GPU: pin AMD's HRX pair and upstream ROCmFPX; our GPU kernel work behind -DONEBIT_GPU_PRIVATE

[bong-water-water-bong]

Our GPU kernel work (ggml-hrx fixes, the HRX prefill split, ROCmI4 on Vulkan) becomes closed source, like the NPU kernels (#85). The engine stops pinning our public forks.

Public pins

Submodule	was	now
third_party/llama.cpp	1bit-MONSTER/llama.cpp 1bit/hrx-vulkan-patched c075cc1	AMD-Ecosystem/llama.cpp hrx-graph-develop-v2 f1a0aca (AMD's pin, unchanged)
third_party/hrx-system	ROCm/hrx-system 51b1739	a351789 (AMD's current pair)
third_party/llama.cpp-rocmfpx	1bit-MONSTER/ROCmFPX 1bit/vulkan-rocmi4 b44bf74	charlie12345/ROCmFPX main fb08d7c

Hook: -DONEBIT_GPU_PRIVATE=<gpu-kernels checkout> includes addons/gpu/gpu.cmake from the private repo, which names the llama.cpp / hrx-system / ROCmFPX trees ONEBIT_HRX and ONEBIT_LEAN build instead. Without it, 1bit serve --prefill-device hrx exits with a message and serve_e2e_vulkan_prefill_hrx is not added.

Workflows: bump-hrx.yml and bump-rocmfpx.yml now only move the public pins (no fork sync, no rebase; HRX_BUMP_TOKEN needs only engine access). The private repo keeps its own pins; a bump flow there is still to do.

Docs: hrx.md (public vs private table, outcome level only; the "Fixed"/"Our patches" internals removed), lean.md, serve.md, PORTING.md, README status, NOTICE.

Verified on Strix Halo (tests/serve_e2e.sh, builds with -DONEBIT_HRX=ON -DONEBIT_LEAN=ON):

Build	Qwen3-0.6B Q4_K_M hrx	vulkan	vulkan --prefill-device hrx	Qwen3.6-35B-A3B Q8_0 hrx
public	PASS	PASS	refused (message)	FAIL (compute error, known)
-DONEBIT_GPU_PRIVATE	PASS	PASS	PASS	PASS

After merge, the fork branches 1bit-MONSTER/llama.cpp 1bit/hrx-vulkan-patched / 1bit/hrx-vulkan and 1bit-MONSTER/ROCmFPX 1bit/vulkan-rocmi4 are no longer used by the engine (private copies are in gpu-kernels).

🤖 Generated with Claude Code

[context7]

Docs7 for 1bit-monster/engine

Result	Status	Action
Deployment	➖ Not used	—
Content review	➖ Did not run. This site has no agent runs available this month. Wait for the monthly reset or check your Docs7 plan.	—

Commit 5f070bc

[github-actions]

PR Reviewer Guide 🔍

(Review updated until commit 5f070bc)

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 85 - Partially compliant Compliant requirements: The Qwen3.6-35B-A3B NPU route is now closed source and moved to a private add-on The engine keeps only a small hook for it Remove the third_party/OpenFlowLM-Next submodule and its NOTICE entry Remove scripts/build-lax.sh Remove npu/lax* Remove app/npu_lax* and the 1bit npu-lax command Remove tests/npu_lax* and tests/golden/npu_lax/ Remove docs/npu-lax.md Remove the ONEBIT_NPU_LAX* CMake options Remove the --npu-kernels, --npu-transport and --npu-snapshots flags Update docs to mention the route now has a short note instead Add npu/private_route.{h,cpp}: a registry of private NPU routes keyed by model_type, plus optional 1bit subcommands Add CMake option -DONEBIT_NPU_PRIVATE=<npu-kernels checkout> (needs -DONEBIT_NPU=ON) Add 1bit serve --npu-opt KEY=VALUE (in unified: --opt) Without the add-on, 1bit serve -m <Qwen3.6-35B-A3B dir> --device npu exits with a specific message Add tests/npu_private_route_test.cpp (ctest npu_private_route), added to the CI target list Update docs/npu.md, "Private routes" Non-compliant requirements: None Requires further human verification: None

⏱️ Estimated effort to review: 4 🔵🔵🔵🔵⚪

🧪 PR contains tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Runtime check for private GPU build The runtime check for the private GPU build (ONEBIT_GPU_PRIVATE) in app/serve.cpp is correctly implemented to prevent usage of --prefill-device hrx without the private build. However, the error message could be more specific about the expected format of the private checkout path to aid in troubleshooting. #ifndef ONEBIT_GPU_PRIVATE if (split) throw std::runtime_error("--prefill-device hrx is not part of this build; build with " "-DONEBIT_GPU_PRIVATE=<gpu-kernels checkout> (docs/hrx.md)"); #endif CMake logic for private sources The CMake logic in cmake/hrx.cmake now uses ONEBIT_GPU_LLAMA_SOURCE and ONEBIT_GPU_HRX_SYSTEM_SOURCE variables to determine which source trees to use for building. This change correctly supports the private build, but it's important to ensure that these variables are properly initialized and validated to prevent build errors. foreach(_src ${ONEBIT_GPU_HRX_SYSTEM_SOURCE} ${ONEBIT_GPU_LLAMA_SOURCE}) if(NOT EXISTS "${_src}/CMakeLists.txt") message(FATAL_ERROR "ONEBIT_HRX needs ${_src}: git submodule update --init --depth 1 in its repository") endif() endforeach() CMake logic for ROCmFPX source Similar to cmake/hrx.cmake, cmake/lean.cmake now uses ONEBIT_GPU_ROCMFPX_SOURCE to determine the ROCmFPX source tree. This change correctly supports the private build, but it's important to ensure that this variable is properly initialized and validated to prevent build errors. # ONEBIT_GPU_ROCMFPX_SOURCE: third_party/llama.cpp-rocmfpx, or the private tree with # -DONEBIT_GPU_PRIVATE (CMakeLists.txt). if(NOT EXISTS "${ONEBIT_GPU_ROCMFPX_SOURCE}/CMakeLists.txt") message(FATAL_ERROR "ONEBIT_LEAN needs ${ONEBIT_GPU_ROCMFPX_SOURCE}: git submodule update --init --depth 1 in its repository") endif() Private GPU build setup The CMakeLists.txt file correctly sets up the private GPU build by defining ONEBIT_GPU_PRIVATE and including the private gpu.cmake file. However, it's important to ensure that the private build is thoroughly tested to prevent regressions in the public build. # ── Private GPU trees (docs/hrx.md, "Private GPU build") ────────────────── # The public pins are AMD's ggml-hrx pair (ONEBIT_HRX) and upstream ROCmFPX (ONEBIT_LEAN). # ONEBIT_GPU_PRIVATE points at a checkout of the private gpu-kernels repository (branch # addons/gpu, submodules initialised); its addons/gpu/gpu.cmake names the llama.cpp, # hrx-system and ROCmFPX trees those builds use instead. set(ONEBIT_GPU_PRIVATE "" CACHE PATH "A gpu-kernels checkout whose trees ONEBIT_HRX and ONEBIT_LEAN build") set(ONEBIT_GPU_LLAMA_SOURCE "${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()

[github-actions]

Persistent review updated to latest commit 5f070bc

[github-actions]

Persistent review updated to latest commit 5f070bc