fix: revert "feat(ascend): op-norm-rope group — Swiglu, SiluAndMul, CausalSoftmax, RmsNorm, AddRmsNorm, ApplyRotaryPosEmb, RotaryEmbedding"

CMakeLists.txt

@@ -18,21 +18,12 @@ option(WITH_ASCEND "Enable Ascend backend" OFF)
 
 option(WITH_TORCH "Enable PyTorch C++ backend" OFF)
 
-# Custom `AscendC` kernels under `src/ascend/custom/`.  `ON` by default
-# so CI and routine dev builds always exercise `implementation_index=1/2`
-# for `RmsNorm` / `AddRmsNorm`.  Gated by `WITH_ASCEND` in
-# `src/CMakeLists.txt` — non-Ascend builds ignore it.  Pass
-# `-DBUILD_ASCEND_CUSTOM=OFF` to skip the `ccec` build on Ascend
-# machines where the custom kernels aren't needed.
-#
-# When `ON`, `src/CMakeLists.txt` drives the standalone
-# `src/ascend/custom/build.sh` via `execute_process` at configure time
-# (sidesteps a `CANN` `extract_host_stub.py` path bug that breaks
-# in-tree `ascendc_library()` under `scikit-build-core` temp-dir builds)
-# and links the produced `libno_workspace_kernel.a` into the `ops`
-# module with `--whole-archive`.  Requires `torch_npu` and the
-# `AscendC` toolchain (`ccec`).
-option(BUILD_ASCEND_CUSTOM "Build custom AscendC kernels" ON)
+# Default OFF until CANN's `extract_host_stub.py` path handling is fixed for
+# `scikit-build-core` temp-dir builds (triggers `KeyError` on the preprocessed
+# object path). Enable explicitly with `-DBUILD_CUSTOM_KERNEL=ON` when the
+# toolchain is compatible or when building via the standalone
+# `src/ascend/custom/build.sh` script.
+option(BUILD_CUSTOM_KERNEL "Build custom AscendC kernel PyTorch extension (requires `torch_npu`)" OFF)
 
 option(AUTO_DETECT_DEVICES "Automatically detect available devices" OFF)
 option(AUTO_DETECT_BACKENDS "Automatically detect available backends" OFF)

pyproject.toml

@@ -7,15 +7,6 @@ name = "InfiniOps"
 version = "0.1.0"
 
 [project.optional-dependencies]
-# TODO: `torch` here is unconstrained.  On Ascend hosts, the working
-# torch is the Ascend-matched `torch 2.9.0+cpu` paired with
-# `torch_npu 2.9.0.post1+…`.  A `pip install -e .[dev] --force-reinstall`
-# will re-resolve `torch` to the latest PyPI version (currently
-# `torch 2.11.0`), which now declares `cuda-toolkit` / `nvidia-cublas` /
-# `nvidia-cudnn` / … as hard deps — downloads GBs of CUDA wheels and
-# kills the `torch_npu` / `vllm-ascend` pairing.  Needs a platform-aware
-# split (e.g. `torch; platform_machine != 'aarch64'`, or move `torch`
-# out of `dev` and require it pre-installed in the container image).
 dev = ["pytest", "pytest-cov", "pytest-xdist", "ruff", "torch", "pyyaml"]
 
 [tool.scikit-build.wheel]

scripts/generate_wrappers.py

@@ -112,29 +112,9 @@ def _find_vector_tensor_params(op_name):
     return set(re.findall(r"std::vector<Tensor>\s+(\w+)", source))
 
 
-def _find_params_with_defaults(op_name):
-    """Return ``{param_name: default_literal}`` for base-header params that
-    carry a `= <literal>` default value.  `libclang`'s cursor API does not
-    expose defaults reliably, so we regex-scan the source.  Only used for
-    plain scalar defaults such as ``bool pre_gathered = false``.
-    """
-    source = (_BASE_DIR / f"{op_name}.h").read_text()
-
-    mapping = {}
-
-    for name, default in re.findall(
-        r"\b(?:bool|int(?:64_t|32_t|8_t|16_t)?|std::size_t|std::uint\w+_t|float|double)\s+(\w+)\s*=\s*([^,\)]+?)\s*(?:,|\))",
-        source,
-    ):
-        mapping[name] = default.strip()
-
-    return mapping
-
-
 def _generate_pybind11(operator):
     optional_tensor_params = _find_optional_tensor_params(operator.name)
     vector_tensor_params = _find_vector_tensor_params(operator.name)
-    params_with_defaults = _find_params_with_defaults(operator.name)
 
     def _is_optional_tensor(arg):
         if arg.spelling in optional_tensor_params:
@@ -206,10 +186,6 @@ def _generate_py_args(node):
 
             if _is_optional(arg):
                 parts.append(f'py::arg("{arg.spelling}") = py::none()')
-            elif arg.spelling in params_with_defaults:
-                parts.append(
-                    f'py::arg("{arg.spelling}") = {params_with_defaults[arg.spelling]}'
-                )
             else:
                 parts.append(f'py::arg("{arg.spelling}")')
 
@@ -281,7 +257,8 @@ def _generate_call(op_name, call, method=True):
       }})
       .def_static("clear_cache", &Self::clear_cache);
 
-{callers}}}
+{callers}
+}}
 
 }}  // namespace infini::ops
 

src/CMakeLists.txt

@@ -241,66 +241,8 @@ if(WITH_ASCEND)
     list(APPEND DEVICE_LIST "ascend")
 
