Upload 22 files

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

ASRDataset.py

@@ -0,0 +1,800 @@
+import datasets
+datasets.config.DOWNLOADED_DATASETS_PATH = "/mnt/jeff/huggingface/data"
+import os
+os.environ['HF_HOME'] = '/mnt/jeff/huggingface'
+
+import json
+import os
+from pathlib import Path
+
+import numpy as np
+import torch
+import sacrebleu
+
+from datasets import load_dataset
+from torch.utils.data import Dataset, ConcatDataset
+from tqdm import tqdm
+from transformers import (
+    BatchFeature,
+)
+import pandas as pd
+import soundfile as sf
+from datasets import Audio
+import random
+from copy import deepcopy
+import torchaudio
+
+ANSWER_SUFFIX = "<|eot_id|>"
+_IGNORE_INDEX = -100
+class BaseAudioDataset(Dataset):
+    def __init__(self, processor, split, sampling_rate=16000, debug=False):
+        self.processor = processor
+        self.training = "train" in split or 'other' in split
+        self.debug = debug
+        self.sampling_rate = sampling_rate
+        self.name = ""
+        
+    def set_dataset_name(self, name):
+        self.name = name
+
+    @staticmethod
+    def filter_corrupted_files(data, audio_field, text_fields, dataset_name, sampling_rate=16000, debug=True):
+        original_size = len(data)
+        
+        data = data.cast_column(audio_field, Audio(decode=False))
+        
+        def identify_corrupted_files(example):
+            try:
+                sf.read(example[audio_field]["path"])
+                
+                for field in text_fields:
+                    if field in example and example[field].replace('"', '') == "":
+                        return False
+                return True
+            except Exception:
+                return False
+        
+        data = data.filter(identify_corrupted_files, num_proc=16)
+        validated_size = len(data)
+        
+        # Audio Decoding
+        data = data.cast_column(audio_field, Audio(sampling_rate=sampling_rate, decode=True))
+        
+        if debug:
+            print(f"Dataset: {dataset_name}")
+            print(f"Original data nums: {original_size}")
+            print(f"After filtering data nums: {validated_size}")
+            print(f"Filtering ratio: {validated_size/original_size:.2%}")
+            
+        return data
+
+    @staticmethod
+    def filter_by_audio_length(data, audio_field, min_sec=2, max_sec=20, debug=True):
+        original_size = len(data)
+        
+        def filter_audio_by_length(example):
+            try:
+                audio = example[audio_field]['array']
+                channel = 1
+                if hasattr(audio, 'ndim') and audio.ndim > 1:
+                    channel = audio.ndim
+                    audio = audio.squeeze()
+                audio_length = len(audio) / example[audio_field]['sampling_rate'] / channel
+                return min_sec <= audio_length <= max_sec
+            except Exception as e:
+                if debug:
+                    print(f"Error : {str(e)[:100]}... - sample excluded")
+                return False
+        
+        data = data.filter(filter_audio_by_length, num_proc=16)
+        filtered_size = len(data)
+        
+        if debug:
+            print(f"Before Length Filtering data nums: {original_size}")
+            print(f"After Length Filtering data nums: {filtered_size}")
+            print(f"Filtering ratio: {filtered_size/original_size:.2%}")
+            
+        return data
+
+    def prepare_model_inputs(self, audio_array, instruction, answer_text):
+        user_message = {
+            'role': 'user',
+            'content': '<start_of_audio>' + instruction,
+        }
+        prompt = self.processor.tokenizer.apply_chat_template(
+            [user_message], tokenize=False, add_generation_prompt=True, add_bos=True
+        )
+        
+        inputs = self.processor(
+            text=prompt, 
+            audio=[audio_array], 
+            return_tensors='pt'
+        )
+        
+        answer = f"{answer_text}{ANSWER_SUFFIX}"
+        answer_ids = self.processor.tokenizer(answer, return_tensors='pt').input_ids
+        
+        if self.debug:
+            self.debug = False
+            task_type = 'AST' if hasattr(self, 'ast') and self.ast else 'ASR'
+            lang_info = f" - {self.lang}" if hasattr(self, 'lang') else ""
+            print(f"{task_type}{lang_info}\nPROMPT: {prompt}\nINPUT: {self.processor.decode(inputs.input_ids[0], skip_special_tokens=False)}\nANSWER: {self.processor.decode(answer_ids[0], skip_special_tokens=False)}\n")
+            print(f"INPUT_MODE: {inputs.input_modes[0].item()}")
+        
+        if self.training:
+            input_ids = torch.cat([inputs.input_ids, answer_ids], dim=1)
+            labels = torch.full_like(input_ids, _IGNORE_INDEX)
+            labels[:, -answer_ids.shape[1]:] = answer_ids
+            padding = torch.zeros((inputs.token_type_ids.shape[0], answer_ids.shape[1]))
+            token_type_ids = torch.cat([inputs.token_type_ids, padding], dim=1)
+        else:
+            input_ids = inputs.input_ids
+            labels = answer_ids
+            token_type_ids = inputs.token_type_ids
+        if self.training:
+            return {
+                'input_ids': input_ids,
+                'labels': labels,
+                'token_type_ids': token_type_ids,
+                'input_audio_embeds': inputs.input_audio_embeds,
+                'audio_embed_sizes': inputs.audio_embed_sizes,
+                'input_modes': inputs.input_modes,
+            }
+        else:
+            return {
+                'input_ids': input_ids,
+                'token_type_ids': token_type_ids,
+                'input_audio_embeds': inputs.input_audio_embeds,
+                'audio_embed_sizes': inputs.audio_embed_sizes,
+                'input_modes': inputs.input_modes,
+                'answer': answer_text,
+            }
+# Libri Speech Dataset Class
+class LibriSpeechDataset(BaseAudioDataset):
+    def __init__(self, processor, subset, split, sampling_rate=16000, debug=False):
+        super().__init__(processor, split, sampling_rate, debug)
+        
+        self.set_dataset_name(f"LibriSpeech_{subset}")
+        # only ASR
+        self.ast = False
+        self.lang = "en"
+        
+        # load dataset
+        self.data = load_dataset("/mnt/jeff/InCar/data/librispeech_asr",
+                            subset,
+                            split=split,
+                            trust_remote_code=True,
+                            cache_dir=Path("/mnt/jeff/InCar/data")
+                            )
+        
+        # (Optional) Audio length Filtering
+        self.data = self.filter_by_audio_length(self.data, "audio")
+            
+        # Instruction Setting
+        self.instruction = random.choice(INSTRUCTION["asr"])
+    
+    def __len__(self):
+        return len(self.data)
+    
+    def __getitem__(self, idx):
+        data = self.data[idx]
+        
+        # Libri Speech is only for ASR
+        answer_text = data["text"].replace('"', '')
+        
+        return self.prepare_model_inputs(
+            data["audio"]["array"],
+            random.choice(INSTRUCTION["asr"]),
+            answer_text
+        )
+    
+# common_voice_16_1 dataset
+class CommonVoiceDataset(BaseAudioDataset):
+    def __init__(self, processor, split, source_lang, sampling_rate=16000, debug=False):
+        super().__init__(processor, split, sampling_rate, debug)
+        
+        self.set_dataset_name(f"CommonVoice_{source_lang}")
+        # only ASR
+        self.ast = False
+        self.lang=source_lang
+        
+        # load dataset
+        if source_lang=="zh-TW":
+            data_path = "/mnt/jeff/InCar/data/common_voice_16_1"
+        else:
+            data_path = "/mnt/jeff/InCar/data/common_voice_17_0"
+        self.data = load_dataset(data_path,
+                            source_lang,
+                            split=split,
+                            trust_remote_code=True,
+                            cache_dir=Path("/mnt/jeff/InCar/data")
+                            )
+        def prepare_dataset(batch):
+            """Function to preprocess the dataset with the .map method"""
+            transcription = batch["sentence"]
+            
+            if transcription.startswith('"') and transcription.endswith('"'):
+                # we can remove trailing quotation marks as they do not affect the transcription
+                transcription = transcription[1:-1]
+            
+            if transcription[-1] not in [".", "?", "!"]:
+                # append a full-stop to sentences that do not end in punctuation
+                transcription = transcription + "."
