md).
 [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/1bit-MONSTER/llama.cpp.git
-	branch = 1bit/hrx-vulkan-patched
+	url = https://github.com/AMD-Ecosystem/llama.cpp.git
+	branch = hrx-graph-develop-v2
 	shallow = true
 [submodule "third_party/xdna-driver"]
 	path = third_party/xdna-driver
@@ -59,11 +60,13 @@
 	path = third_party/llama.cpp-vulkan
 	url = https://github.com/ggml-org/llama.cpp.git
 	shallow = true
+# The lean option (docs/lean.md): upstream charlie12345/ROCmFPX;
+# .github/workflows/bump-rocmfpx.yml keeps it on its main.
 [submodule "third_party/llama.cpp-rocmfpx"]
 	path = third_party/llama.cpp-rocmfpx
-	url = https://github.com/1bit-MONSTER/ROCmFPX.git
+	url = https://github.com/charlie12345/ROCmFPX.git
 	shallow = true
-	branch = 1bit/vulkan-rocmi4
+	branch = main
 [submodule "third_party/comfyui.cpp"]
 	path = third_party/comfyui.cpp
 	url = https://github.com/1bit-MONSTER/comfyui.cpp.git

diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -78,6 +78,24 @@ if(ONEBIT_VULKAN)
     endif()
 endif()
 
+# ── Private GPU trees (docs/hrx.md, "Private GPU build") ──────────────────
+# The public pins are AMD's ggml-hrx pair (ONEBIT_HRX) and upstream ROCmFPX (ONEBIT_LEAN).
+# ONEBIT_GPU_PRIVATE points at a checkout of the private gpu-kernels repository (branch
+# addons/gpu, submodules initialised); its addons/gpu/gpu.cmake names the llama.cpp,
+# hrx-system and ROCmFPX trees those builds use instead.
+set(ONEBIT_GPU_PRIVATE "" CACHE PATH "A gpu-kernels checkout whose trees ONEBIT_HRX and ONEBIT_LEAN build")
+set(ONEBIT_GPU_LLAMA_SOURCE "${CMAKE_SOURCE_DIR}/third_party/llama.cpp")
+set(ONEBIT_GPU_HRX_SYSTEM_SOURCE "${CMAKE_SOURCE_DIR}/third_party/hrx-system")
+set(ONEBIT_GPU_ROCMFPX_SOURCE "${CMAKE_SOURCE_DIR}/third_party/llama.cpp-rocmfpx")
+if(ONEBIT_GPU_PRIVATE)
+    if(NOT EXISTS "${ONEBIT_GPU_PRIVATE}/addons/gpu/gpu.cmake")
+        message(FATAL_ERROR "ONEBIT_GPU_PRIVATE=${ONEBIT_GPU_PRIVATE} has no addons/gpu/gpu.cmake")
+    endif()
+    include(${ONEBIT_GPU_PRIVATE}/addons/gpu/gpu.cmake)
+    message(STATUS "ONEBIT_GPU_PRIVATE: llama.cpp ${ONEBIT_GPU_LLAMA_SOURCE}, hrx-system ${ONEBIT_GPU_HRX_SYSTEM_SOURCE}, ROCmFPX ${ONEBIT_GPU_ROCMFPX_SOURCE}")
+    target_compile_definitions(onebit PRIVATE ONEBIT_GPU_PRIVATE)
+endif()
+
 # ── The lean option: ROCmFPX's formats (docs/lean.md) ─────────────────────
 # Off by default. Builds llama-server + llama-quantize from third_party/llama.cpp-rocmfpx
 # on Vulkan: `1bit serve --lean` runs ROCmFP4 files with it. ONEBIT_LEAN_ROCM adds the
@@ -118,9 +136,12 @@ if(ONEBIT_HRX)
             add_test(NAME serve_e2e_${_dev}
                      COMMAND ${CMAKE_SOURCE_DIR}/tests/serve_e2e.sh $<TARGET_FILE:onebit> ${ONEBIT_SERVE_TEST_GGUF} ${_dev})
         endforeach()
-        # Vulkan decode with prompt prefill on HRX over one shared KV cache (docs/hrx.md)
-        add_test(NAME serve_e2e_vulkan_prefill_hrx
-                 COMMAND ${CMAKE_SOURCE_DIR}/tests/serve_e2e.sh $<TARGET_FILE:onebit> ${ONEBIT_SERVE_TEST_GGUF} vulkan --prefill-device hrx)
+        # Vulkan decode with prompt prefill on HRX over one shared KV cache (docs/hrx.md);
+        # only the private trees have it
+        if(ONEBIT_GPU_PRIVATE)
+            add_test(NAME serve_e2e_vulkan_prefill_hrx
+                     COMMAND ${CMAKE_SOURCE_DIR}/tests/serve_e2e.sh $<TARGET_FILE:onebit> ${ONEBIT_SERVE_TEST_GGUF} vulkan --prefill-device hrx)
+        endif()
     endif()
 endif()
 

diff --git a/NOTICE b/NOTICE
@@ -52,7 +52,7 @@ Peano, llvm-aie (third_party/llvm-aie)
   Copyright the LLVM Project authors and Xilinx, Inc. / Advanced Micro Devices, Inc.
   License: Apache-2.0 WITH LLVM-exception
 
-llama.cpp (third_party/llama.cpp: 1bit-MONSTER/llama.cpp, a fork of
+llama.cpp (third_party/llama.cpp: AMD-Ecosystem/llama.cpp, a fork of
 ggml-org/llama.cpp carrying AMD's HRX integration; third_party/llama.cpp-vulkan:
 ggml-org/llama.cpp, an upstream release)
   https://github.com/ggml-org/llama.cpp

diff --git a/README.md b/README.md
@@ -43,7 +43,10 @@ serves each model behind an OpenAI-compatible API (`1bit serve`), whatever devic
 > upstream XDNA stack pinned ([docs/npu.md](docs/npu.md); its layer kernel is not yet built from
 > source), ZINC ([docs/zinc.md](docs/zinc.md)) and MLX ([docs/apple.md](docs/apple.md)). Experimental,
 > and closed source: Qwen3.6-35B-A3B on the NPU through a private add-on, parity against fp64 passes,
-> 16.3-16.5 tok/s decode ([docs/npu.md](docs/npu.md#private-routes)). Next is step 4,
+> 16.3-16.5 tok/s decode ([docs/npu.md](docs/npu.md#private-routes)). Our GPU kernel work is
+> closed source as well: the public build runs AMD's HRX pair and upstream ROCmFPX as they are,
+> and `-DONEBIT_GPU_PRIVATE` builds in our fixes (MoE models above 128 experts on HRX, the HRX
+> prefill split, ROCmI4 on Vulkan; [docs/hrx.md](docs/hrx.md#private-gpu-build)). Next is step 4,
 > the Laya router. The working engine is being ported from 1bit-MONSTER, our private development
 > repository; see [docs/PORTING.md](docs/PORTING.md).
 > Measured results are on the [wiki](https://github.com/1bit-MONSTER/engine/wiki). This repository