eustlb

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Reproduce expected outputs for each VoxtralRealtime HF integration test.

Uses vLLM offline inference (as in run_eval.py) to generate reference
transcriptions for every @slow integration test in
test_modeling_voxtral_realtime.py, then saves them to a JSON file.
"""
import json

eustlb

import json
import os

os.environ["CUDA_VISIBLE_DEVICES"] = "2"

import jiwer
import torch
from datasets import Audio, load_dataset
from transformers import (
    VoxtralRealtimeForConditionalGeneration,

eustlb

from transformers import LasrTokenizer
from transformers.tokenization_utils_sentencepiece import SentencePieceExtractor
from huggingface_hub import hf_hub_download
import sentencepiece
from datasets import load_dataset
from tqdm import tqdm

path = hf_hub_download(repo_id='wuketest/lasr_test', filename='spiece.model')
# vocab_ids, vocab_scores, merges = SentencePieceExtractor(path).extract()

eustlb

from transformers import LasrTokenizer, LasrFeatureExtractor, LasrProcessor
from transformers.tokenization_utils_sentencepiece import SentencePieceExtractor
from huggingface_hub import hf_hub_download
import sentencepiece
from datasets import load_dataset
from tqdm import tqdm
import unicodedata
import re

path = hf_hub_download(repo_id='wuketest/lasr_test', filename='spiece.model')

eustlb

from transformers import AutoProcessor, HiggsAudioForConditionalGeneration

model_id = "eustlb/higgs-v2"
processor = AutoProcessor.from_pretrained(model_id, device_map="cuda")
processor.tokenizer.pad_token = processor.tokenizer.eos_token
model = HiggsAudioForConditionalGeneration.from_pretrained(model_id, device_map="cuda")

# single speaker smart voice
conversation = [
    {

eustlb

from transformers import AutoProcessor
from transformers.models.mllama.image_processing_mllama import convert_aspect_ratios_to_ids

# Load the chat template from file
with open("/Users/eustachelebihan/dev/add-higgs-v2/tmp/chat_template.jinja", "r") as f:
    chat_template = f.read()

# Load expected outputs for comparison
with open("/Users/eustachelebihan/dev/add-higgs-v2/expected/single_speaker_with_smart_voice.txt", "r") as f:
    expected_single_speaker_with_smart_voice = f.read()

eustlb

#run: uv pip install nemo_toolkit[asr]
from nemo.collections.asr.modules import AudioToMelSpectrogramPreprocessor
from transformers import ParakeetFeatureExtractor, ParakeetProcessor
from datasets import load_dataset, Audio
import torch
import numpy as np


torch.use_deterministic_algorithms(True)

eustlb

from datasets import load_dataset, Audio
from transformers import VoxtralForConditionalGeneration, VoxtralProcessor
import os
import torch
from whisper.normalizers import EnglishTextNormalizer
import jiwer

os.environ["CUDA_VISIBLE_DEVICES"] = "0"
torch_device = "cuda" if torch.cuda.is_available() else "cpu"

eustlb

STATE_DICT_MAPPING = {
    # Subsampling layer
    r"encoder\.pre_encode\.": r"encoder.subsampling.",
    # Subsampling specific mappings
    r"encoder\.subsampling\.conv\.": r"encoder.subsampling.layers.",
    r"encoder\.subsampling\.out\.": r"encoder.subsampling.linear.",
    # # Positional encoding (skip pe buffer)
    # r"encoder\.pos_enc\.pe$": None,  # Skip buffer
    r"encoder\.pos_enc\.": r"encoder.encode_positions.",
    # Conformer layers - attention (NeMo already uses self_attn)

eustlb

# To install NeMo, run:
# uv pip install git+https://github.com/NVIDIA/NeMo.git@b97e42b3dd1c9bcdf37c81c63220744af474c9c0

from nemo.collections.asr.models import ASRModel
import torch
import os
from datasets import load_dataset
import soundfile as sf

TMP_DIR = "./tmp"