+            
+            batch["sentence"] = transcription
+            
+            return batch
+
+        
+        import opencc
+        converter = opencc.OpenCC('s2tw.json')
+        def To_zhTW(batch):
+            
+            transcription = converter.convert(batch["sentence"])
+            batch["sentence"] = transcription
+            
+            return batch
+        self.data = self.data.map(prepare_dataset, desc="preprocess dataset")
+        if source_lang=='zh-CN':
+            self.data = self.data.map(To_zhTW, desc="preprocess dataset To_zhTW")
+        
+        
+        # (Optional) Audio length Filtering
+        self.data = self.filter_by_audio_length(self.data, "audio")
+
+        if source_lang == "zh-TW" and split=='train':
+            import torchaudio
+            from torchaudio import transforms
+            import copy
+            import pickle
+            import os
+            def subsample(batch):
+                batch['audio']['array']=torchaudio.functional.resample(torch.FloatTensor(batch['audio']['array']), orig_freq=batch['audio']['sampling_rate'], new_freq=16000)
+                batch['audio']['sampling_rate']=16000
+                return batch
+            def TW_data_augment_fast(batch):
+                speed_perturb_fast = transforms.SpeedPerturbation(batch['audio']['sampling_rate'], [1.1])
+                new_array_fast = speed_perturb_fast(torch.FloatTensor(batch['audio']['array']))[0]
+                batch['audio']['array'] = new_array_fast
+                return batch
+            def TW_data_augment_slow(batch):
+                speed_perturb_slow = transforms.SpeedPerturbation(batch['audio']['sampling_rate'], [0.9])
+                new_array_slow = speed_perturb_slow(torch.FloatTensor(batch['audio']['array']))[0]
+                batch['audio']['array'] = new_array_slow
+                return batch
+            # data = self.data.map(subsample, num_proc=1, desc="subsample")
+            fast_path = '/mnt/jeff/InCar/data/tw_fast.pkl'
+            if not os.path.exists(fast_path):
+                data_fast = self.data.map(TW_data_augment_fast, num_proc=1, desc="augment fast")
+                with open(fast_path,'wb') as f:
+                    pickle.dump(data_fast,f)
+            else:
+                with open(fast_path,'rb') as f:
+                    data_fast=pickle.load(f)
+            
+            slow_path = '/mnt/jeff/InCar/data/data_slow.pkl'
+            if not os.path.exists(slow_path):
+                data_slow = self.data.map(TW_data_augment_slow, num_proc=1, desc="augment slow")
+                with open(slow_path,'wb') as f:
+                    pickle.dump(data_slow,f)
+            else:
+                with open(slow_path,'rb') as f:
+                    data_slow=pickle.load(f)
+            self.data = [d for d in self.data]+[d for d in data_fast]+[d for d in data_slow]
+            
+        # Instruction Setting
+        self.instruction = random.choice(INSTRUCTION["asr"])
+    
+    def __len__(self):
+        return len(self.data)
+    
+    def __getitem__(self, idx):
+        data = self.data[idx]
+        
+        answer_text = data["sentence"]
+        return self.prepare_model_inputs(
+            data["audio"]["array"],
+            random.choice(INSTRUCTION["asr"]),
+            answer_text
+        )
+    
+
+# Fleurs Dataset Class
+class FleursDataset(BaseAudioDataset):
+    def __init__(self, processor, split, source_lang, target_lang=None, 
+                 mode="asr", sampling_rate=16000, debug=False):
+        super().__init__(processor, split, sampling_rate, debug)
+        
+        self.set_dataset_name("Fleurs")
+        # Mode Setting (ASR or AST)
+        if mode not in ["asr", "ast"]:
+            raise ValueError("mode must be 'asr' or 'ast'.")
+        
+        self.mode = mode
+        self.ast = (mode == "ast")
+        self.source_lang = source_lang
+        
+        # Language name mapping (expand if needed)
+        self.lang_names = {
+            'en_us': 'English', 'cmn_hans': 'Mandarin Chinese'
+        }
+        
+        # load dataset - source language dataset
+        self.data = load_dataset("/mnt/jeff/InCar/data/fleurs",
+                            source_lang,
+                            split=split,
+                            trust_remote_code=True,
+                            cache_dir=Path("/mnt/jeff/InCar/data")
+                            )
+        import opencc
+        converter = opencc.OpenCC('s2tw.json')
+        def prepare_dataset(batch):
+            transcription = converter.convert(batch["transcription"])
+            batch["transcription"] = transcription
+            
+            return batch
+        if (source_lang=="cmn_hans_cn"):
+            self.data = self.data.map(prepare_dataset, desc="preprocess dataset")
+
+        # (Optional) Audio length Filtering
+        self.data = self.filter_by_audio_length(self.data, "audio")
+        self.target_lang_name = ""
+        # When AST mode, load target language dataset.
+        if self.ast:
+            if target_lang is None:
+                raise ValueError("AST mode requires target_lang.")
+                
+            self.target_lang = target_lang
+            self.lang = f"{source_lang}_{target_lang}"
+            
+            # load dataset - target language dataset (for translation)
+            target_data = load_dataset("/mnt/jeff/InCar/data/fleurs",
+                                target_lang,
+                                split=split,
+                                trust_remote_code=True,
+                                cache_dir=Path("/mnt/jeff/InCar/data")
+                                )
+            if target_lang=="cmn_hans_cn":
+                target_data=target_data.map(prepare_dataset, desc="preprocess dataset")
+            source_dict = {item['id']: item for item in self.data}
+            target_dict = {item['id']: item for item in target_data}
+            
+            # only Common ID, add translation fields
+            common_ids = set(source_dict.keys()) & set(target_dict.keys())
+            print(f"FLEURS AST Common data filtering: {len(self.data)} -> {len(common_ids)}")
+            self.data = [
+                {**source_dict[id], 'translation': target_dict[id]['transcription']}
+                for id in common_ids
+            ]
+
+            # Instruction Setting - use target language name
+            self.target_lang_name = self.lang_names.get(target_lang, target_lang.capitalize())
+            self.instruction = random.choice(INSTRUCTION["ast"])
+        else:
+            # ASR mode
+            self.lang = source_lang
+            self.instruction = random.choice(INSTRUCTION["asr"])
+
+        if self.debug:
+            print(f"FLEURS dataset loaded: {self.mode.upper()} mode")
+            print(f"source lang: {source_lang} ({self.lang_names.get(source_lang, source_lang)})")
+            if self.ast:
+                print(f"target lang: {target_lang} ({self.lang_names.get(target_lang, target_lang)})")
+            print(f"dataset size: {len(self.data)}")
+    
+    def __len__(self):
+        return len(self.data)
+    
+    def __getitem__(self, idx):
+        data = self.data[idx]
+        audio_array = data["audio"]["array"]
+
+        if self.ast:
+            answer_text = data["translation"]
+            instruction = random.choice(INSTRUCTION["ast"]).format(self.target_lang_name)
+        else:
+            answer_text = data["transcription"]
+            instruction = random.choice(INSTRUCTION["asr"])
+        
+        return self.prepare_model_inputs(
+            audio_array,
+            instruction,
+            answer_text
+        )
+
+class TWCostumData(BaseAudioDataset):
+    
+    def __init__(self, processor, split="train", sampling_rate=16000,csv_path="", debug=False):
+        super().__init__(processor, split, sampling_rate, debug)
+        import pandas as pd
+        from datasets import Dataset, Audio 
+        
+
+        df = pd.read_csv(csv_path).fillna('')
+        
+
+        self.set_dataset_name(f"TWCostumData")
+        self.data = Dataset.from_dict(
+                                    {
+                                        "audio": [audio for audio in df['audio']],
+                                        "sentence": [text for text in df['text']]
+                                    }
+                                ).cast_column("audio", Audio(sampling_rate=16000))
+
+        # Instruction Setting
+        self.instruction = random.choice(INSTRUCTION["asr"])
+    
+    def __len__(self):
+        return len(self.data)
+    
+    def __getitem__(self, idx):
+        data = self.data[idx]
+        
+        answer_text = data["sentence"]
+        return self.prepare_model_inputs(
+            data["audio"]["array"],
+            random.choice(INSTRUCTION["asr"]),
+            answer_text
+        )
+def covost_collate_fn(batch):
+    input_ids_list = []
+    labels_list = []
+    token_type_ids_list = []
+    input_audio_embeds_list = []
+    audio_embed_sizes_list = []
+    audio_attention_mask_list = []
+    input_modes_list = []
+    audio_paths = []
+    for inputs in batch:
+        if 'audio_path' in inputs:
+            audio_paths.append(inputs['audio_path'])
+        input_ids_list.append(inputs['input_ids'][0])
+        labels_list.append(inputs['labels'][0])
+        token_type_ids_list.append(inputs['token_type_ids'][0])
+        if inputs['input_modes']==2:
+            input_audio_embeds_list.append(inputs['input_audio_embeds'])
+            audio_embed_sizes_list.append(inputs['audio_embed_sizes'])
+            audio_attention_mask_list.append(
+                inputs['input_audio_embeds'].new_full((inputs['input_audio_embeds'].size(1),), True, dtype=torch.bool)
+            )
+        # else:
+        #     input_audio_embeds_list.append(None)
+        #     audio_embed_sizes_list.append(None)
+        #     audio_attention_mask_list.append(None)
+        input_modes_list.append(inputs['input_modes'])
+    # try:
+    token_type_ids = pad_sequence(token_type_ids_list, padding_side='left', padding_value=0)
+    input_ids = pad_sequence(input_ids_list, padding_side='left', padding_value=0)
+    labels = pad_sequence(labels_list, padding_side='left', padding_value=0)
+    audio_attention_mask = (
+        pad_sequence(audio_attention_mask_list, padding_side='left', padding_value=False)
+        if len(audio_attention_mask_list) > 1
+        else None
+    )
+    # except Exception as e:
+    #     print(e)
+    #     print(input_ids_list)
+    #     print(labels_list)
+    #     raise
+    attention_mask = (input_ids != 0).long()
+    input_audio_embeds = cat_with_pad(input_audio_embeds_list, dim=0) if len(input_audio_embeds_list)>0 else None
+    audio_embed_sizes = torch.cat(audio_embed_sizes_list) if len(audio_embed_sizes_list)>0 else None
+    input_modes = torch.cat(input_modes_list)
+    if len(audio_paths)>0:
+        return BatchFeature(
+            {
+                "audio_path": audio_paths,
+                'input_ids': input_ids,
+                'labels': labels,
+                'token_type_ids': token_type_ids,
+                'attention_mask': attention_mask,
+                'input_audio_embeds': input_audio_embeds,
+                'audio_embed_sizes': audio_embed_sizes,
+                'audio_attention_mask': audio_attention_mask,
+                'input_modes': input_modes,
+            }
+        )
+    else:
+        return BatchFeature(
+            {
+                'input_ids': input_ids,
+                'labels': labels,
+                'token_type_ids': token_type_ids,
+                'attention_mask': attention_mask,
+                'input_audio_embeds': input_audio_embeds,
+                'audio_embed_sizes': audio_embed_sizes,
+                'audio_attention_mask': audio_attention_mask,
+                'input_modes': input_modes,
+            }
+        )
+
+def pad_sequence(sequences, padding_side='left', padding_value=0):
+    """
+    Pad a list of sequences to the same length.
