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 10, 17:45 UTC
[Misc] Enable Paligemma's PrefixLM attention mask computation (#31725)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Isotr0py committed
Jan 6, 2026 at 03:31 UTC
51e38a8e3010ef20c6cb2f33f27fff0a2753aef4
2 files changed
+1
-7
tests/models/multimodal/generation/test_common.py
-4
@@ -121,10 +121,6 @@ VLM_TEST_SETTINGS = {
121
),
122
auto_cls=AutoModelForImageTextToText,
123
vllm_output_post_proc=model_utils.paligemma_vllm_to_hf_output,
124
- dtype="bfloat16",
125
- marks=[
126
- pytest.mark.skip(reason="vLLM does not support PrefixLM attention mask")
127
- ],
124
),
125
"qwen2_5_vl": VLMTestInfo(
126
models=["Qwen/Qwen2.5-VL-3B-Instruct"],
vllm/config/model.py
+1
-3
@@ -1097,9 +1097,7 @@ class ModelConfig:
1097
"""Whether to use bidirectional attention for mm positions."""
1098
MM_PREFIX_LM_MODELS = (
1099
"gemma3",
1100
- # TODO(Isotr0py): Disable paligemma for now before
1101
- # we supports soft cap attention for FlexAttention
1102
- # "paligemma",
1100
+ "paligemma",
1101
)
1102
if not hasattr(self.hf_config, "model_type"):
1103
return False