[bong-water-water-bong]

Root cause

.github/workflows/ci.yml:63-64 configures without -DONEBIT_NPU, so the build job compiles none of npu/forward's forward: npu/forward/src/q4nx_forward.cpp is only in the ONEBIT_NPU branch (CMakeLists.txt:802). Two neighbouring C sources in that directory are compiled, by tests — model.c (:641), q4nx_pack.c (:529, :667) — which is why the hole looked smaller than it is: the file that carries the math, and all five #808 A1 defects plus #874's missing q4nx_elf_arg_sizes declaration, was the one nobody built.

Why not "turn ONEBIT_NPU on", and not a stub either

ONEBIT_NPU needs the private NPU runtime (ONEBIT_NPU_RT_LIB/_RT_DIR), which the CI image does not have. But compiling it needs no runtime: the in-tree npu/rt.h is an alias of npu_rt_api.h behind ONEBIT_NPU_RUNTIME_OWN — no XRT, no private header, no stub that could drift from the real interface. (My own earlier note on this board said a stub rt.h was needed; that was wrong, and this PR is the measurement that says so.)

The change

• onebit_npu_forward_syntax: an OBJECT library over npu/forward/src/q4nx_forward.cpp with the real target's include dirs (npu/forward/include, npu/forward/engine-src, npu), ONEBIT_NPU_RUNTIME_OWN, nlohmann_json and the same guarded OpenMP flags — so #pragma omp clauses are parsed exactly as in onebit_npu_forward. It links nothing.
• add_dependencies(onebit onebit_npu_forward_syntax), so the existing CI build job (cmake --build build --target onebit …) compiles it.

No workflow file is touched: this box's gh token carries repo only, no workflow scope. If a lane that has that scope prefers it, adding onebit_npu_forward_syntax to ci.yml's explicit target list is a one-line follow-up; it would merely make the step visible by name in the log.

Measured (this tree, Release, gcc-16, on strixhalo)

configure                                   rc 0
ninja -n onebit                             -> [23/27] Building CXX object
                            CMakeFiles/onebit_npu_forward_syntax.dir/npu/forward/src/q4nx_forward.cpp.o
clean tree, ninja onebit_npu_forward_syntax  -> object built, rc 0
#808's rebase defect injected at :1894      -> FAILED:
  npu/forward/src/q4nx_forward.cpp:1894:77: error: invalid conversion from 'int' to
  'const char*' [-fpermissive]   (initializing argument 8 of `npu_gemm(I8Ctx&, const float*, int,
  int, const float*, std::vector<float>&, int, const char*, int)`)

The negative control is the defect verbatim: npu_gemm(..., 0, "attn_qkv", g->real_k) → "attn_qkv" replaced by g->real_k, which is what #808's rebase did at 16 call sites (7 of them that hard error). Restored before commit; the only files in this PR are CMakeLists.txt.

What this does not do

The same hole exists for the ONNX EP leg (ONEBIT_EP_NPU off — the #874 class) and it is not fixed here; npu/ort/onebit_npu_gemm.cpp still is not compiled by CI. That is a second, separate change.

closes #898

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 808 - Partially compliant Compliant requirements: Split the XDNA launch layer behind NpuTarget interface Move device and launch layer code to npu/target/xdna/ directory Extract NpuTarget interface in npu/target/target.h Update q4nx_forward.cpp to use npu_gemm() calls instead of direct kernel launches Non-compliant requirements: Ensure no numerics or kernel changes in this phase (this is a mechanical split, so no changes expected) Requires further human verification: Verification that the interface is correctly implemented and all launch calls are properly routed through NpuTarget 874 - Partially compliant Compliant requirements: Fix compilation error in ONNX EP NPU leg where q4nx_elf_arg_sizes is called without declaration Non-compliant requirements: None Requires further human verification: None 898 - Partially compliant Compliant requirements: Add a compile-only job that builds npu/forward/src/q4nx_forward.cpp to catch C++ errors in CI Ensure that changes to npu/forward are compiled in CI Non-compliant requirements: None Requires further human verification: None

⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Missing OpenMP Include The find_package(OpenMP) call in CMakeLists.txt may not be sufficient to ensure OpenMP is properly enabled for the onebit_npu_forward_syntax library. If OpenMP is not available or not properly configured in the build environment, the #pragma omp clauses in the source code may not be processed correctly. This could lead to subtle performance issues or incorrect parallelization behavior. find_package(OpenMP) if(OpenMP_CXX_FOUND) target_link_libraries(onebit_npu_forward_syntax PRIVATE OpenMP::OpenMP_CXX) endif()

[bong-water-water-bong]

VERDICT: READY at b06ab9d21 — reproduced the whole premise here, in a CI-shaped configure, including the negative control.

Reproduced (fresh configure exactly as ci.yml does it, i.e. without -DONEBIT_NPU):

cmake -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_C_COMPILER=gcc-16 -DCMAKE_CXX_COMPILER=g++-16   rc 0
ninja -n onebit    → [23/27] Building CXX object CMakeFiles/onebit_npu_forward_syntax.dir/npu/forward/src/q4nx_forward.cpp.o
ninja onebit_npu_forward_syntax → built, rc 0, q4nx_forward.cpp.o 1,040,952 B

So add_dependencies(onebit …) really does put it on the path of the target the Build step names, with no workflow change.

Negative control, the #808 defect verbatim — :1894 "attn_qkv" → g->real_k, the substitution that broke 16 call sites in #808's rebase:

FAILED: [code=1] CMakeFiles/onebit_npu_forward_syntax.dir/npu/forward/src/q4nx_forward.cpp.o
npu/forward/src/q4nx_forward.cpp:1894:77: error: invalid conversion from 'int' to 'const char*' [-fpermissive]

Restored with git checkout -- afterwards; the tree I built from is the PR head.

The "it was the only uncovered file" claim, checked rather than accepted — the four C siblings are each compiled through the add_dependencies workaround, so the .cpp really was the hole:
q4nx_dequant.c → npu_dequant_bf16_test (:691, bound at :695), q4nx_semantics.c → npu_rmsnorm_offset_test (:701, bound at :706), q4nx_pack.c → npu_elf_args_test/npu_design_grid_test (:547/:685), model.c → npu_forward_loader_test (:659).

Two nits, neither a gate:

1. The syntax target compiles a different branch of npu/rt.h than the shipped target. ONEBIT_NPU_RUNTIME_OWN is defined for it (CMakeLists.txt:370) and not for onebit_npu_forward (:815-839), and npu/rt.h:26 is #if !defined(ONEBIT_NPU_RUNTIME_OWN). That is what makes it buildable here without XRT, and it is right — but it means CI typechecks q4nx_forward.cpp against the in-tree API alias, not against the private runtime's header, so a signature drift between the two stays invisible. One clause in the comment block would stop someone reading this as "CI validates the file against the real runtime".
2. NPU_INFER_DEBUG_DUMPS stays uncompiled (4 sites in this file; the real target defines it only under ONEBIT_NPU_DEBUG_DUMPS). Fine, but it is the one region of the file the new target still does not see.

Agreed with the body on the ONNX EP leg: npu/ort/onebit_npu_gemm.cpp is the same class of hole and is not covered here.