+    sequences: list of tensors in [seq_len, *] shape
+    """
+    assert padding_side in ['right', 'left']
+    max_size = sequences[0].size()
+    trailing_dims = max_size[1:]
+    max_len = max(len(seq) for seq in sequences)
+    batch_size = len(sequences)
+    output = sequences[0].new_full((batch_size, max_len) + trailing_dims, padding_value)
+    for i, seq in enumerate(sequences):
+        length = seq.size(0)
+        if padding_side == 'right':
+            output.data[i, :length] = seq
+        else:
+            output.data[i, -length:] = seq
+    return output
+
+def cat_with_pad(tensors, dim, padding_value=0):
+    """
+    cat along dim, while pad to max for all other dims
+    """
+    ndim = tensors[0].dim()
+    assert all(
+        t.dim() == ndim for t in tensors[1:]
+    ), 'All tensors must have the same number of dimensions'
+
+    out_size = [max(t.shape[i] for t in tensors) for i in range(ndim)]
+    out_size[dim] = sum(t.shape[dim] for t in tensors)
+    output = tensors[0].new_full(out_size, padding_value)
+
+    index = 0
+    for t in tensors:
+        # Create a slice list where every dimension except dim is full slice
+        slices = [slice(0, t.shape[d]) for d in range(ndim)]
+        # Update only the concat dimension slice
+        slices[dim] = slice(index, index + t.shape[dim])
+
+        output[slices] = t
+        index += t.shape[dim]
+
+    return output
+
+
+
+class MultiturnAudioDataset(BaseAudioDataset):
+    def __init__(self, processor, split="train", sampling_rate=16000,json_path="",text_only=False, debug=False):
+        super().__init__(processor, split, sampling_rate, debug)
+        from llamafactory.data.template import Llama2Template,parse_template
+        from llamafactory.data.formatter import EmptyFormatter, FunctionFormatter, StringFormatter, ToolFormatter
+        from llamafactory.data.mm_plugin import get_mm_plugin
+        import json 
+        self.train=False
+        self.text_only=text_only
+        with open(json_path) as f:
+            js_data = json.load(f)
+        if split=='train':
+            self.train=True
+            js_data = js_data[:int(len(js_data)*0.8)]
+        else:
+            js_data = js_data[-int(len(js_data)*0.2):]
+        for conv in js_data:
+            for mess in conv['conversations']:
+                if 'audio_path' in mess:
+                    mess['audio_path'] = mess['audio_path'].replace('/home/jeff/codes/llm/InCar/srdc_generate_tts/','/mnt/jeff/InCar/data/multiturn_data/')
+        default_system = ""#"""You are a helpful assistant that determines how to solve problems based on user needs and converts user speech into text.\n"""
+        self.template=Llama2Template(
+            format_user=StringFormatter(slots=["<|begin_of_text|>user\n{{content}}<|end_of_text|>\n<|begin_of_text|>model\n"]),
+            format_assistant=StringFormatter(slots=["{{content}}<end_of_turn>\n"]),
+            format_system=StringFormatter(slots=["{{content}}\n\n"]),
+            format_function=FunctionFormatter(slots=["{{content}}", {"eos_token"}], tool_format="default"),
+            format_tools = ToolFormatter(tool_format="default"),
+            format_observation=StringFormatter(
+                slots=["<|begin_of_text|>tool\n{{content}}<|end_of_text|>\n<|begin_of_text|>model\n"]
+            ),
+            default_system=default_system,
+            thought_words=("<think>", "</think>"),
+            efficient_eos=False,
+            replace_eos=False,
+            replace_jinja_template=False,
+            format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
+            stop_words=["<|end_of_text|>"],
+            mm_plugin=get_mm_plugin(name="base"),
+            enable_thinking=False
+        )
+
+        self.set_dataset_name(f"MultiturnCostumData")
+        
+        
+        self.data = []
+        self.text_only_data = []
+        for conv in js_data:
+            tools = conv['tools'] if 'tools' in conv else ""
+            system = conv['system'] if 'system' in conv else default_system
+            tmp = {
+                'tools':tools,
+                'system':system,
+                'messages':[],
+            }
+            for i,mess in enumerate(conv['conversations']):
+                tmp['messages'].append(mess)
+                if mess['from']=='human':
+                    tmp['messages'].append(conv['conversations'][i+1])
+                    d = deepcopy(tmp)
+                    d['audio_array'] = torchaudio.load(mess['audio_path'])[0][0]
+                    self.data.append(d)
+                    if self.text_only:
+                        self.text_only_data.append(deepcopy(tmp))
+                    tmp['messages'].pop()
+                elif mess['from']=='observation':
+                    tmp['messages'].append(conv['conversations'][i+1])
+                    d = deepcopy(tmp)
+                    self.text_only_data.append(d)
+                    tmp['messages'].pop()
+        if text_only:
+            self.data=self.text_only_data
+            
+
+    def prepare_multiturn_model_inputs(self, audio_array, messages, system="", tools=""):
+        prompt = ""
+        answer_text = ""
+        user_transcribe = ""
+        audio_paths = []
+        for i, message in enumerate(messages):
+            elements = []
+            
+            system_text = ""
+            if i == 0:
+                elements += self.template.format_prefix.apply()
+                if system or tools:
+                    tool_text = self.template.format_tools.apply(content=tools)[0] if tools else ""
+                    system_text = self.template.format_system.apply(content=(system + tool_text))[0]
+
+            if message["from"] == "human":
+                if i==len(messages)-2 and not self.text_only:
+                    user_transcribe =  message["value"]
+                    elements += self.template.format_user.apply(content=system_text+'<start_of_audio>')
+                else:
+                    elements += self.template.format_user.apply(content=system_text + message["value"])
+                audio_paths.append(message['audio_path'])
+            elif message["from"] == "gpt":
+                elements += self.template.format_assistant.apply(content=message["value"])
+            elif message["from"] == "observation":
+                elements += self.template.format_observation.apply(content=message["value"])
+            elif message["from"] == "function_call":
+                elements += self.template.format_function.apply(content=message["value"])
+            else:
+                raise NotImplementedError("Unexpected role: {}".format(message["from"]))
+            
+            
+            for elem in elements:
+                ele_str = ""
+                if isinstance(elem, str):
+                    ele_str=elem
+                elif isinstance(elem, set):
+                    if "bos_token" in elem and self.processor.tokenizer.bos_token_id is not None:
+                        ele_str = self.processor.tokenizer.bos_token
+                    elif "eos_token" in elem and self.processor.tokenizer.eos_token_id is not None:
+                        ele_str = self.processor.tokenizer.eos_token
+                if i == len(messages)-1:
+                    answer_text+=ele_str
+                else:
+                    prompt+=ele_str
+            
+
+        if type(audio_array)!=type(None):
+            inputs = self.processor(
+                text=prompt, 
+                audio=[audio_array], 
+                return_tensors='pt'
+            )
+            answer = "\nUser transcribe is : {};\nGPT output is : {}{}".format(user_transcribe,answer_text,ANSWER_SUFFIX)
+        else:
+            inputs = self.processor(
+                text=prompt, 
+                audio=None, 
+                return_tensors='pt'
+            )
+            answer = f"{answer_text}{ANSWER_SUFFIX}"
+        # print('user_transcribe',user_transcribe)
+        # print('answer_text', answer)
+        # print('prompt',prompt)
+        answer_ids = self.processor.tokenizer(answer, return_tensors='pt').input_ids
+        
+        if self.debug:
+            self.debug = False
+            task_type = 'AST' if hasattr(self, 'ast') and self.ast else 'ASR'
+            lang_info = f" - {self.lang}" if hasattr(self, 'lang') else ""
+            print(f"{task_type}{lang_info}\nPROMPT: {prompt}\nINPUT: {self.processor.decode(inputs.input_ids[0], skip_special_tokens=False)}\nANSWER: {self.processor.decode(answer_ids[0], skip_special_tokens=False)}\n")
+            print(f"INPUT_MODE: {inputs.input_modes[0].item()}")
+        
+        if self.training:
+            input_ids = torch.cat([inputs.input_ids, answer_ids], dim=1)
+            labels = torch.full_like(input_ids, _IGNORE_INDEX)
+            labels[:, -answer_ids.shape[1]:] = answer_ids
+            padding = torch.zeros((inputs.token_type_ids.shape[0], answer_ids.shape[1]))
+            token_type_ids = torch.cat([inputs.token_type_ids, padding], dim=1)
+        else:
+            input_ids = inputs.input_ids
+            labels = answer_ids
+            token_type_ids = inputs.token_type_ids
+        if type(audio_array)!=type(None):
+            if not self.train:
+                return {
+                    "audio_path": audio_paths,
+                    'input_ids': input_ids,
+                    'labels': labels,
+                    'token_type_ids': token_type_ids,
+                    'input_audio_embeds': inputs.input_audio_embeds,
+                    'audio_embed_sizes': inputs.audio_embed_sizes,
+                    'input_modes': inputs.input_modes,
+                }
+            else:
+                return {
+                    'input_ids': input_ids,
+                    'labels': labels,
+                    'token_type_ids': token_type_ids,
+                    'input_audio_embeds': inputs.input_audio_embeds,
+                    'audio_embed_sizes': inputs.audio_embed_sizes,
+                    'input_modes': inputs.input_modes,
+                }
+        else:
+            return {
+                'input_ids': input_ids,
+                'labels': labels,
+                'token_type_ids': token_type_ids,
+                'input_audio_embeds': None,
+                'audio_embed_sizes': None,
+                'input_modes': inputs.input_modes,
+            }
+    def __len__(self):
+        return len(self.data)
+    
+    def __getitem__(self, idx):
+        data = self.data[idx]
+        return self.prepare_multiturn_model_inputs(
+            audio_array=data["audio_array"] if "audio_array" in data else None,
+            messages=data['messages'],
+            system=data["system"],
+            tools=data["tools"]
+        )
+
+
+
+os.environ["TOKENIZERS_PARALLELISM"] = "false" 
+
+INSTRUCTION = {
+    "ast": [
+        "Translate the audio to {0}.",
+        "Translate the audio clip into {0}.",
+        "Based on the attached audio, generate a comprehensive {0} translation of the spoken content.",
+        "Translate the provided audio file into {0}.",
+        "Convert the audio speech to {0} text.",
+        "Write an {0} translation of the audio file.",
+        "Translate spoken words from the audio into {0}.",
+        "Create an {0} version of the audio content.",
+        "Produce an accurate {0} translation of the audio.",
+        "Extract speech from the audio and translate it to {0}.",
+        "Turn the audio into readable {0} text.",
+        "Write all spoken content from the audio in {0}.",
+        "Generate an {0} translation of the speech in the file.",
+        "Convert the recording into {0} text.",
+        "Accurately translate the audio recording to {0}.",
+        "Write down dialogue from the given audio in {0}.",
+        "Translate all speech in this audio file to {0}.",
+        "Create an accurate {0} version of the speech.",
+        "Perform a complete {0} translation of the audio."
+    ],
+    "asr": [
+        "Transcribe the audio clip into text.",
+        "Based on the attached audio, generate a comprehensive text transcription of the spoken content.",
+        "Transcribe the provided audio file into text.",
+        "Convert the audio speech to text.",
+        "Write a transcript of the audio file.",
+        "Transcribe spoken words from the audio.",
+        "Create a text version of the audio content.",
+        "Produce a verbatim transcript of the audio.",
+        "Extract and transcribe speech from the audio.",
+        "Turn the audio into readable text.",
+        "Write all spoken words from the audio.",
+        "Generate a transcript of the speech in the file.",
+        "Convert the recording into a text transcript.",
+        "Accurately transcribe the audio recording.",
+        "Write down dialogue from the given audio.",
