[CI/Build] Improve stability of CPU tests (#39966)

Signed-off-by: jiang1.li <jiang1.li@intel.com>

.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: []

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:

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,

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