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 15, 17:15 UTC
[CI][TPU] Temporarily Disable Quant Test on TPU (#15649)
Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com>
Robert Shaw committed
Mar 27, 2025 at 22:45 UTC
8a49eea74bb1e664381d32f1d041b5d1e651664d
2 files changed
+5
-7
.buildkite/run-tpu-v1-test.sh
+5
-4
@@ -28,15 +28,16 @@ docker run --privileged --net host --shm-size=16G -it \
28
&& echo TEST_3 \
29
&& pytest -v -s /workspace/vllm/tests/entrypoints/llm/test_accuracy.py::test_lm_eval_accuracy_v1_engine \
30
&& echo TEST_4 \
31
- && pytest -s -v /workspace/vllm/tests/tpu/test_quantization_accuracy.py \
32
- && echo TEST_5 \
31
&& python3 /workspace/vllm/examples/offline_inference/tpu.py \
34
- && echo TEST_6 \
32
+ && echo TEST_5 \
33
&& pytest -s -v /workspace/vllm/tests/tpu/worker/test_tpu_model_runner.py \
36
- && echo TEST_7 \
34
+ && echo TEST_6 \
35
&& pytest -s -v /workspace/vllm/tests/v1/tpu/test_sampler.py" \
36
37
38
# TODO: This test fails because it uses RANDOM_SEED sampling
39
# && VLLM_USE_V1=1 pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py \
40
41
+# TODO: Re-enable this after fixing recompilation in quantization.
42
+# && echo TEST_4 \
43
+# && pytest -s -v /workspace/vllm/tests/tpu/test_quantization_accuracy.py \
tests/v1/tpu/test_basic.py
-3
@@ -31,14 +31,12 @@ TENSOR_PARALLEL_SIZES = [1]
31
reason="This is a basic test for TPU only")
32
@pytest.mark.parametrize("model", MODELS)
33
@pytest.mark.parametrize("max_tokens", [5])
34
-@pytest.mark.parametrize("enforce_eager", [True])
34
@pytest.mark.parametrize("tensor_parallel_size", TENSOR_PARALLEL_SIZES)
35
def test_models(
36
vllm_runner: type[VllmRunner],
37
monkeypatch: pytest.MonkeyPatch,
38
model: str,
39
max_tokens: int,
41
- enforce_eager: bool,
40
tensor_parallel_size: int,
41
) -> None:
42
prompt = "The next numbers of the sequence " + ", ".join(
@@ -51,7 +49,6 @@ def test_models(
49
with vllm_runner(
50
model,
51
max_model_len=8192,
54
- enforce_eager=enforce_eager,
52
gpu_memory_utilization=0.7,
53
max_num_seqs=16,
54
tensor_parallel_size=tensor_parallel_size) as vllm_model: