style(comments): wrap remaining technical identifiers in Markdown backticks

Scan-and-fix pass for identifiers in comments that still lack Markdown
backticks, matching reviewer comments 9, 11, 13, 14 on PR #64. Applied
only to files authored / modified by this PR (leaves custom/cmake/
config_envs.cmake and similar vllm-ascend-verbatim content untouched to
stay consistent with the upstream it was adapted from).

- `CMakeLists.txt`: `pybind11` (line 7).
- `src/ascend/common.h`: `shape`, `strides`, `storage_shape`, `dtype`
  in the `AclTensorCache` class doc.
- `src/ascend/custom/CMakeLists.txt`: `AscendC` toolchain reference.
- `src/ascend/custom/build.sh`: `AscendC`, `libascend_kernel.so`.
- `src/ascend/custom/cmake/config_ascend.cmake`: `SOC_VERSION`,
  `CANN`, `AscendC`.

Author: zhangyue

CMakeLists.txt

@@ -4,7 +4,7 @@ project(InfiniOps LANGUAGES CXX)
 set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 
-# Internal variable to control pybind11's automatic optimization flags (like `-flto`).
+# Internal variable to control `pybind11`'s automatic optimization flags (like `-flto`).
 set(PYBIND11_ENABLE_EXTRAS ON)
 
 # Options for backends.

src/CMakeLists.txt

@@ -178,7 +178,7 @@ if(WITH_ASCEND)
         "ascend/*.cc"
         "ascend/*.cpp"
     )
-    # Exclude kernel_impl.cpp — AscendC device code, not compiled by the host C++ compiler.
+    # Exclude `kernel_impl.cpp` — `AscendC` device code, not compiled by the host C++ compiler.
     list(FILTER ASCEND_SOURCES EXCLUDE REGEX ".*kernel_impl\\.cpp$")
     # Exclude custom/ — standalone PyTorch extension, built separately.
     list(FILTER ASCEND_SOURCES EXCLUDE REGEX ".*/custom/.*")
@@ -218,7 +218,7 @@ if(WITH_ASCEND)
         "${ASCEND_HAL_LIB}")
 
     # ATB (Ascend Transformer Boost) — provides fused operators like
-    # PagedAttention and ReshapeAndCache that are graph-capture safe.
+    # `PagedAttention` and `ReshapeAndCache` that are graph-capture safe.
     set(ATB_HOME_DIR "$ENV{ATB_HOME_PATH}")
     if(NOT ATB_HOME_DIR)
         # Default search path under CANN nnal directory.
@@ -240,12 +240,12 @@ if(WITH_ASCEND)
 
     list(APPEND DEVICE_LIST "ascend")
 
-    # Custom AscendC kernels (PyTorch extension, requires torch_npu).
+    # Custom `AscendC` kernels (PyTorch extension, requires `torch_npu`).
     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
+        # Link the compiled `AscendC` kernel objects into `infiniops` so that
+        # custom kernel implementations (e.g. `RmsNorm` index 1) can call
         # them via the generated launch functions.
         target_compile_definitions(infiniops PUBLIC INFINI_HAS_CUSTOM_KERNELS=1)
     endif()
@@ -373,8 +373,8 @@ if(GENERATE_PYTHON_BINDINGS)
     target_include_directories(ops PRIVATE ${PROJECT_SOURCE_DIR})
     target_link_libraries(ops PRIVATE infiniops)
 
-    # Custom AscendC kernel objects must be linked directly into ops
-    # because the AscendC toolchain compiles host stubs with hidden
+    # Custom `AscendC` kernel objects must be linked directly into ops
+    # because the `AscendC` toolchain compiles host stubs with hidden
     # visibility — `libinfiniops.so` cannot re-export those symbols.
     # The `Operator<..., 1>` template instantiations that call
     # `aclrtlaunch_*` live in `ops.cc`, so link here with

src/ascend/common.h

@@ -65,10 +65,10 @@ inline aclTensor* buildAclTensor(const Tensor& t,
 
 // Pre-computed tensor metadata for descriptor reuse.
 //
-// Stores shape, strides, storage_shape, and dtype once (avoiding per-call heap
-// allocations).  The `aclTensor` descriptor is created on the first `get()`
-// call and its data pointer is updated in-place via `aclSetRawTensorAddr` on
-// subsequent calls.
+// Stores `shape`, `strides`, `storage_shape`, and `dtype` once (avoiding
+// per-call heap allocations).  The `aclTensor` descriptor is created on the
+// first `get()` call and its data pointer is updated in-place via
+// `aclSetRawTensorAddr` on subsequent calls.
 class AclTensorCache {
  public:
   AclTensorCache() = default;

src/ascend/custom/CMakeLists.txt

@@ -56,7 +56,7 @@ file(GLOB OP_SRCS
 # Python side via `import ascend_kernel`.
 set(OP_PLUGIN_NAME ascend_kernel)
 
-# Kernel side files (device code compiled by AscendC toolchain).
+# Kernel side files (device code compiled by `AscendC` toolchain).
 ascendc_library(no_workspace_kernel STATIC
     ${PROJECT_OP_SRC_BASE}/rms_norm/op_kernel/rms_norm.cpp
 )

src/ascend/custom/build.sh

@@ -1,5 +1,5 @@
 #!/bin/bash
-# Build custom AscendC kernels into libascend_kernel.so.
+# Build custom `AscendC` kernels into `libascend_kernel.so`.
 set -e
 
 SOC_VERSION="${1:-Ascend910_9382}"

src/ascend/custom/cmake/config_ascend.cmake

@@ -6,8 +6,8 @@ endif()
 
 set(ASCEND_CANN_PACKAGE_PATH ${ASCEND_HOME_PATH})
 
-# Auto-detect SOC_VERSION from `npu-smi info` if not set externally.  Required
-# by CANN's `ascendc.cmake` for AscendC kernel compilation.
+# Auto-detect `SOC_VERSION` from `npu-smi info` if not set externally.
+# Required by `CANN`'s `ascendc.cmake` for `AscendC` kernel compilation.
 if(NOT DEFINED SOC_VERSION OR "${SOC_VERSION}" STREQUAL "")
     execute_process(
         COMMAND bash -c "npu-smi info 2>/dev/null | awk '/910B|910A|310/ {for (i=1;i<=NF;i++) if ($i ~ /^(910|310)/) {print \"Ascend\" $i; exit}}'"