From 8d09cdbef4f645001cc6e6bfc45eb1ca86d8ba27 Mon Sep 17 00:00:00 2001
From: Digant Desai <digantdesai@meta.com>
Date: Fri, 30 Oct 2026 16:28:42 -0700
Subject: [PATCH] Update

[ghstack-poisoned]
---
 .ci/scripts/test-native.sh             |  5 +++--
 .github/workflows/test-native.yml      |  9 ++++++---
 backends/native/test/CMakeLists.txt    | 11 +++++++----
 backends/native/test/runtime/ggml/BUCK | 10 ++++++----
 4 files changed, 22 insertions(+), 13 deletions(-)

diff --git a/.ci/scripts/test-native.sh b/.ci/scripts/test-native.sh
index 98139a453e2..fadbfe75840 100644
--- a/.ci/scripts/test-native.sh
+++ b/.ci/scripts/test-native.sh
@@ -36,8 +36,9 @@ case "${component}" in
       -DEXECUTORCH_BUILD_NATIVE_VISION_RUNNER=ON
       -DGGML_NATIVE=OFF
     )
-    targets+=(run_hf_llm_ggml native_ggml_engine_test native_llm_runner_utils_test run_classification_ggml_cpu)
-    test_regex='^native_(ggml_engine|llm_runner_utils)_test$'
+    # TODO: run the other sample tests once they pass.
+    targets+=(run_hf_llm_ggml native_ggml_engine_test native_llm_runner_utils_test run_classification_ggml_cpu native_output_check_test native_ggml_sample_test)
+    test_regex='^native_(ggml_engine|llm_runner_utils|output_check)_test$|^native_ggml_sample_test_act_quant$'
     ;;
   vulkan)
     cmake_args+=(
diff --git a/.github/workflows/test-native.yml b/.github/workflows/test-native.yml
index c7a7aae2f0c..71d051f1679 100644
--- a/.github/workflows/test-native.yml
+++ b/.github/workflows/test-native.yml
@@ -17,6 +17,8 @@ on:
       - backends/cpu/CMakeLists.txt
       - backends/cpu/runtime/**
       - backends/native/**
+      - backends/cadence/utils/FACTO
+      - backends/test/suite/operators/test_linear.py
       - backends/vulkan/CMakeLists.txt
       - backends/vulkan/cmake/**
       - backends/vulkan/runtime/api/**
@@ -113,9 +115,10 @@ jobs:
 
           if [[ "${NATIVE_COMPONENT}" == "ggml" ]]; then
             ./install_requirements.sh --use-pt-pinned-commit
-            # The LLM runner test imports executorch from src/ without installing the
-            # wheel, so install its export-time dependencies directly.
-            pip install flatbuffers ruamel.yaml tabulate \
+            # The LLM runner test and the op sample generator import executorch from
+            # src/ without installing the wheel, so install their export-time
+            # dependencies directly.
+            pip install flatbuffers gguf ruamel.yaml tabulate \
               "$(grep -E '^transformers\b' requirements-examples.txt)"
           fi
 
diff --git a/backends/native/test/CMakeLists.txt b/backends/native/test/CMakeLists.txt
index 3fdc35066d2..6efaef06591 100644
--- a/backends/native/test/CMakeLists.txt
+++ b/backends/native/test/CMakeLists.txt
@@ -154,10 +154,13 @@ if(EXECUTORCH_BUILD_NATIVE_GGML)
                    --device=cpu
   )
   # native_ggml_sample_test_act_quant lets ggml quantize the activations of
-  # weight-only quantized linears, judged against golden-int8_act.
-  add_test(NAME native_ggml_sample_test_act_quant
-           COMMAND native_ggml_sample_test --packages=${_native_op_samples}
-                   --device=cpu --allow_activation_quant_for_weight_only
+  # weight-only quantized linears, judged against golden-int8_act. ggml
+  # quantizes them symmetrically per block, not asymmetrically per token like
+  # the golden, so the biases get one more standard error.
+  add_test(
+    NAME native_ggml_sample_test_act_quant
+    COMMAND native_ggml_sample_test --packages=${_native_op_samples}
+            --device=cpu --allow_activation_quant_for_weight_only --sigmas=7
   )
   target_compile_options(
     native_ggml_sample_test
diff --git a/backends/native/test/runtime/ggml/BUCK b/backends/native/test/runtime/ggml/BUCK
index 064ee4a1254..84696757b65 100644
--- a/backends/native/test/runtime/ggml/BUCK
+++ b/backends/native/test/runtime/ggml/BUCK
@@ -94,7 +94,9 @@ _VARIANTS = [
 ]
 
 # ggml_sample_test_act_quant lets ggml quantize the activations of weight-only
-# quantized linears, judged against golden-int8_act.
+# quantized linears, judged against golden-int8_act. ggml quantizes them
+# symmetrically per block, not asymmetrically per token like the golden, so the
+# biases get one more standard error.
 [
     fbcode_target(
         _kind = runtime.cxx_test,
@@ -105,8 +107,8 @@ _VARIANTS = [
         ] + flags,
         env = v.get("env", {}) | v.get("sample_env", {}),
         remote_execution = v.get("remote_execution"),
-        # TODO: run in CI once they pass.
-        labels = ci.labels(ci.map(ci.skip_test())),
+        # TODO: run the others in CI once they pass.
+        labels = [] if v["suffix"] + name == "_act_quant" else ci.labels(ci.map(ci.skip_test())),
         test_main = "//executorch/backends/native/test/samples:sample_test_main",
         deps = [":ggml_engine_host" + v["suffix"]],
         **v.get("compat", {})
@@ -114,7 +116,7 @@ _VARIANTS = [
     for v in _VARIANTS
     for name, flags in [
         ("", []),
-        ("_act_quant", ["--allow_activation_quant_for_weight_only"]),
+        ("_act_quant", ["--allow_activation_quant_for_weight_only", "--sigmas=7"]),
     ]
 ]