[cpu][perf] Fix low CPU utilization with VLLM_CPU_OMP_THREADS_BIND on AArch64 (#27415)

Signed-off-by: Fadi Arafeh <fadi.arafeh@arm.com>

cmake/cpu_extension.cmake

@@ -212,11 +212,24 @@ if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR (ASIMD_FOUND AND NOT APPLE_SILICON
         # Build ACL with scons
         include(ProcessorCount)
         ProcessorCount(_NPROC)
+        set(_scons_cmd
+        scons -j${_NPROC}
+            Werror=0 debug=0 neon=1 examples=0 embed_kernels=0 os=linux
+            arch=armv8.2-a build=native benchmark_examples=0 fixed_format_kernels=1
+            multi_isa=1 openmp=1 cppthreads=0
+        )
+
+        # locate PyTorch's libgomp (e.g. site-packages/torch.libs/libgomp-947d5fa1.so.1.0.0)
+        # and create a local shim dir with it
+        include("${CMAKE_CURRENT_LIST_DIR}/utils.cmake")
+        vllm_prepare_torch_gomp_shim(VLLM_TORCH_GOMP_SHIM_DIR)
+
+        if(NOT VLLM_TORCH_GOMP_SHIM_DIR STREQUAL "")
+            list(APPEND _scons_cmd extra_link_flags=-L${VLLM_TORCH_GOMP_SHIM_DIR})
+        endif()
+
         execute_process(
-            COMMAND scons -j${_NPROC}
-                    Werror=0 debug=0 neon=1 examples=0 embed_kernels=0 os=linux
-                    arch=armv8.2-a build=native benchmark_examples=0 fixed_format_kernels=1
-                    multi_isa=1 openmp=1 cppthreads=0
+            COMMAND ${_scons_cmd}
             WORKING_DIRECTORY "$ENV{ACL_ROOT_DIR}"
             RESULT_VARIABLE _acl_rc
         )

cmake/utils.cmake

@@ -129,6 +129,44 @@ function (get_torch_gpu_compiler_flags OUT_GPU_FLAGS GPU_LANG)
   set(${OUT_GPU_FLAGS} ${GPU_FLAGS} PARENT_SCOPE)
 endfunction()
 
+# Find libgomp that gets shipped with PyTorch wheel and create a shim dir with:
+#   libgomp.so    -> libgomp-<hash>.so...
+#   libgomp.so.1  -> libgomp-<hash>.so...
+# OUTPUT: TORCH_GOMP_SHIM_DIR  ("" if not found)
+function(vllm_prepare_torch_gomp_shim TORCH_GOMP_SHIM_DIR)
+  set(${TORCH_GOMP_SHIM_DIR} "" PARENT_SCOPE)
+
+  # Use run_python to locate vendored libgomp; never throw on failure.
+  run_python(_VLLM_TORCH_GOMP_PATH
+    "
+import os, glob
+try:
+  import torch
+  torch_pkg = os.path.dirname(torch.__file__)
+  site_root = os.path.dirname(torch_pkg)
+  torch_libs = os.path.join(site_root, 'torch.libs')
+  print(glob.glob(os.path.join(torch_libs, 'libgomp-*.so*'))[0])
+except:
+  print('')
+"
+    "failed to probe torch.libs for libgomp")
+
+  if(_VLLM_TORCH_GOMP_PATH STREQUAL "" OR NOT EXISTS "${_VLLM_TORCH_GOMP_PATH}")
+    return()
+  endif()
+
+  # Create shim under the build tree
+  set(_shim "${CMAKE_BINARY_DIR}/gomp_shim")
+  file(MAKE_DIRECTORY "${_shim}")
+
+  execute_process(COMMAND ${CMAKE_COMMAND} -E rm -f "${_shim}/libgomp.so")
+  execute_process(COMMAND ${CMAKE_COMMAND} -E rm -f "${_shim}/libgomp.so.1")
+  execute_process(COMMAND ${CMAKE_COMMAND} -E create_symlink "${_VLLM_TORCH_GOMP_PATH}" "${_shim}/libgomp.so")
+  execute_process(COMMAND ${CMAKE_COMMAND} -E create_symlink "${_VLLM_TORCH_GOMP_PATH}" "${_shim}/libgomp.so.1")
+
+  set(${TORCH_GOMP_SHIM_DIR} "${_shim}" PARENT_SCOPE)
+endfunction()
+
 # Macro for converting a `gencode` version number to a cmake version number.
 macro(string_to_ver OUT_VER IN_STR)
   string(REGEX REPLACE "\([0-9]+\)\([0-9]\)" "\\1.\\2" ${OUT_VER} ${IN_STR})

vllm/platforms/cpu.py

@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
+import glob
 import json
 import os
 import platform
@@ -301,8 +302,8 @@ class CpuPlatform(Platform):
         os.environ["VLLM_DISABLE_SHARED_EXPERTS_STREAM"] = "0"
 
         # Intel OpenMP setting
-        ld_prealod_str = os.getenv("LD_PRELOAD", "")
-        if "libiomp5.so" in ld_prealod_str:
+        ld_preload_str = os.getenv("LD_PRELOAD", "")
+        if "libiomp5.so" in ld_preload_str:
             # The time(milliseconds) that a thread should wait after
             # completing the execution of a parallel region, before sleeping.
             os.environ["KMP_BLOCKTIME"] = "1"
@@ -313,6 +314,31 @@ class CpuPlatform(Platform):
             os.environ["KMP_PLAIN_BARRIER_PATTERN"] = "dist,dist"
             os.environ["KMP_REDUCTION_BARRIER_PATTERN"] = "dist,dist"
 
+        if (
+            platform.system() == "Linux"
+            and Platform.get_cpu_architecture() == CpuArchEnum.ARM
+            and not ("libomp" in ld_preload_str or "libgomp" in ld_preload_str)
+        ):
+            # We need to LD_PRELOAD PyTorch's libgomp, otherwise only
+            # one core will be properly utilized when we thread-bind
+            # See: https://github.com/vllm-project/vllm/issues/27369
+            # TODO: Remove once:
+            # https://github.com/pytorch/pytorch/issues/166087 is fixed
+
+            # We need to find the location of PyTorch's libgomp
+            torch_pkg = os.path.dirname(torch.__file__)
+            site_root = os.path.dirname(torch_pkg)
+            torch_libs = os.path.join(site_root, "torch.libs")
+            pytorch_libgomp_so_candidates = glob.glob(
+                os.path.join(torch_libs, "libgomp-*.so*")
+            )
+            if pytorch_libgomp_so_candidates:
+                pytorch_libgomp_so = pytorch_libgomp_so_candidates[0]
+                if ld_preload_str:
+                    ld_preload_str += ":"
+                ld_preload_str += pytorch_libgomp_so
+                os.environ["LD_PRELOAD"] = ld_preload_str
+
         # To hint IPEX uses shared memory based AllReduce
         os.environ["LOCAL_WORLD_SIZE"] = str(
             vllm_config.parallel_config.tensor_parallel_size