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, 02:45 UTC
[ROCm][CI] Enable AITER for failing `test_gpt_oss` test case on MI355 (#36174)
Micah Williamson committed
Mar 7, 2026 at 15:50 UTC
fc4657756ff01fec770433530a5dd2a238e7e034
1 file changed
+8
-1
tests/models/quantization/test_gpt_oss.py
+8
-1
@@ -21,6 +21,7 @@ import lm_eval
21
import pytest
22
from packaging import version
23
24
+from vllm.platforms.rocm import on_gfx950
25
from vllm.utils.torch_utils import cuda_device_count_stateless
26
27
MODEL_ACCURACIES = {
@@ -83,11 +84,17 @@ class EvaluationConfig:
84
@pytest.mark.parametrize("tp_size", [1, 2, 4, 8])
85
@pytest.mark.parametrize("model_name, expected_accuracy", MODEL_ACCURACIES.items())
86
def test_gpt_oss_attention_quantization(
86
- model_name: str, tp_size: int, expected_accuracy: float
87
+ model_name: str,
88
+ tp_size: int,
89
+ expected_accuracy: float,
90
+ monkeypatch: pytest.MonkeyPatch,
91
):
92
if tp_size > cuda_device_count_stateless():
93
pytest.skip("Not enough GPUs to run this test case")
94
95
+ if "amd/gpt-oss-20b-MoE-Quant-W-MXFP4-A-FP8-KV-FP8" in model_name and on_gfx950():
96
+ monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1")
97
+
98
model_args = EvaluationConfig(model_name).get_model_args(tp_size)
99
100
extra_run_kwargs = {