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 7, 09:00 UTC
[ROCm][CI] Fix TP size issue for `test_gpt_oss` (#35887)
Signed-off-by: Micah Williamson <micah.williamson@amd.com>
Micah Williamson committed
Mar 3, 2026 at 14:57 UTC
e7213003cbf64d3f35b97d711eb595aa9e47039c
1 file changed
+5
tests/models/quantization/test_gpt_oss.py
+5
@@ -21,6 +21,8 @@ import lm_eval
21
import pytest
22
from packaging import version
23
24
+from vllm.utils.torch_utils import cuda_device_count_stateless
25
+
26
MODEL_ACCURACIES = {
27
# Full quantization: attention linears and MoE linears
28
"amd/gpt-oss-20b-WFP8-AFP8-KVFP8": 0.89,
@@ -83,6 +85,9 @@ class EvaluationConfig:
85
def test_gpt_oss_attention_quantization(
86
model_name: str, tp_size: int, expected_accuracy: float
87
):
88
+ if tp_size > cuda_device_count_stateless():
89
+ pytest.skip("Not enough GPUs to run this test case")
90
+
91
model_args = EvaluationConfig(model_name).get_model_args(tp_size)
92
93
extra_run_kwargs = {