A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 5, 02:15 UTC
[CI/Build] Improve stability of CPU tests (#39966)
Signed-off-by: jiang1.li <jiang1.li@intel.com>
Li, Jiang committed
Apr 16, 2026 at 21:50 UTC
324a3d2bd8b5fc3f38c6d2f2cc243f747800ba28
4 files changed
+8
-6
.buildkite/hardware_tests/cpu.yaml
+2
-2
@@ -46,7 +46,7 @@ steps:
46
- tests/models/language/pooling/
47
commands:
48
- |
49
- bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 30m "
49
+ bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 40m "
50
pytest -x -v -s tests/models/language/generation -m cpu_model
51
pytest -x -v -s tests/models/language/pooling -m cpu_model"
52
@@ -99,7 +99,7 @@ steps:
99
- |
100
bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 45m "
101
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"
102
- parallelism: 2
102
+ parallelism: 3
103
104
- label: "Arm CPU Test"
105
depends_on: []
tests/models/language/generation/test_common.py
+4
-4
@@ -100,7 +100,7 @@ AITER_MODEL_LIST = [
100
pytest.param("bigcode/starcoder2-3b"), # starcoder2
101
pytest.param(
102
"TitanML/tiny-mixtral", # mixtral
103
- marks=[pytest.mark.core_model, pytest.mark.cpu_model],
103
+ marks=[pytest.mark.core_model],
104
),
105
pytest.param("swiss-ai/Apertus-8B-Instruct-2509"), # apertus
106
pytest.param(
@@ -143,9 +143,9 @@ def test_models(
143
# in parts of the operators
144
pytest.skip(f"Skipping '{model}' model test with AITER kernel.")
145
146
- if current_platform.is_cpu() and model == "TitanML/tiny-mixtral":
147
- # This untrained model is sensitive to the rounding error
148
- # Fuse ops to reduce bfloat16 rounding
146
+ if current_platform.is_cpu() and model in ("openai-community/gpt2",):
147
+ # These models are sensitive to the rounding error
148
+ # Fuse ops to reduce rounding
149
monkeypatch.setenv("VLLM_CPU_CI_ENV", "0")
150
151
with hf_runner(model) as hf_model:
tests/models/language/generation/test_granite.py
+1
@@ -15,6 +15,7 @@ MODELS = [
15
@pytest.mark.parametrize("dtype", ["bfloat16"])
16
@pytest.mark.parametrize("max_tokens", [64])
17
@pytest.mark.parametrize("num_logprobs", [5])
18
+@pytest.mark.cpu_model
19
def test_models(
20
hf_runner,
21
vllm_runner,
vllm/platforms/cpu.py
+1
@@ -242,6 +242,7 @@ class CpuPlatform(Platform):
242
"cpp.dynamic_threads": True,
243
}
244
)
245
+ compilation_config.ir_enable_torch_wrap = False
246
247
if vllm_config.lora_config is not None:
248
compilation_config.mode = CompilationMode.NONE