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"

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"

eustlb

import torch
import torch.nn as nn
from torch.nn.utils.parametrizations import weight_norm
from torch.nn.utils.parametrize import remove_parametrizations

# Define dtypes to test
dtypes_to_test = [torch.float64, torch.float32, torch.float16]

for dtype in dtypes_to_test:
    print(f"\nTesting with dtype: {dtype}")

eustlb

import torch
import torch.nn as nn
from torch.nn.utils.parametrizations import weight_norm
from torch.nn.utils.parametrize import remove_parametrizations

# Check if CUDA is available
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(f"Using device: {device}")

# 1. Create conv layer and move to device

eustlb

from mistral_common.protocol.instruct.messages import TextChunk, AudioChunk, UserMessage, AssistantMessage, RawAudio
from mistral_common.audio import Audio
from huggingface_hub import hf_hub_download
from openai import OpenAI

# Modify OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = "http://0.0.0.0:8000/v1"
client = OpenAI(
    api_key=openai_api_key,

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" # "cpu"

eustlb

from datasets import load_dataset, Audio
from transformers import KyutaiSpeechToTextProcessor, KyutaiSpeechToTextForConditionalGeneration
import os
import torch
from whisper.normalizers import EnglishTextNormalizer
import jiwer

os.environ["CUDA_VISIBLE_DEVICES"] = "3"
torch_device = "cuda" if torch.cuda.is_available() else "cpu" # "cpu"

eustlb

# ------ install moshi ------
# git clone https://github.com/kyutai-labs/moshi.git
# cd moshi && git checkout 0395bd6c9a95e899c397a68c75f300f3b5409b2c
# uv pip install -e .
# ----------------------------

import torch
from moshi import run_inference

args = {