fix bug

vllm-project · Oct 18, 2024 · 5dff6ea · 5dff6ea
1 parent 2eb71ac
commit 5dff6ea
Show file tree

Hide file tree

Showing 2 changed files with 12 additions and 9 deletions.
diff --git a/examples/offline_inference_audio_language.py b/examples/offline_inference_audio_language.py
@@ -51,10 +51,11 @@ def run_qwen2_audio(question, audio_count):
               max_num_seqs=5,
               limit_mm_per_prompt={"audio": audio_count})
 
+    audio_in_prompt = "".join([f"Audio {idx+1}: <|audio_bos|><|AUDIO|><|audio_eos|>\n" for idx in range(audio_count)])
+
     prompt = ("<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
               "<|im_start|>user\n"
-              "Audio 1: <|audio_bos|><|AUDIO|><|audio_eos|>\n"
-              f"{question}<|im_end|>\n"
+              f"{audio_in_prompt}{question}<|im_end|>\n"
               "<|im_start|>assistant\n")
     stop_token_ids = None
     return llm, prompt, stop_token_ids

diff --git a/vllm/model_executor/models/qwen2_audio.py b/vllm/model_executor/models/qwen2_audio.py
@@ -92,13 +92,13 @@ def dummy_data_for_qwen2_audio(ctx: InputContext, seq_len: int,
 
     audio_token_index = ctx.model_config.hf_config.audio_token_index
 
-    dummy_seqdata = SequenceData.from_token_counts(
+    dummy_seqdata = SequenceData.from_prompt_token_counts(
         (audio_token_index, max_llm_audio_tokens),
         (0, seq_len - max_llm_audio_tokens),
     )
     dummy_audio = np.full((max_llm_audio_tokens * 2 * 2 * 160, ), 0.)
     return dummy_seqdata, {
-        "audio": dummy_audio if num_audios == 1 else [dummy_audio] * num_audios
+        "audio": [(dummy_audio, 16000)] * num_audios
     }
 
 
@@ -165,11 +165,12 @@ def input_processor_for_qwen2_audio(ctx: InputContext,
     multi_modal_data = llm_inputs.get("multi_modal_data")
     if multi_modal_data is None or "audio" not in multi_modal_data:
         return llm_inputs
+    if len(multi_modal_data["audio"]) == 0:
+        return llm_inputs
+    assert isinstance(multi_modal_data['audio'], list) and isinstance(multi_modal_data['audio'][0], tuple)
 
-    audios = multi_modal_data['audio']
     processor = cached_get_processor(ctx.model_config.model)
-    if len(audios) == 0:
-        return llm_inputs
+    audios = [_[0] for _ in multi_modal_data['audio']]
 
     audio_inputs = processor.feature_extractor(audios,
                                                sampling_rate=16000,
@@ -227,14 +228,15 @@ def input_mapper_for_qwen2_audio(
         }
         return batch_data
     try:
-        batch_data = audio_feature_extractor(multi_modal_data,
+        audios = [_[0] for _ in multi_modal_data]
+        batch_data = audio_feature_extractor(audios,
                                              sampling_rate=16000,
                                              return_attention_mask=True,
                                              padding="max_length",
                                              return_tensors="pt").data
         batch_data["feature_attention_mask"] = batch_data.pop("attention_mask")
     except Exception:
-        logger.error("Failed to process audio (%s)", multi_modal_data)
+        logger.error("Failed to process audio (%s)", audios)
         raise
 
     return MultiModalInputs(batch_data)