[ROCm][CI] Avoid forcing FlashAttention in the ColPali pooling test (#52565)
Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>
Andreas Karatzas committed
Aug 16, 2026 at 23:08 UTC
311b3513af33bc29b4acb2fde2e9313e5e9966a0
1 file changed
+3
-1
tests/models/multimodal/pooling/test_colpali.py
+3
-1
@@ -19,6 +19,7 @@ from vllm.entrypoints.chat_utils import (
19
ChatCompletionContentPartTextParam,
20
)
21
from vllm.entrypoints.pooling.scoring.typing import ScoreMultiModalParam
22
+from vllm.platforms import current_platform
23
24
from ....conftest import VllmRunner
25
@@ -215,6 +216,7 @@ def _run_multimodal_text_query_image_docs_test(
216
_make_image_mm_param(red_image),
217
_make_image_mm_param(blue_image),
218
]
219
+ attention_backend = "FLASH_ATTN" if current_platform.is_cuda() else None
220
221
with vllm_runner(
222
model,
@@ -223,7 +225,7 @@ def _run_multimodal_text_query_image_docs_test(
225
max_model_len=4096,
226
enforce_eager=True,
227
gpu_memory_utilization=GPU_MEMORY_UTILIZATION,
226
- attention_backend="FLASH_ATTN",
228
+ attention_backend=attention_backend,
229
kernel_config={"enable_flashinfer_autotune": False},
230
) as vllm_model:
231
assert vllm_model.llm.llm_engine.vllm_config.use_v2_model_runner