Support batch processing for whisper example (#1723)

Author: jiafatom

benchmark/python/benchmark_multimodal.py

@@ -157,7 +157,11 @@ def run_benchmark(args, model, processor, image, audio, generation_length, max_l
         main_prompt = "What is the meaning of life?"
         prompt = f'{user_prompt}{main_prompt}{prompt_suffix}{assistant_prompt}'        
 
-    prompts = [prompt]
+    if model_type == "whisper":
+        decoder_prompt_tokens = ["<|startoftranscript|>", "<|en|>", "<|transcribe|>", "<|notimestamps|>"]
+        prompts = ["".join(decoder_prompt_tokens)]
+    else:
+        prompts = [prompt]
     inputs = processor(prompts, images=image, audios=audio)
     prompt_length = inputs['input_ids'].shape()[1]
     if args.verbose: print(f"Prompt used: {prompt}")

examples/python/whisper.py

@@ -36,7 +36,7 @@ def run(args: argparse.Namespace):
         readline.set_completer(_complete)
 
         if args.non_interactive:
-            audio_paths = [args.audio]
+            audio_paths = [audio_path.strip() for audio_path in args.audio.split(",")]
         else:
             audio_paths = [audio_path.strip() for audio_path in input("Audio Paths (comma separated): ").split(",")]
         if len(audio_paths) == 0:
@@ -60,6 +60,7 @@ def run(args: argparse.Namespace):
             num_beams=args.num_beams,
             num_return_sequences=args.num_beams,
             max_length=448,
+            batch_size=batch_size,
         )
 
         generator = og.Generator(model, params)