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 8, 03:45 UTC
[ROCm][CI] Removed hard-coded attn backend requirement for Qwen VL (#34753)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Andreas Karatzas committed
Feb 17, 2026 at 21:59 UTC
cef65f0715927d5a5137c34f9a95d6aa5be26cc2
1 file changed
-7
tests/models/multimodal/generation/test_common.py
-7
@@ -169,13 +169,6 @@ VLM_TEST_SETTINGS = {
169
auto_cls=AutoModelForImageTextToText,
170
vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output,
171
patch_hf_runner=model_utils.qwen3_vl_patch_hf_runner,
172
- vllm_runner_kwargs={
173
- "attention_config": {
174
- "backend": "ROCM_AITER_FA",
175
- },
176
- }
177
- if current_platform.is_rocm()
178
- else None,
172
image_size_factors=[(0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
173
marks=[
174
pytest.mark.core_model,