From d95dd67e2f89ebfb4102831b515452ca023a96e9 Mon Sep 17 00:00:00 2001
From: Digant Desai <digantdesai@meta.com>
Date: Tue, 3 Nov 2026 06:24:23 -0800
Subject: [PATCH] [executorch][native] Pin eager kernels when generating op
 samples

Pull Request resolved: https://github.com/pytorch/executorch/pull/24358

Eager golden outputs differed by an ulp or two between generating hosts, because MKL and oneDNN pick kernels by ISA and MKL splits work by thread count. The `op_samples` genrule and the `native_op_samples` CMake target now pin MKL to its reproducible path, oneDNN and ATen to AVX2, and OpenMP and MKL to one thread. `MKL_NUM_THREADS` is set as well as `OMP_NUM_THREADS` so a host's own value cannot override it.

Authored with Claude Code.

@bypass-github-export-checks
@bypass-github-pytorch-ci-checks
@bypass-github-executorch-ci-checks
ghstack-source-id: 463657753
@exported-using-ghexport

Differential Revision: [D128254894](https://our.internmc.facebook.com/intern/diff/D128254894/)
---
 backends/native/test/CMakeLists.txt      | 11 +++++++----
 backends/native/test/samples/BUCK        |  6 ++++--
 backends/native/test/samples/generate.py |  7 ++++++-
 3 files changed, 17 insertions(+), 7 deletions(-)

diff --git a/backends/native/test/CMakeLists.txt b/backends/native/test/CMakeLists.txt
index eb404c4cba9..4e9c7b28837 100644
--- a/backends/native/test/CMakeLists.txt
+++ b/backends/native/test/CMakeLists.txt
@@ -29,7 +29,9 @@ target_compile_options(
                               $<$<NOT:$<CXX_COMPILER_ID:MSVC>>:-fexceptions>
 )
 
-# Op packages with bundled samples, one per op and quantization scheme.
+# Op packages with bundled samples, one per op and quantization scheme. The
+# environment pins the eager kernels' ISA and threads, so the golden is the same
+# on any x86 host with AVX2. Other hosts can differ.
 set(_native_op_samples ${CMAKE_CURRENT_BINARY_DIR}/op_samples)
 file(GLOB _native_op_samples_srcs CONFIGURE_DEPENDS
      ${CMAKE_CURRENT_SOURCE_DIR}/samples/*.py
@@ -37,9 +39,10 @@ file(GLOB _native_op_samples_srcs CONFIGURE_DEPENDS
 add_custom_command(
   OUTPUT ${_native_op_samples}.stamp
   COMMAND
-    ${CMAKE_COMMAND} -E env "PYTHONPATH=${EXECUTORCH_ROOT}/src"
-    ${PYTHON_EXECUTABLE} -m executorch.backends.native.test.samples.generate
-    --out ${_native_op_samples}
+    ${CMAKE_COMMAND} -E env MKL_CBWR=COMPATIBLE ONEDNN_MAX_CPU_ISA=AVX2
+    ATEN_CPU_CAPABILITY=avx2 OMP_NUM_THREADS=1 MKL_NUM_THREADS=1
+    "PYTHONPATH=${EXECUTORCH_ROOT}/src" ${PYTHON_EXECUTABLE} -m
+    executorch.backends.native.test.samples.generate --out ${_native_op_samples}
   COMMAND ${CMAKE_COMMAND} -E touch ${_native_op_samples}.stamp
   DEPENDS ${_native_op_samples_srcs}
           ${EXECUTORCH_ROOT}/backends/test/suite/operators/test_linear.py
diff --git a/backends/native/test/samples/BUCK b/backends/native/test/samples/BUCK
index 97b4b66b7db..48e83efc2d9 100644
--- a/backends/native/test/samples/BUCK
+++ b/backends/native/test/samples/BUCK
@@ -40,11 +40,13 @@ fbcode_target(
     ],
 )
 
-# Op packages with bundled samples, one per op and quantization scheme.
+# Op packages with bundled samples, one per op and quantization scheme. The
+# environment pins the eager kernels' ISA and threads, so the golden is the same
+# on any x86 host with AVX2. Other hosts can differ.
 fbcode_target(
     _kind = runtime.genrule,
     name = "op_samples",
     out = "op_samples",
-    cmd = "$(exe :generate) --out $OUT",
+    cmd = "MKL_CBWR=COMPATIBLE ONEDNN_MAX_CPU_ISA=AVX2 ATEN_CPU_CAPABILITY=avx2 OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 $(exe :generate) --out $OUT",
     visibility = ["//executorch/backends/native/test/..."],
 )
diff --git a/backends/native/test/samples/generate.py b/backends/native/test/samples/generate.py
index 665a17aa22b..f98fe7e91cb 100644
--- a/backends/native/test/samples/generate.py
+++ b/backends/native/test/samples/generate.py
@@ -5,7 +5,12 @@
 # LICENSE file in the root directory of this source tree.
 
 """Write <op>_<scheme>.ptn for each op and quantization scheme that fits a case
-of the op."""
+of the op.
+
+Golden outputs depend on the eager kernels' ISA and threads; the op_samples
+genrule and the native_op_samples CMake target pin them, so build one of those
+rather than running this module to reproduce their packages.
+"""
 
 import argparse
 import logging