     # Custom `AscendC` kernels (PyTorch extension, requires `torch_npu`).
-    if(BUILD_ASCEND_CUSTOM)
-        # In-tree `ascendc_library()` trips the `CANN` `extract_host_stub.py`
-        # path-handling bug under `scikit-build-core`'s temp-dir builds
-        # (`KeyError` on `/./workspace/...` paths in `$<TARGET_OBJECTS>`).
-        # Work around it by driving the standalone `src/ascend/custom/build.sh`
-        # — that script invokes a separate `cmake` with
-        # `src/ascend/custom/` as its `SOURCE_DIR`, avoiding the buggy
-        # path shape.  The produced `.a` is imported and linked into
-        # `ops` with `--whole-archive`.
-        set(_custom_build_dir "${CMAKE_SOURCE_DIR}/build/build_ascend_custom")
-        set(_custom_lib "${_custom_build_dir}/lib/libno_workspace_kernel.a")
-
-        if(NOT DEFINED SOC_VERSION OR "${SOC_VERSION}" STREQUAL "")
-            include(${CMAKE_CURRENT_SOURCE_DIR}/ascend/custom/cmake/detect_soc.cmake)
-            infiniops_detect_soc(SOC_VERSION)
-        endif()
-
-        # Drive `build.sh` as a build-phase target with explicit source
-        # dependencies so that editing any `op_host/` or `op_kernel/`
-        # source re-triggers the build (plain `execute_process` at
-        # configure time would only gate on file existence and leave
-        # stale `.a` files in place).
-        file(GLOB_RECURSE _custom_srcs CONFIGURE_DEPENDS
-            "${CMAKE_CURRENT_SOURCE_DIR}/ascend/custom/*.cpp"
-            "${CMAKE_CURRENT_SOURCE_DIR}/ascend/custom/*.h"
-            "${CMAKE_CURRENT_SOURCE_DIR}/ascend/custom/build.sh")
-
-        # Scrub env inherited from the outer `scikit-build-core` invocation
-        # before handing control to `build.sh`:
-        #  * `CMAKE_GENERATOR` / `CMAKE_EXPORT_COMPILE_COMMANDS` leaking
-        #    into the inner `cmake` change the path format passed to
-        #    `ninja`'s `_host_cpp` rule and re-trigger the `CANN`
-        #    `extract_host_stub.py` `KeyError` (`/./workspace/...`) that
-        #    standalone `build.sh` avoids.
-        #  * `PYTHONPATH` from `pip`'s build-isolation overlay makes the
-        #    child `python3` skip the system `site-packages` — child
-        #    `cmake` modules that `import torch` (`config_envs.cmake`)
-        #    then fail with `ModuleNotFoundError` even though `torch` is
-        #    installed.
-        add_custom_command(
-            OUTPUT ${_custom_lib}
-            COMMAND ${CMAKE_COMMAND} -E env
-                    --unset=CMAKE_GENERATOR
-                    --unset=CMAKE_EXPORT_COMPILE_COMMANDS
-                    --unset=CMAKE_BUILD_PARALLEL_LEVEL
-                    --unset=PYTHONPATH
-                    "BUILD_DIR=${_custom_build_dir}"
-                    "CMAKE_EXE=${CMAKE_COMMAND}"
-                    bash ${CMAKE_CURRENT_SOURCE_DIR}/ascend/custom/build.sh ${SOC_VERSION}
-            WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/ascend/custom
-            DEPENDS ${_custom_srcs}
-            COMMENT "Building custom AscendC kernels (SOC_VERSION=${SOC_VERSION})"
-            VERBATIM)
-
-        add_custom_target(no_workspace_kernel_build ALL DEPENDS ${_custom_lib})
-
-        add_library(no_workspace_kernel STATIC IMPORTED GLOBAL)
-        set_target_properties(no_workspace_kernel PROPERTIES
-            IMPORTED_LOCATION "${_custom_lib}")
-        add_dependencies(no_workspace_kernel no_workspace_kernel_build)
+    if(BUILD_CUSTOM_KERNEL)
+        add_subdirectory(ascend/custom)
 
         # Link the compiled `AscendC` kernel objects into `infiniops` so that
         # custom kernel implementations (e.g. `RmsNorm` index 1) can call
@@ -437,13 +379,9 @@ if(GENERATE_PYTHON_BINDINGS)
     # The `Operator<..., 1>` template instantiations that call
     # `aclrtlaunch_*` live in `ops.cc`, so link here with
     # `--whole-archive` to ensure all launch functions are available.
-    # `$<TARGET_FILE>` works for both real `ascendc_library()` targets and
-    # `IMPORTED` targets pointing at a pre-built `.a`.
-    if(BUILD_ASCEND_CUSTOM)
+    if(BUILD_CUSTOM_KERNEL)
         target_link_libraries(ops PRIVATE
-            -Wl,--whole-archive $<TARGET_FILE:no_workspace_kernel> -Wl,--no-whole-archive)
-        # `ops` link step must wait for `build.sh` to produce the `.a`.
-        add_dependencies(ops no_workspace_kernel_build)
+            -Wl,--whole-archive no_workspace_kernel -Wl,--no-whole-archive)
     endif()
 
     set_target_properties(infiniops PROPERTIES INSTALL_RPATH "$ORIGIN")