Remove usage of torch.autograd.profiler_legacy for benchmarks; remove IPEX (#2149)

Signed-off-by: Anatoly Myachev <[email protected]>

Author: anmyachev

.github/actions/setup-ipex/action.yml

@@ -16,7 +16,6 @@ on:
         description: The method used to obtain performance numbers
         type: choice
         options:
-          - PYTORCH_LEGACY_PROFILER_USING_IPEX
           - ELAPSED_TIME
           - UPSTREAM_PYTORCH_PROFILER
         default: UPSTREAM_PYTORCH_PROFILER
@@ -54,7 +53,6 @@ permissions: read-all
 env:
   PYTHON_VERSION: "3.10"
   BENCHMARKING_METHOD: ${{ inputs.benchmarking_method || 'UPSTREAM_PYTORCH_PROFILER' }}
-  USE_IPEX: ${{ (inputs.benchmarking_method || 'UPSTREAM_PYTORCH_PROFILER') == 'PYTORCH_LEGACY_PROFILER_USING_IPEX' && '1' || '0' }}
   TAG: ${{ inputs.tag || (github.event_name == 'pull_request' && format('pr-{0}', github.event.number)) || (github.event_name == 'schedule' && 'ci') || 'test' }}
 
 jobs:
@@ -113,22 +111,11 @@ jobs:
         run: |
           pip install wheel cmake
 
-      - name: Setup PyTorch with IPEX
-        if: ${{ env.USE_IPEX == '1' }}
-        uses: ./.github/actions/setup-pytorch
-        with:
-          repository: Stonepia/pytorch
-
-      - name: Setup PyTorch without IPEX
-        if: ${{ env.USE_IPEX == '0' }}
+      - name: Setup PyTorch
         uses: ./.github/actions/setup-pytorch
         with:
           repository: pytorch/pytorch
 
-      - name: Setup IPEX
-        if: ${{ env.USE_IPEX == '1' }}
-        uses: ./.github/actions/setup-ipex
-
       - name: Build Triton wheels
         uses: ./.github/actions/setup-triton
         with:

.github/actions/setup-no-op-ipex/action.yml

@@ -67,8 +67,6 @@ repos:
       - id: pylint
         name: pylint for scripts
         files: '^scripts/.*\.py$'
-        # FIXME: will be deleted soon
-        exclude: '^scripts/create-noop-ipex\.py$'
         args:
           - --check-quote-consistency=true
           # Max line length 120 is consistent with ruff's one in upstream

.github/workflows/triton-benchmarks.yml

@@ -4,8 +4,6 @@ set(CMAKE_CXX_STANDARD 20)
 
 project(TritonBenchmark)
 
-option(USE_IPEX "Use IPEX" ON)
-
 if(NOT WIN32)
     list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake")
 endif()
@@ -16,11 +14,5 @@ find_package(Torch REQUIRED)
 find_library(TORCH_PYTHON_LIBRARY torch_python PATH "${TORCH_INSTALL_PREFIX}/lib")
 find_package(XeTLALibrary REQUIRED)
 
-if(USE_IPEX)
-  string(APPEND CMAKE_CXX_FLAGS " -DUSE_IPEX")
-  find_package(IPEX REQUIRED)
-endif()
-
-
 # add the XeTLA kernel.
 add_subdirectory(xetla_kernel)

.pre-commit-config.yaml

@@ -1,12 +1,7 @@
-import os
-
 import torch
 import triton
 import triton.language as tl
 
-if os.getenv('USE_IPEX', '1') == '1':
-    import intel_extension_for_pytorch  # type: ignore # noqa: F401
-
 
 @triton.jit
 def float_trunc_kernel(

benchmarks/CMakeLists.txt

@@ -28,29 +28,15 @@ def __init__(self, build_lib, build_temp, debug=False, dry_run=False):
         self.extdir = build_lib + "/triton_kernels_benchmark"
         self.build_type = self.get_build_type(debug)
         self.cmake_prefix_paths = [torch.utils.cmake_prefix_path]
-        self.use_ipex = False
         self.dry_run = dry_run
 
     def get_build_type(self, debug):
         DEBUG_OPTION = os.getenv("DEBUG", "0")
         return "Debug" if debug or (DEBUG_OPTION == "1") else "Release"
 
     def run(self):
-        self.check_ipex()
         self.build_extension()
 
