From 02bd65484630fc024a14a7cb3c5cb3c8b0ff81b8 Mon Sep 17 00:00:00 2001 From: Alex Brooks Date: Wed, 30 Apr 2025 20:51:36 -0600 Subject: [PATCH] [Misc] Rename Audios -> Audio in Qwen2audio Processing (#17507) Signed-off-by: Alex-Brooks --- vllm/model_executor/models/qwen2_audio.py | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-) diff --git a/vllm/model_executor/models/qwen2_audio.py b/vllm/model_executor/models/qwen2_audio.py index 0cb541c6cbb2c..f30bf08ab18bf 100644 --- a/vllm/model_executor/models/qwen2_audio.py +++ b/vllm/model_executor/models/qwen2_audio.py @@ -150,8 +150,15 @@ class Qwen2AudioMultiModalProcessor( mm_data: Mapping[str, object], mm_kwargs: Mapping[str, Any], ) -> BatchFeature: + # NOTE - we rename audios -> audio in mm data because transformers has + # deprecated audios for the qwen2audio processor and will remove + # support for it in transformers 4.54. + audios = mm_data.pop("audios", []) + if audios: + mm_data["audio"] = audios + # Text-only input not supported in composite processor - if not mm_data.get("audios", []): + if not mm_data.get("audio", []): prompt_ids = self.info.get_tokenizer().encode(prompt) prompt_ids = self._apply_hf_processor_tokens_only(prompt_ids) return BatchFeature(dict(input_ids=[prompt_ids]), tensor_type="pt")