onnx EP NPU leg does not compile on main since #869: q4nx_elf_arg_sizes is called without its declaration #874

Activity

1. bong-water-water-bong

   added

   bugSomething isn't workingSomething isn't working

   npuThe NPU (XDNA2) forward/serve routeThe NPU (XDNA2) forward/serve route

   on Nov 3, 2026
2. bong-water-water-bong commented on Nov 3, 2026

   

   bong-water-water-bong

   on Nov 3, 2026

   CollaboratorAuthor

   More actions

   Reproduced on tonight's main (23d6972d9), with the issue's own command, in a clean worktree at that sha:

   $ cd ~/wt/iss874   # origin/main, 23d6972d
   $ g++-16 -DONEBIT_NPU_RUNTIME_OWN=1 -Inpu/forward/engine-src -Inpu -Inpu/forward/include -Inpu/ort \
            -std=c++26 -O1 -fsyntax-only npu/ort/onebit_npu_gemm.cpp
   npu/forward/engine-src/npu_engine_i8ctx_inc.h:219:13: error: ‘q4nx_elf_arg_sizes’ was not declared in this scope
     219 |         if (q4nx_elf_arg_sizes(raw.data(), raw.size(), declared) < 0) return true;   // not our shape
   

   Both halves of the report check out at the pin:

   • the declaration is only at npu/forward/include/q4nx_pack.h:257 (int q4nx_elf_arg_sizes(const void*, size_t, uint64_t[3]);)
     and npu/forward/engine-src/ contains no include of that header;
   • .github/workflows/ci.yml sets no ONEBIT_EP_NPU / ONEBIT_NPU — the only cmake invocation is
     cmake -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_C{,_XX}_COMPILER=gcc-16 and the only target built is onebit
     (:64,70), while the option itself lives at onnx/ep/CMakeLists.txt:27 and needs -DONEBIT_NPU_RT_DIR=<dir with npu_rt.cpp>
     (:42). So green CI on a red main for both NPU legs — the test-backend-ops (fork pin 40c4ccb): a run that executes zero cases exits 0 and prints OK #657 class, second instance.

   The proposed fix, tested rather than asserted

   -include npu/forward/include/q4nx_pack.h on the same command → rc 0. So the include form works for the EP leg, and it is
   the right one (the header is extern "C"-guarded, and a bare forward declaration would have to be kept in sync by hand).

   One thing to widen before writing it: there are three consumers of that inc header, not two

   git grep -ln npu_engine_i8ctx_inc.h -- .
   npu/forward/src/q4nx_forward.cpp          <- includes q4nx_pack.h, compiles
   npu/ort/onebit_npu_gemm.cpp               <- this issue
   tools/npu-arena/runner/runner.cpp         <- includes only "rt.h": SAME BREAK, and no CI job builds it either
   onnx/ep/CMakeLists.txt                    <- lists the header as a source
   

   Fixing only the EP leg leaves the arena runner broken in the same way; the include belongs in the header (which fixes both),
   and the arena runner is worth one -fsyntax-only line too if you want the class closed rather than this instance.

   CI gap — the cheap version

   A syntax-only job costs one step and no device: g++-16 -std=c++26 -fsyntax-only -Inpu/forward/engine-src -Inpu -Inpu/forward/include -Inpu/ort -DONEBIT_NPU_RUNTIME_OWN=1 npu/ort/*.cpp npu/forward/src/q4nx_forward.cpp would have caught
   this in the second between #869's merge and now, without npu_rt sources or an NPU. The full ONEBIT_EP_NPU=ON build needs
   the private runtime directory, so it is not a CI candidate; the syntax job is.

   Owner pi-ogaload (author of #869, the commit that introduced the call), bucket A, one-line fix + one CI step.
   #853 and #862 are merged, so their ctest entries are the ones currently unreproducible as builds on main — this is the
   reason to land it before the morning.
3. bong-water-water-bong

   closed this as completedin #878on Nov 3, 2026
4. bong-water-water-bong commented on Nov 3, 2026

   

   bong-water-water-bong

   on Nov 3, 2026

   CollaboratorAuthor

   More actions

   Closing note (issue keeper): closed by #878 → cc526384b — and the fix is in the header, which is where it had to be

   #878's body ends Closes #874, so GitHub closed this at 10:35:01Z (merge 10:34:59Z, squash). Verified device-free at 2fc5c4fd (origin/main):

   1. the declaration is reachable from all three consumers: npu/forward/engine-src/npu_engine_i8ctx_inc.h:39 now includes q4nx_pack.h with the comment naming args_fit_elf(), so npu/forward/src/q4nx_forward.cpp, npu/ort/onebit_npu_gemm.cpp and tools/npu-arena/runner/runner.cpp inherit it rather than each adding an include.
   2. g++-16 -std=c++20 -fsyntax-only -I npu/forward/include -I npu/forward/engine-src npu/ort/onebit_npu_gemm.cpp (with an empty stub rt.h, since the real one is the private runtime): 0 errors mentioning q4nx_elf_arg_sizes. The only errors left are the stub's (:52, xrt), i.e. past the point this issue was about. I could not compile the leg further on this box and am not claiming it.

   Residual, on the board, not held here: .github/workflows/ci.yml still sets no ONEBIT_EP_NPU/ONEBIT_NPU, so this configuration can go red on main with green CI again — exactly what happened. #878 deliberately did not add the job because both configurations need the private rt.h/npu_rt.h; a -fsyntax-only job needs a stub runtime header first. That is the #657 class, second instance, and it belongs to the CI gap list rather than to this row.
5. bong-water-water-bong

   added 4 commits that reference this issue on Nov 3, 2026

   ci: compile npu/forward's forward C++ as part of `onebit`, which CI b…

   ...

   ae0ee19

   ci: compile and link the ONNX EP's NPU leg, so a broken ONEBIT_EP_NPU…

   ...

   f2d19ee

   onnx EP: the caller's NPU tree must win the include order (Copilot on #…

   ...

   e76573e

   ci: compile and link the ONNX EP's NPU leg, so a broken ONEBIT_EP_NPU…

   ...

   14fcce4