+        "Transcribe all speech in this audio file.",
+        "Create an accurate text version of the speech.",
+        "Perform a complete transcription of the audio."
+    ],
+}

add_token.ipynb

@@ -0,0 +1,748 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "import transformers\n",
+    "\n",
+    "model_id = \"./\"\n",
+    "tokenizer = transformers.AutoTokenizer.from_pretrained(model_id)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "tokenizer.pad_token_id"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "128256"
+      ]
+     },
+     "execution_count": 5,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "tokenizer.audio_token_id"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "128009"
+      ]
+     },
+     "execution_count": 2,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "tokenizer.eos_token_id"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "PreTrainedTokenizerFast(name_or_path='/mnt/jeff/InCar/LLamaNemotronOmni/NemotronOmni', vocab_size=128000, model_max_length=131072, is_fast=True, padding_side='right', truncation_side='right', special_tokens={'bos_token': '<|begin_of_text|>', 'eos_token': '<|eot_id|>', 'pad_token': '<|eot_id|>', 'additional_special_tokens': ['<audio_soft_token>', '<start_of_audio>', '<end_of_audio>', '']}, clean_up_tokenization_spaces=True, added_tokens_decoder={\n",
+       "\t128000: AddedToken(\"<|begin_of_text|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128001: AddedToken(\"<|end_of_text|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128002: AddedToken(\"<|reserved_special_token_0|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128003: AddedToken(\"<|reserved_special_token_1|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128004: AddedToken(\"<|finetune_right_pad_id|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128005: AddedToken(\"<|reserved_special_token_2|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128006: AddedToken(\"<|start_header_id|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128007: AddedToken(\"<|end_header_id|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128008: AddedToken(\"<|eom_id|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128009: AddedToken(\"<|eot_id|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128010: AddedToken(\"<|python_tag|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128011: AddedToken(\"<|reserved_special_token_3|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128012: AddedToken(\"<|reserved_special_token_4|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128013: AddedToken(\"<|reserved_special_token_5|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128014: AddedToken(\"<|reserved_special_token_6|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128015: AddedToken(\"<|reserved_special_token_7|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128016: AddedToken(\"<|reserved_special_token_8|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128017: AddedToken(\"<|reserved_special_token_9|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128018: AddedToken(\"<|reserved_special_token_10|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128019: AddedToken(\"<|reserved_special_token_11|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128020: AddedToken(\"<|reserved_special_token_12|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128021: AddedToken(\"<|reserved_special_token_13|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128022: AddedToken(\"<|reserved_special_token_14|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128023: AddedToken(\"<|reserved_special_token_15|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128024: AddedToken(\"<|reserved_special_token_16|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128025: AddedToken(\"<|reserved_special_token_17|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128026: AddedToken(\"<|reserved_special_token_18|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128027: AddedToken(\"<|reserved_special_token_19|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128028: AddedToken(\"<|reserved_special_token_20|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128029: AddedToken(\"<|reserved_special_token_21|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128030: AddedToken(\"<|reserved_special_token_22|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128031: AddedToken(\"<|reserved_special_token_23|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128032: AddedToken(\"<|reserved_special_token_24|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128033: AddedToken(\"<|reserved_special_token_25|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128034: AddedToken(\"<|reserved_special_token_26|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128035: AddedToken(\"<|reserved_special_token_27|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128036: AddedToken(\"<|reserved_special_token_28|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128037: AddedToken(\"<|reserved_special_token_29|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128038: AddedToken(\"<|reserved_special_token_30|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128039: AddedToken(\"<|reserved_special_token_31|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128040: AddedToken(\"<|reserved_special_token_32|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128041: AddedToken(\"<|reserved_special_token_33|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128042: AddedToken(\"<|reserved_special_token_34|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128043: AddedToken(\"<|reserved_special_token_35|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128044: AddedToken(\"<|reserved_special_token_36|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128045: AddedToken(\"<|reserved_special_token_37|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128046: AddedToken(\"<|reserved_special_token_38|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128047: AddedToken(\"<|reserved_special_token_39|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128048: AddedToken(\"<|reserved_special_token_40|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128049: AddedToken(\"<|reserved_special_token_41|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128050: AddedToken(\"<|reserved_special_token_42|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128051: AddedToken(\"<|reserved_special_token_43|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128052: AddedToken(\"<|reserved_special_token_44|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128053: AddedToken(\"<|reserved_special_token_45|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128054: AddedToken(\"<|reserved_special_token_46|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128055: AddedToken(\"<|reserved_special_token_47|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128056: AddedToken(\"<|reserved_special_token_48|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128057: AddedToken(\"<|reserved_special_token_49|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128058: AddedToken(\"<|reserved_special_token_50|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128059: AddedToken(\"<|reserved_special_token_51|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128060: AddedToken(\"<|reserved_special_token_52|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128061: AddedToken(\"<|reserved_special_token_53|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128062: AddedToken(\"<|reserved_special_token_54|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128063: AddedToken(\"<|reserved_special_token_55|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128064: AddedToken(\"<|reserved_special_token_56|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128065: AddedToken(\"<|reserved_special_token_57|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128066: AddedToken(\"<|reserved_special_token_58|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128067: AddedToken(\"<|reserved_special_token_59|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128068: AddedToken(\"<|reserved_special_token_60|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128069: AddedToken(\"<|reserved_special_token_61|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128070: AddedToken(\"<|reserved_special_token_62|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128071: AddedToken(\"<|reserved_special_token_63|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128072: AddedToken(\"<|reserved_special_token_64|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128073: AddedToken(\"<|reserved_special_token_65|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128074: AddedToken(\"<|reserved_special_token_66|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128075: AddedToken(\"<|reserved_special_token_67|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128076: AddedToken(\"<|reserved_special_token_68|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128077: AddedToken(\"<|reserved_special_token_69|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128078: AddedToken(\"<|reserved_special_token_70|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128079: AddedToken(\"<|reserved_special_token_71|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128080: AddedToken(\"<|reserved_special_token_72|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128081: AddedToken(\"<|reserved_special_token_73|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128082: AddedToken(\"<|reserved_special_token_74|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128083: AddedToken(\"<|reserved_special_token_75|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128084: AddedToken(\"<|reserved_special_token_76|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128085: AddedToken(\"<|reserved_special_token_77|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128086: AddedToken(\"<|reserved_special_token_78|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128087: AddedToken(\"<|reserved_special_token_79|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128088: AddedToken(\"<|reserved_special_token_80|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128089: AddedToken(\"<|reserved_special_token_81|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128090: AddedToken(\"<|reserved_special_token_82|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128091: AddedToken(\"<|reserved_special_token_83|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128092: AddedToken(\"<|reserved_special_token_84|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128093: AddedToken(\"<|reserved_special_token_85|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128094: AddedToken(\"<|reserved_special_token_86|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128095: AddedToken(\"<|reserved_special_token_87|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128096: AddedToken(\"<|reserved_special_token_88|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128097: AddedToken(\"<|reserved_special_token_89|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128098: AddedToken(\"<|reserved_special_token_90|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128099: AddedToken(\"<|reserved_special_token_91|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128100: AddedToken(\"<|reserved_special_token_92|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128101: AddedToken(\"<|reserved_special_token_93|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128102: AddedToken(\"<|reserved_special_token_94|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128103: AddedToken(\"<|reserved_special_token_95|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128104: AddedToken(\"<|reserved_special_token_96|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128105: AddedToken(\"<|reserved_special_token_97|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128106: AddedToken(\"<|reserved_special_token_98|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128107: AddedToken(\"<|reserved_special_token_99|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128108: AddedToken(\"<|reserved_special_token_100|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128109: AddedToken(\"<|reserved_special_token_101|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128110: AddedToken(\"<|reserved_special_token_102|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128111: AddedToken(\"<|reserved_special_token_103|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128112: AddedToken(\"<|reserved_special_token_104|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128113: AddedToken(\"<|reserved_special_token_105|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128114: AddedToken(\"<|reserved_special_token_106|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128115: