[WhisperPipeline] Batch Size Mismatch Error When Using Beam Search (with iGPU)

[e950280]

Problem Description

When running speech recognition with OpenVINO WhisperPipeline, setting num_beams=2 or higher causes a batch size mismatch error.

Error Message

Traceback (most recent call last):
File "", line 1, in
RuntimeError: Exception from src/core/src/runtime/tensor.cpp:121:
Exception from src/inference/dev_api/openvino/runtime/iremote_tensor.hpp:32:
Not Implemented

Code

from librosa import load
from openvino_genai import WhisperPipeline

audio, _ = load("./recording.wav", sr=16000)

pipe = WhisperPipeline("./whisper_model/whisper-medium/", device="GPU")

results = pipe.generate(
audio,
language="<|en|>",
num_beams=2, # Setting this value to 2 or higher causes the error
)

Expected Behavior

Speech recognition should work normally on the iGPU when Beam Search is enabled (num_beams = 2 or higher).

Confirmation of other information.

1. Using pipe = WhisperPipeline("./whisper_model/whisper-medium/", device="CPU") works.
   #[https://github.com/[WhisperPipeline] Batch Size Mismatch Error When Using Beam Search #2069]
2. Using GPU without beam works. (num_beams = 1)

CPU Information

Intel Core ultra 5 125H

OS Information

PRETTY_NAME="Ubuntu 24.04.1 LTS"

Python Version

python 3.12.3