eustlb

"""
Benchmark TDT loss: PyTorch vs CUDA kernel vs NeMo (Numba).

Sweeps over batch sizes and sequence lengths, measuring speed and peak memory.

Usage:
    /raid/eustache/venvs/pr-44171/bin/python benchmark_tdt_loss.py
    /home/eustache_lebihan/pr-44171/NeMo/.venv/bin/python benchmark_tdt_loss.py --nemo-worker
"""

eustlb

"""
Generate the expected TDT loss reference value using NeMo's GPU kernel.

Runs nvidia/parakeet-tdt-0.6b-v3 in eval mode on 2 LibriSpeech samples
with sigma=0 and computes the HF-style mean reduction (per-sample loss
divided by target length, then averaged across the batch).

NeMo commit: 16f469b122 (v2.8.0rc0)
    https://github.com/NVIDIA/NeMo/tree/16f469b122

eustlb

"""
Convert a CohereAsr SentencePiece tokenizer (.model) to a HuggingFace fast tokenizer.

Downloads the tokenizer files from CohereLabs/cohere-transcribe-03-2026 on HuggingFace Hub.
"""

import json
from pathlib import Path

from huggingface_hub import snapshot_download

eustlb

# Reproduce expected outputs for test_integration_longform in
# transformers/tests/models/mimi/test_modeling_mimi.py
#
# This uses the original moshi codebase (https://github.com/kyutai-labs/moshi)
# to generate reference values.
#
# Installation:
#   git clone https://github.com/kyutai-labs/moshi.git
#   uv pip install -e moshi/moshi/
#   uv pip install librosa

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()