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 12, 06:00 UTC
[multimodal][test] Reduce memory utilization for test_siglip to avoid OOM (#29504)
Signed-off-by: zhxchen17 <zhxchen17@fb.com> Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
Zhengxu Chen committed
Dec 1, 2025 at 07:41 UTC
ad9d656bfa19d937b3ecb4b33b2824a8440bd8f5
1 file changed
+7
-1
tests/models/multimodal/pooling/test_siglip.py
+7
-1
@@ -37,7 +37,12 @@ def _run_test(
37
dtype: str,
38
) -> None:
39
with vllm_runner(
40
- model, runner="pooling", dtype=dtype, enforce_eager=True, max_model_len=64
40
+ model,
41
+ runner="pooling",
42
+ dtype=dtype,
43
+ enforce_eager=True,
44
+ max_model_len=64,
45
+ gpu_memory_utilization=0.7,
46
) as vllm_model:
47
vllm_outputs = vllm_model.embed(input_texts, images=input_images)
48
@@ -134,6 +139,7 @@ def test_models_text_image_no_crash(
139
dtype=dtype,
140
enforce_eager=True,
141
max_model_len=64,
142
+ gpu_memory_utilization=0.7,
143
) as vllm_model:
144
with pytest.raises(ValueError, match="not both"):
145
vllm_model.embed(texts, images=images)