AddedToken(\"<|reserved_special_token_107|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128116: AddedToken(\"<|reserved_special_token_108|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128117: AddedToken(\"<|reserved_special_token_109|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128118: AddedToken(\"<|reserved_special_token_110|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128119: AddedToken(\"<|reserved_special_token_111|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128120: AddedToken(\"<|reserved_special_token_112|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128121: AddedToken(\"<|reserved_special_token_113|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128122: AddedToken(\"<|reserved_special_token_114|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128123: AddedToken(\"<|reserved_special_token_115|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128124: AddedToken(\"<|reserved_special_token_116|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128125: AddedToken(\"<|reserved_special_token_117|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128126: AddedToken(\"<|reserved_special_token_118|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128127: AddedToken(\"<|reserved_special_token_119|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128128: AddedToken(\"<|reserved_special_token_120|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128129: AddedToken(\"<|reserved_special_token_121|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128130: AddedToken(\"<|reserved_special_token_122|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128131: AddedToken(\"<|reserved_special_token_123|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128132: AddedToken(\"<|reserved_special_token_124|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128133: AddedToken(\"<|reserved_special_token_125|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128134: AddedToken(\"<|reserved_special_token_126|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128135: AddedToken(\"<|reserved_special_token_127|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128136: AddedToken(\"<|reserved_special_token_128|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128137: AddedToken(\"<|reserved_special_token_129|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128138: AddedToken(\"<|reserved_special_token_130|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128139: AddedToken(\"<|reserved_special_token_131|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128140: AddedToken(\"<|reserved_special_token_132|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128141: AddedToken(\"<|reserved_special_token_133|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128142: AddedToken(\"<|reserved_special_token_134|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128143: AddedToken(\"<|reserved_special_token_135|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128144: AddedToken(\"<|reserved_special_token_136|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128145: AddedToken(\"<|reserved_special_token_137|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128146: AddedToken(\"<|reserved_special_token_138|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128147: AddedToken(\"<|reserved_special_token_139|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128148: AddedToken(\"<|reserved_special_token_140|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128149: AddedToken(\"<|reserved_special_token_141|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128150: AddedToken(\"<|reserved_special_token_142|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128151: AddedToken(\"<|reserved_special_token_143|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128152: AddedToken(\"<|reserved_special_token_144|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128153: AddedToken(\"<|reserved_special_token_145|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128154: AddedToken(\"<|reserved_special_token_146|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128155: AddedToken(\"<|reserved_special_token_147|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128156: AddedToken(\"<|reserved_special_token_148|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128157: AddedToken(\"<|reserved_special_token_149|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128158: AddedToken(\"<|reserved_special_token_150|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128159: AddedToken(\"<|reserved_special_token_151|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128160: AddedToken(\"<|reserved_special_token_152|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128161: AddedToken(\"<|reserved_special_token_153|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128162: AddedToken(\"<|reserved_special_token_154|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128163: AddedToken(\"<|reserved_special_token_155|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128164: AddedToken(\"<|reserved_special_token_156|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128165: AddedToken(\"<|reserved_special_token_157|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128166: AddedToken(\"<|reserved_special_token_158|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128167: AddedToken(\"<|reserved_special_token_159|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128168: AddedToken(\"<|reserved_special_token_160|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128169: AddedToken(\"<|reserved_special_token_161|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128170: AddedToken(\"<|reserved_special_token_162|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128171: AddedToken(\"<|reserved_special_token_163|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128172: AddedToken(\"<|reserved_special_token_164|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128173: AddedToken(\"<|reserved_special_token_165|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128174: AddedToken(\"<|reserved_special_token_166|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128175: AddedToken(\"<|reserved_special_token_167|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128176: AddedToken(\"<|reserved_special_token_168|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128177: AddedToken(\"<|reserved_special_token_169|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128178: AddedToken(\"<|reserved_special_token_170|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128179: AddedToken(\"<|reserved_special_token_171|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128180: AddedToken(\"<|reserved_special_token_172|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128181: AddedToken(\"<|reserved_special_token_173|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128182: AddedToken(\"<|reserved_special_token_174|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128183: AddedToken(\"<|reserved_special_token_175|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128184: AddedToken(\"<|reserved_special_token_176|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128185: AddedToken(\"<|reserved_special_token_177|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128186: AddedToken(\"<|reserved_special_token_178|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128187: AddedToken(\"<|reserved_special_token_179|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128188: AddedToken(\"<|reserved_special_token_180|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128189: AddedToken(\"<|reserved_special_token_181|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128190: AddedToken(\"<|reserved_special_token_182|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128191: AddedToken(\"<|reserved_special_token_183|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128192: AddedToken(\"<|reserved_special_token_184|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128193: AddedToken(\"<|reserved_special_token_185|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128194: AddedToken(\"<|reserved_special_token_186|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128195: AddedToken(\"<|reserved_special_token_187|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128196: AddedToken(\"<|reserved_special_token_188|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128197: AddedToken(\"<|reserved_special_token_189|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128198: AddedToken(\"<|reserved_special_token_190|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128199: AddedToken(\"<|reserved_special_token_191|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128200: AddedToken(\"<|reserved_special_token_192|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128201: AddedToken(\"<|reserved_special_token_193|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128202: AddedToken(\"<|reserved_special_token_194|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128203: AddedToken(\"<|reserved_special_token_195|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128204: AddedToken(\"<|reserved_special_token_196|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128205: AddedToken(\"<|reserved_special_token_197|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128206: AddedToken(\"<|reserved_special_token_198|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128207: AddedToken(\"<|reserved_special_token_199|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128208: AddedToken(\"<|reserved_special_token_200|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128209: AddedToken(\"<|reserved_special_token_201|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128210: AddedToken(\"<|reserved_special_token_202|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128211: AddedToken(\"<|reserved_special_token_203|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128212: AddedToken(\"<|reserved_special_token_204|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128213: AddedToken(\"<|reserved_special_token_205|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128214: AddedToken(\"<|reserved_special_token_206|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128215: AddedToken(\"<|reserved_special_token_207|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128216: AddedToken(\"<|reserved_special_token_208|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128217: AddedToken(\"<|reserved_special_token_209|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128218: AddedToken(\"<|reserved_special_token_210|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128219: AddedToken(\"<|reserved_special_token_211|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128220: AddedToken(\"<|reserved_special_token_212|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128221: AddedToken(\"<|reserved_special_token_213|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128222: AddedToken(\"<|reserved_special_token_214|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128223: AddedToken(\"<|reserved_special_token_215|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128224: AddedToken(\"<|reserved_special_token_216|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128225: AddedToken(\"<|reserved_special_token_217|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128226: AddedToken(\"<|reserved_special_token_218|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128227: AddedToken(\"<|reserved_special_token_219|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128228: AddedToken(\"<|reserved_special_token_220|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128229: AddedToken(\"<|reserved_special_token_221|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128230: AddedToken(\"<|reserved_special_token_222|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128231: AddedToken(\"<|reserved_special_token_223|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128232: AddedToken(\"<|reserved_special_token_224|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128233: AddedToken(\"<|reserved_special_token_225|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128234: AddedToken(\"<|reserved_special_token_226|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128235: AddedToken(\"<|reserved_special_token_227|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128236: AddedToken(\"<|reserved_special_token_228|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128237: AddedToken(\"<|reserved_special_token_229|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128238: AddedToken(\"<|reserved_special_token_230|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128239: AddedToken(\"<|reserved_special_token_231|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128240: AddedToken(\"<|reserved_special_token_232|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128241: AddedToken(\"<|reserved_special_token_233|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128242: AddedToken(\"<|reserved_special_token_234|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128243: AddedToken(\"<|reserved_special_token_235|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128244: AddedToken(\"<|reserved_special_token_236|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128245: AddedToken(\"<|reserved_special_token_237|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128246: AddedToken(\"<|reserved_special_token_238|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128247: AddedToken(\"<|reserved_special_token_239|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128248: AddedToken(\"<|reserved_special_token_240|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128249: AddedToken(\"<|reserved_special_token_241|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128250: AddedToken(\"<|reserved_special_token_242|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128251: AddedToken(\"<|reserved_special_token_243|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128252: AddedToken(\"<|reserved_special_token_244|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128253: AddedToken(\"<|reserved_special_token_245|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128254: AddedToken(\"<|reserved_special_token_246|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128255: AddedToken(\"<|reserved_special_token_247|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128256: AddedToken(\"<audio_soft_token>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128257: AddedToken(\"<start_of_audio>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128258: AddedToken(\"<end_of_audio>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "}\n",