-    def check_ipex(self):
-        self.use_ipex = os.getenv("USE_IPEX", "1") == "1"
-        if not self.use_ipex:
-            return
-        try:
-            import intel_extension_for_pytorch
-        except ImportError:
-            log.warn("ipex is not installed trying to build without ipex")
-            self.use_ipex = False
-            return
-        self.cmake_prefix_paths.append(intel_extension_for_pytorch.cmake_prefix_path)
-
     def check_call(self, *popenargs, **kwargs):
         log.info(" ".join(popenargs[0]))
         if not self.dry_run:
@@ -67,7 +53,6 @@ def build_extension(self):
             "-DCMAKE_MAKE_PROGRAM=" +
             ninja_dir,  # Pass explicit path to ninja otherwise cmake may cache a temporary path
             "-DCMAKE_PREFIX_PATH=" + ";".join(self.cmake_prefix_paths),
-            f"-DUSE_IPEX={int(self.use_ipex)}",
             "-DCMAKE_INSTALL_PREFIX=" + self.extdir,
             "-DPython3_ROOT_DIR:FILEPATH=" + sys.exec_prefix,
             "-DCMAKE_C_COMPILER=icx",
@@ -141,7 +126,6 @@ def get_git_commit_hash(length=8):
     },
     ext_modules=[CMakeExtension("triton_kernels_benchmark.xetla_kernel")],
     extras_require={
-        "ipex": ["numpy<=2.0", "intel-extension-for-pytorch==2.1.10"],
         "pytorch": ["torch>=2.6"],
     },
 )

benchmarks/micro_benchmarks/conversion/float_conversion/float_conversion.py

@@ -1,6 +1,6 @@
-from .benchmark_testing import do_bench, assert_close, perf_report, Benchmark, USE_IPEX_OPTION, BENCHMARKING_METHOD  # type: ignore # noqa: F401
+from .benchmark_testing import do_bench, assert_close, perf_report, Benchmark, BENCHMARKING_METHOD  # type: ignore # noqa: F401
 
-if USE_IPEX_OPTION or BENCHMARKING_METHOD == "UPSTREAM_PYTORCH_PROFILER":
+if BENCHMARKING_METHOD == "UPSTREAM_PYTORCH_PROFILER":
     from triton.runtime import driver
     from . import benchmark_driver
     # replace the launcher with the profilier hook.

benchmarks/setup.py

@@ -11,8 +11,6 @@
 
 import torch
 
-from .benchmark_testing import USE_IPEX_OPTION
-
 _dirname = os.getenv("ZE_PATH", default="/usr/local")
 
 include_dir = [
@@ -31,13 +29,6 @@
 library_dir = [os.path.join(_dirname, "lib"), os.path.join(torch.utils.cmake_prefix_path, "../../lib")]
 libraries = ["ze_loader", "sycl", "torch"]
 
-if USE_IPEX_OPTION:
-    import intel_extension_for_pytorch
-
-    include_dir.append(os.path.join(intel_extension_for_pytorch.cmake_prefix_path, "../../include"))
-    library_dir.append(os.path.join(intel_extension_for_pytorch.cmake_prefix_path, "../../lib"))
-    libraries.append("intel-ext-pt-gpu")
-
 
 def compile_module_from_src(src, name):
     key = hashlib.sha256(src.encode("utf-8")).hexdigest()
@@ -144,14 +135,6 @@ def format_of(ty):
     fmt = "iiiOOOOOO" + args_format
     args_list = ", " + ", ".join(f"&_arg{i}" for i, ty in signature.items()) if len(signature) > 0 else ""
 
-    record_function_header = "#include <ATen/record_function.h>"
-    ipex_header = ""
-    xpu_profiler_record = ""
-    if USE_IPEX_OPTION:
-        record_function_header = "#include <torch/extension.h>"
-        ipex_header = "#include <ipex.h>"
-        xpu_profiler_record = "xpu::profiler_record(kernel_name, event);"
-
     # generate glue code
     src = f"""
     #include <cstddef>
@@ -160,8 +143,7 @@ def format_of(ty):
     #include <iomanip>
     #include <level_zero/ze_api.h>
     #include <sycl/sycl.hpp>
-    {record_function_header}
-    {ipex_header}
+    #include <ATen/record_function.h>
 
     #define NPY_NO_DEPRECATED_API NPY_1_7_API_VERSION
     #include <Python.h>
@@ -302,7 +284,6 @@ def format_of(ty):
       }}
       }};
     auto event = stream.submit(cgf);
-    {xpu_profiler_record}
   }}
 // end sycl
     static PyObject* launch(PyObject* self, PyObject* args) {{