[Frontend] Add automatic language detection for Whisper transcription (#34342)

Signed-off-by: space_check <roman.vuskov@rwth-aachen.de>
Signed-off-by: Roman <45857014+spacecheck@users.noreply.github.com>
Co-authored-by: Nicolò Lucchesi <nicolo.lucchesi@gmail.com>
Co-authored-by: Nicolò Lucchesi <nlucches@redhat.com>
This commit is contained in:
Roman
2026-02-21 04:49:41 -08:00
committed by GitHub
co-authored by Nicolò Lucchesi Nicolò Lucchesi
parent 272b535ab3
commit 98b0205c3c
5 changed files with 249 additions and 10 deletions
@@ -111,6 +111,47 @@ def check_model_available(model: str) -> None:
model_info.check_transformers_version(on_fail="skip")
def test_parse_language_detection_output():
"""Unit test for WhisperForConditionalGeneration.parse_language_detection_output.
No GPU or model loading required.
"""
from unittest.mock import MagicMock
from vllm.model_executor.models.whisper import (
WhisperForConditionalGeneration,
)
cls = WhisperForConditionalGeneration
def make_tokenizer(return_value: str) -> MagicMock:
tok = MagicMock()
tok.decode = MagicMock(return_value=return_value)
return tok
# English
assert (
cls.parse_language_detection_output([50259], make_tokenizer("<|en|>")) == "en"
)
# German
assert (
cls.parse_language_detection_output([50261], make_tokenizer("<|de|>")) == "de"
)
# Unsupported language code
with pytest.raises(AssertionError):
cls.parse_language_detection_output([99999], make_tokenizer("<|xx|>"))
# No special token format
with pytest.raises(AssertionError):
cls.parse_language_detection_output([1], make_tokenizer("hello"))
# Empty token_ids
with pytest.raises((AssertionError, IndexError)):
cls.parse_language_detection_output([], make_tokenizer("anything"))
@pytest.mark.core_model
@pytest.mark.cpu_model
@pytest.mark.parametrize("model", ["openai/whisper-large-v3-turbo"])