From 324a3d2bd8b5fc3f38c6d2f2cc243f747800ba28 Mon Sep 17 00:00:00 2001
From: "Li, Jiang" <jiang1.li@intel.com>
Date: Thu, 16 Apr 2026 21:50:36 +0800
Subject: [PATCH] [CI/Build] Improve stability of CPU tests (#39966)

Signed-off-by: jiang1.li <jiang1.li@intel.com>
---
 .buildkite/hardware_tests/cpu.yaml               | 4 ++--
 tests/models/language/generation/test_common.py  | 8 ++++----
 tests/models/language/generation/test_granite.py | 1 +
 vllm/platforms/cpu.py                            | 1 +
 4 files changed, 8 insertions(+), 6 deletions(-)

diff --git a/.buildkite/hardware_tests/cpu.yaml b/.buildkite/hardware_tests/cpu.yaml
index e466e2a5244..9b104444378 100644
--- a/.buildkite/hardware_tests/cpu.yaml
+++ b/.buildkite/hardware_tests/cpu.yaml
@@ -46,7 +46,7 @@ steps:
   - tests/models/language/pooling/
   commands:
     - |
-      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 30m "
+      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 40m "
       pytest -x -v -s tests/models/language/generation -m cpu_model
       pytest -x -v -s tests/models/language/pooling -m cpu_model"
 
@@ -99,7 +99,7 @@ steps:
     - |
       bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 45m "
       pytest -x -v -s tests/models/multimodal/generation --ignore=tests/models/multimodal/generation/test_pixtral.py -m cpu_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB"
-  parallelism: 2
+  parallelism: 3
 
 - label: "Arm CPU Test"
   depends_on: []
diff --git a/tests/models/language/generation/test_common.py b/tests/models/language/generation/test_common.py
index b276f37a2a3..ebf847b706c 100644
--- a/tests/models/language/generation/test_common.py
+++ b/tests/models/language/generation/test_common.py
@@ -100,7 +100,7 @@ AITER_MODEL_LIST = [
         pytest.param("bigcode/starcoder2-3b"),  # starcoder2
         pytest.param(
             "TitanML/tiny-mixtral",  # mixtral
-            marks=[pytest.mark.core_model, pytest.mark.cpu_model],
+            marks=[pytest.mark.core_model],
         ),
         pytest.param("swiss-ai/Apertus-8B-Instruct-2509"),  # apertus
         pytest.param(
@@ -143,9 +143,9 @@ def test_models(
         # in parts of the operators
         pytest.skip(f"Skipping '{model}' model test with AITER kernel.")
 
-    if current_platform.is_cpu() and model == "TitanML/tiny-mixtral":
-        # This untrained model is sensitive to the rounding error
-        # Fuse ops to reduce bfloat16 rounding
+    if current_platform.is_cpu() and model in ("openai-community/gpt2",):
+        # These models are sensitive to the rounding error
+        # Fuse ops to reduce rounding
         monkeypatch.setenv("VLLM_CPU_CI_ENV", "0")
 
     with hf_runner(model) as hf_model:
diff --git a/tests/models/language/generation/test_granite.py b/tests/models/language/generation/test_granite.py
index e569e75ff3a..c0498b2f7de 100644
--- a/tests/models/language/generation/test_granite.py
+++ b/tests/models/language/generation/test_granite.py
@@ -15,6 +15,7 @@ MODELS = [
 @pytest.mark.parametrize("dtype", ["bfloat16"])
 @pytest.mark.parametrize("max_tokens", [64])
 @pytest.mark.parametrize("num_logprobs", [5])
+@pytest.mark.cpu_model
 def test_models(
     hf_runner,
     vllm_runner,
diff --git a/vllm/platforms/cpu.py b/vllm/platforms/cpu.py
index f319dbc497a..86f07b1032e 100644
--- a/vllm/platforms/cpu.py
+++ b/vllm/platforms/cpu.py
@@ -242,6 +242,7 @@ class CpuPlatform(Platform):
                     "cpp.dynamic_threads": True,
                 }
             )
+            compilation_config.ir_enable_torch_wrap = False
 
         if vllm_config.lora_config is not None:
             compilation_config.mode = CompilationMode.NONE