+       ")"
+      ]
+     },
+     "execution_count": 3,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "tokenizer"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "3"
+      ]
+     },
+     "execution_count": 7,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "tokenizer.add_special_tokens({'additional_special_tokens':['<audio_soft_token>','<start_of_audio>','<end_of_audio>','']})"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "PreTrainedTokenizerFast(name_or_path='/mnt/jeff/InCar/LLamaNemotronOmni/Llama-3.1-Nemotron-Nano-4B-v1.1', vocab_size=128000, model_max_length=131072, is_fast=True, padding_side='right', truncation_side='right', special_tokens={'bos_token': '<|begin_of_text|>', 'eos_token': '<|eot_id|>', 'pad_token': '<|eot_id|>', 'additional_special_tokens': ['<audio_soft_token>', '<start_of_audio>', '<end_of_audio>', '']}, clean_up_tokenization_spaces=True, added_tokens_decoder={\n",
+       "\t128000: AddedToken(\"<|begin_of_text|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128001: AddedToken(\"<|end_of_text|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128002: AddedToken(\"<|reserved_special_token_0|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128003: AddedToken(\"<|reserved_special_token_1|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128004: AddedToken(\"<|finetune_right_pad_id|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128005: AddedToken(\"<|reserved_special_token_2|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128006: AddedToken(\"<|start_header_id|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128007: AddedToken(\"<|end_header_id|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128008: AddedToken(\"<|eom_id|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128009: AddedToken(\"<|eot_id|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128010: AddedToken(\"<|python_tag|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128011: AddedToken(\"<|reserved_special_token_3|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128012: AddedToken(\"<|reserved_special_token_4|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128013: AddedToken(\"<|reserved_special_token_5|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128014: AddedToken(\"<|reserved_special_token_6|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128015: AddedToken(\"<|reserved_special_token_7|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128016: AddedToken(\"<|reserved_special_token_8|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128017: AddedToken(\"<|reserved_special_token_9|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128018: AddedToken(\"<|reserved_special_token_10|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128019: AddedToken(\"<|reserved_special_token_11|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128020: AddedToken(\"<|reserved_special_token_12|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128021: AddedToken(\"<|reserved_special_token_13|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128022: AddedToken(\"<|reserved_special_token_14|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128023: AddedToken(\"<|reserved_special_token_15|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128024: AddedToken(\"<|reserved_special_token_16|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128025: AddedToken(\"<|reserved_special_token_17|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128026: AddedToken(\"<|reserved_special_token_18|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128027: AddedToken(\"<|reserved_special_token_19|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128028: AddedToken(\"<|reserved_special_token_20|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128029: AddedToken(\"<|reserved_special_token_21|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128030: AddedToken(\"<|reserved_special_token_22|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128031: AddedToken(\"<|reserved_special_token_23|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128032: AddedToken(\"<|reserved_special_token_24|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128033: AddedToken(\"<|reserved_special_token_25|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128034: AddedToken(\"<|reserved_special_token_26|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128035: AddedToken(\"<|reserved_special_token_27|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128036: AddedToken(\"<|reserved_special_token_28|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128037: AddedToken(\"<|reserved_special_token_29|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128038: AddedToken(\"<|reserved_special_token_30|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128039: AddedToken(\"<|reserved_special_token_31|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128040: AddedToken(\"<|reserved_special_token_32|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128041: AddedToken(\"<|reserved_special_token_33|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128042: AddedToken(\"<|reserved_special_token_34|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128043: AddedToken(\"<|reserved_special_token_35|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128044: AddedToken(\"<|reserved_special_token_36|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128045: AddedToken(\"<|reserved_special_token_37|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128046: AddedToken(\"<|reserved_special_token_38|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128047: AddedToken(\"<|reserved_special_token_39|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128048: AddedToken(\"<|reserved_special_token_40|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128049: AddedToken(\"<|reserved_special_token_41|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128050: AddedToken(\"<|reserved_special_token_42|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128051: AddedToken(\"<|reserved_special_token_43|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128052: AddedToken(\"<|reserved_special_token_44|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128053: AddedToken(\"<|reserved_special_token_45|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128054: AddedToken(\"<|reserved_special_token_46|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128055: AddedToken(\"<|reserved_special_token_47|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128056: AddedToken(\"<|reserved_special_token_48|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128057: AddedToken(\"<|reserved_special_token_49|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128058: AddedToken(\"<|reserved_special_token_50|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128059: AddedToken(\"<|reserved_special_token_51|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128060: AddedToken(\"<|reserved_special_token_52|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128061: AddedToken(\"<|reserved_special_token_53|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128062: AddedToken(\"<|reserved_special_token_54|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128063: AddedToken(\"<|reserved_special_token_55|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128064: AddedToken(\"<|reserved_special_token_56|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128065: AddedToken(\"<|reserved_special_token_57|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128066: AddedToken(\"<|reserved_special_token_58|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128067: AddedToken(\"<|reserved_special_token_59|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128068: AddedToken(\"<|reserved_special_token_60|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128069: AddedToken(\"<|reserved_special_token_61|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128070: AddedToken(\"<|reserved_special_token_62|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128071: AddedToken(\"<|reserved_special_token_63|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128072: AddedToken(\"<|reserved_special_token_64|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128073: AddedToken(\"<|reserved_special_token_65|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128074: AddedToken(\"<|reserved_special_token_66|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128075: AddedToken(\"<|reserved_special_token_67|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128076: AddedToken(\"<|reserved_special_token_68|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128077: AddedToken(\"<|reserved_special_token_69|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128078: AddedToken(\"<|reserved_special_token_70|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128079: AddedToken(\"<|reserved_special_token_71|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128080: AddedToken(\"<|reserved_special_token_72|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128081: AddedToken(\"<|reserved_special_token_73|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128082: AddedToken(\"<|reserved_special_token_74|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128083: AddedToken(\"<|reserved_special_token_75|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128084: AddedToken(\"<|reserved_special_token_76|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128085: AddedToken(\"<|reserved_special_token_77|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128086: AddedToken(\"<|reserved_special_token_78|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128087: AddedToken(\"<|reserved_special_token_79|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128088: AddedToken(\"<|reserved_special_token_80|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128089: AddedToken(\"<|reserved_special_token_81|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128090: AddedToken(\"<|reserved_special_token_82|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128091: AddedToken(\"<|reserved_special_token_83|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128092: AddedToken(\"<|reserved_special_token_84|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128093: AddedToken(\"<|reserved_special_token_85|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128094: AddedToken(\"<|reserved_special_token_86|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128095: AddedToken(\"<|reserved_special_token_87|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128096: AddedToken(\"<|reserved_special_token_88|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128097: AddedToken(\"<|reserved_special_token_89|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128098: AddedToken(\"<|reserved_special_token_90|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128099: AddedToken(\"<|reserved_special_token_91|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128100: AddedToken(\"<|reserved_special_token_92|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128101: AddedToken(\"<|reserved_special_token_93|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128102: AddedToken(\"<|reserved_special_token_94|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128103: AddedToken(\"<|reserved_special_token_95|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128104: AddedToken(\"<|reserved_special_token_96|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128105: AddedToken(\"<|reserved_special_token_97|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128106: AddedToken(\"<|reserved_special_token_98|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128107: AddedToken(\"<|reserved_special_token_99|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128108: AddedToken(\"<|reserved_special_token_100|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128109: AddedToken(\"<|reserved_special_token_101|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128110: AddedToken(\"<|reserved_special_token_102|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128111: AddedToken(\"<|reserved_special_token_103|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128112: AddedToken(\"<|reserved_special_token_104|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128113: AddedToken(\"<|reserved_special_token_105|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128114: AddedToken(\"<|reserved_special_token_106|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128115: AddedToken(\"<|reserved_special_token_107|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128116: AddedToken(\"<|reserved_special_token_108|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128117: AddedToken(\"<|reserved_special_token_109|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128118: AddedToken(\"<|reserved_special_token_110|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128119: AddedToken(\"<|reserved_special_token_111|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128120: AddedToken(\"<|reserved_special_token_112|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128121: AddedToken(\"<|reserved_special_token_113|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128122: AddedToken(\"<|reserved_special_token_114|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128123: AddedToken(\"<|reserved_special_token_115|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128124: AddedToken(\"<|reserved_special_token_116|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128125: AddedToken(\"<|reserved_special_token_117|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128126: AddedToken(\"<|reserved_special_token_118|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128127: AddedToken(\"<|reserved_special_token_119|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128128: AddedToken(\"<|reserved_special_token_120|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128129: AddedToken(\"<|reserved_special_token_121|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128130: AddedToken(\"<|reserved_special_token_122|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128131: AddedToken(\"<|reserved_special_token_123|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128132: AddedToken(\"<|reserved_special_token_124|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128133: AddedToken(\"<|reserved_special_token_125|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128134: AddedToken(\"<|reserved_special_token_126|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128135: AddedToken(\"<|reserved_special_token_127|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128136: AddedToken(\"<|reserved_special_token_128|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128137: AddedToken(\"<|reserved_special_token_129|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128138: AddedToken(\"<|reserved_special_token_130|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128139: AddedToken(\"<|reserved_special_token_131|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128140: AddedToken(\"<|reserved_special_token_132|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128141: AddedToken(\"<|reserved_special_token_133|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128142: AddedToken(\"<|reserved_special_token_134|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128143: AddedToken(\"<|reserved_special_token_135|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128144: AddedToken(\"<|reserved_special_token_136|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128145: AddedToken(\"<|reserved_special_token_137|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128146: AddedToken(\"<|reserved_special_token_138|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128147: AddedToken(\"<|reserved_special_token_139|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128148: AddedToken(\"<|reserved_special_token_140|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128149: AddedToken(\"<|reserved_special_token_141|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128150: AddedToken(\"<|reserved_special_token_142|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128151: AddedToken(\"<|reserved_special_token_143|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128152: AddedToken(\"<|reserved_special_token_144|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128153: AddedToken(\"<|reserved_special_token_145|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128154: AddedToken(\"<|reserved_special_token_146|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128155: AddedToken(\"<|reserved_special_token_147|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128156: AddedToken(\"<|reserved_special_token_148|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128157: AddedToken(\"<|reserved_special_token_149|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128158: AddedToken(\"<|reserved_special_token_150|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128159: AddedToken(\"<|reserved_special_token_151|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128160: AddedToken(\"<|reserved_special_token_152|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128161: AddedToken(\"<|reserved_special_token_153|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128162: AddedToken(\"<|reserved_special_token_154|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128163: AddedToken(\"<|reserved_special_token_155|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128164: AddedToken(\"<|reserved_special_token_156|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128165: AddedToken(\"<|reserved_special_token_157|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128166: AddedToken(\"<|reserved_special_token_158|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128167: AddedToken(\"<|reserved_special_token_159|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128168: AddedToken(\"<|reserved_special_token_160|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128169: AddedToken(\"<|reserved_special_token_161|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128170: AddedToken(\"<|reserved_special_token_162|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128171: AddedToken(\"<|reserved_special_token_163|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128172: AddedToken(\"<|reserved_special_token_164|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128173: AddedToken(\"<|reserved_special_token_165|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128174: AddedToken(\"<|reserved_special_token_166|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128175: AddedToken(\"<|reserved_special_token_167|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128176: AddedToken(\"<|reserved_special_token_168|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128177: AddedToken(\"<|reserved_special_token_169|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128178: AddedToken(\"<|reserved_special_token_170|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128179: AddedToken(\"<|reserved_special_token_171|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128180: AddedToken(\"<|reserved_special_token_172|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128181: AddedToken(\"<|reserved_special_token_173|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128182: AddedToken(\"<|reserved_special_token_174|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128183: AddedToken(\"<|reserved_special_token_175|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128184: AddedToken(\"<|reserved_special_token_176|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128185: AddedToken(\"<|reserved_special_token_177|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128186: AddedToken(\"<|reserved_special_token_178|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128187: AddedToken(\"<|reserved_special_token_179|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128188: AddedToken(\"<|reserved_special_token_180|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128189: AddedToken(\"<|reserved_special_token_181|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128190: AddedToken(\"<|reserved_special_token_182|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128191: AddedToken(\"<|reserved_special_token_183|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128192: AddedToken(\"<|reserved_special_token_184|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128193: AddedToken(\"<|reserved_special_token_185|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128194: AddedToken(\"<|reserved_special_token_186|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128195: AddedToken(\"<|reserved_special_token_187|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128196: AddedToken(\"<|reserved_special_token_188|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128197: AddedToken(\"<|reserved_special_token_189|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128198: AddedToken(\"<|reserved_special_token_190|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128199: AddedToken(\"<|reserved_special_token_191|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128200: AddedToken(\"<|reserved_special_token_192|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128201: AddedToken(\"<|reserved_special_token_193|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128202: AddedToken(\"<|reserved_special_token_194|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128203: AddedToken(\"<|reserved_special_token_195|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128204: AddedToken(\"<|reserved_special_token_196|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128205: AddedToken(\"<|reserved_special_token_197|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128206: AddedToken(\"<|reserved_special_token_198|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128207: AddedToken(\"<|reserved_special_token_199|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128208: AddedToken(\"<|reserved_special_token_200|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128209: AddedToken(\"<|reserved_special_token_201|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128210: AddedToken(\"<|reserved_special_token_202|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128211: AddedToken(\"<|reserved_special_token_203|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128212: AddedToken(\"<|reserved_special_token_204|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128213: AddedToken(\"<|reserved_special_token_205|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128214: AddedToken(\"<|reserved_special_token_206|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128215: AddedToken(\"<|reserved_special_token_207|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128216: AddedToken(\"<|reserved_special_token_208|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128217: AddedToken(\"<|reserved_special_token_209|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128218: AddedToken(\"<|reserved_special_token_210|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128219: AddedToken(\"<|reserved_special_token_211|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128220: AddedToken(\"<|reserved_special_token_212|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128221: AddedToken(\"<|reserved_special_token_213|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128222: AddedToken(\"<|reserved_special_token_214|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128223: AddedToken(\"<|reserved_special_token_215|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128224: AddedToken(\"<|reserved_special_token_216|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128225: AddedToken(\"<|reserved_special_token_217|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128226: AddedToken(\"<|reserved_special_token_218|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128227: AddedToken(\"<|reserved_special_token_219|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128228: AddedToken(\"<|reserved_special_token_220|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128229: AddedToken(\"<|reserved_special_token_221|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128230: AddedToken(\"<|reserved_special_token_222|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128231: AddedToken(\"<|reserved_special_token_223|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128232: AddedToken(\"<|reserved_special_token_224|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128233: AddedToken(\"<|reserved_special_token_225|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128234: AddedToken(\"<|reserved_special_token_226|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128235: AddedToken(\"<|reserved_special_token_227|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128236: AddedToken(\"<|reserved_special_token_228|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128237: AddedToken(\"<|reserved_special_token_229|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128238: AddedToken(\"<|reserved_special_token_230|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128239: AddedToken(\"<|reserved_special_token_231|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128240: AddedToken(\"<|reserved_special_token_232|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128241: AddedToken(\"<|reserved_special_token_233|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128242: AddedToken(\"<|reserved_special_token_234|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128243: AddedToken(\"<|reserved_special_token_235|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128244: AddedToken(\"<|reserved_special_token_236|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128245: AddedToken(\"<|reserved_special_token_237|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128246: AddedToken(\"<|reserved_special_token_238|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128247: AddedToken(\"<|reserved_special_token_239|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128248: AddedToken(\"<|reserved_special_token_240|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128249: AddedToken(\"<|reserved_special_token_241|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128250: AddedToken(\"<|reserved_special_token_242|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128251: AddedToken(\"<|reserved_special_token_243|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128252: AddedToken(\"<|reserved_special_token_244|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128253: AddedToken(\"<|reserved_special_token_245|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128254: AddedToken(\"<|reserved_special_token_246|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128255: AddedToken(\"<|reserved_special_token_247|>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128256: AddedToken(\"<audio_soft_token>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128257: AddedToken(\"<start_of_audio>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "\t128258: AddedToken(\"<end_of_audio>\", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True),\n",
+       "}\n",
+       ")"
+      ]
+     },
+     "execution_count": 8,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "tokenizer"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "tokenizer.audio_token = \"<audio_soft_token>\"\n",
+    "tokenizer.boa_token = \"<start_of_audio>\"\n",
+    "tokenizer.eoa_token = \"<end_of_audio>\"\n",
+    "tokenizer.audio_token_id = 128256"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "('../tokenizer/tokenizer_config.json',\n",
+       " '../tokenizer/special_tokens_map.json',\n",
+       " '../tokenizer/tokenizer.json')"
+      ]
+     },
+     "execution_count": 7,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "tokenizer.save_pretrained('../tokenizer')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "You are using a model of type NemotronOmni to instantiate a model of type nemotronOmni. This is not supported for all configurations of models and can yield errors.\n"
+     ]
+    }
+   ],
+   "source": [
+    "from transformers import AutoConfig\n",
+    "config = AutoConfig.from_pretrained('/mnt/jeff/InCar/LlamaNemotronOmni/Llama-3.1-NemotronOmni')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "'llama'"
+      ]
+     },
+     "execution_count": 7,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "config.model_type"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "base",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.12.7"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}

build_model.ipynb

@@ -0,0 +1,525 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "[2025-06-18 08:09:49,587] [INFO] [real_accelerator.py:239:get_accelerator] Setting ds_accelerator to cuda (auto detect)\n"
+     ]
+    },
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "/mnt/jeff/anaconda/compiler_compat/ld: cannot find -laio: No such file or directory\n",
+      "collect2: error: ld returned 1 exit status\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: warning: librt.so.1, needed by /usr/local/cuda/lib64/libcufile.so, not found (try using -rpath or -rpath-link)\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: warning: libpthread.so.0, needed by /usr/local/cuda/lib64/libcufile.so, not found (try using -rpath or -rpath-link)\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: warning: libstdc++.so.6, needed by /usr/local/cuda/lib64/libcufile.so, not found (try using -rpath or -rpath-link)\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: warning: libm.so.6, needed by /usr/local/cuda/lib64/libcufile.so, not found (try using -rpath or -rpath-link)\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::runtime_error::~runtime_error()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `__gxx_personality_v0@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream::tellp()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::chrono::_V2::steady_clock::now()@GLIBCXX_3.4.19'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::_M_replace_aux(unsigned long, unsigned long, unsigned long, char)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for bool@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__throw_logic_error(char const*)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `VTT for std::basic_ostringstream<char, std::char_traits<char>, std::allocator<char> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::logic_error@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::locale::~locale()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_string<char, std::char_traits<char>, std::allocator<char> >::basic_string(std::string const&, unsigned long, unsigned long)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `__cxa_end_catch@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `VTT for std::basic_ofstream<char, std::char_traits<char> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::logic_error::~logic_error()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for __cxxabiv1::__si_class_type_info@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ios<char, std::char_traits<char> >::_M_cache_locale(std::locale const&)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `VTT for std::basic_stringstream<char, std::char_traits<char>, std::allocator<char> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `operator new[](unsigned long)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::_M_leak_hard()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_ifstream<char, std::char_traits<char> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_streambuf<wchar_t, std::char_traits<wchar_t> >::basic_streambuf(std::basic_streambuf<wchar_t, std::char_traits<wchar_t> > const&)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::append(char const*, unsigned long)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_string<char, std::char_traits<char>, std::allocator<char> >::basic_string(std::string const&)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for unsigned short@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::resize(unsigned long, char)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for char const*@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ctype<char>::_M_widen_init() const@GLIBCXX_3.4.11'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__throw_invalid_argument(char const*)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::locale::operator=(std::locale const&)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ios<wchar_t, std::char_traits<wchar_t> >::_M_cache_locale(std::locale const&)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::_Rb_tree_decrement(std::_Rb_tree_node_base const*)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `__cxa_free_exception@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::condition_variable::notify_one()@GLIBCXX_3.4.11'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ios_base::Init::~Init()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_string<char, std::char_traits<char>, std::allocator<char> >::~basic_string()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `__cxa_pure_virtual@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream::flush()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for __cxxabiv1::__class_type_info@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `__cxa_rethrow@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_stringbuf<char, std::char_traits<char>, std::allocator<char> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_fstream<char, std::char_traits<char> >::~basic_fstream()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::compare(char const*) const@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `VTT for std::basic_ostringstream<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::locale::locale()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::chrono::_V2::system_clock::now()@GLIBCXX_3.4.19'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `VTT for std::basic_ifstream<char, std::char_traits<char> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::_Hash_bytes(void const*, unsigned long, unsigned long)@CXXABI_1.3.5'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream& std::ostream::_M_insert<long long>(long long)@GLIBCXX_3.4.9'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for char*@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__detail::_Prime_rehash_policy::_M_need_rehash(unsigned long, unsigned long, unsigned long) const@GLIBCXX_3.4.18'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::out_of_range@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream& std::ostream::_M_insert<unsigned long>(unsigned long)@GLIBCXX_3.4.9'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::_Rb_tree_increment(std::_Rb_tree_node_base const*)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ios_base::~ios_base()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::range_error::~range_error()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__basic_file<char>::~__basic_file()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `__cxa_guard_acquire@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream& std::ostream::_M_insert<bool>(bool)@GLIBCXX_3.4.9'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::overflow_error@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `VTT for std::basic_fstream<char, std::char_traits<char> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::range_error@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_ios<char, std::char_traits<char> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_filebuf<char, std::char_traits<char> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `operator delete[](void*)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_stringstream<char, std::char_traits<char>, std::allocator<char> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_string<char, std::char_traits<char>, std::allocator<char> >::basic_string(unsigned long, char, std::allocator<char> const&)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__detail::_List_node_base::_M_transfer(std::__detail::_List_node_base*, std::__detail::_List_node_base*)@GLIBCXX_3.4.15'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::replace(unsigned long, unsigned long, char const*, unsigned long)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for std::exception@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_string<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >::_Rep::_M_destroy(std::allocator<wchar_t> const&)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::istream& std::istream::_M_extract<double>(double&)@GLIBCXX_3.4.9'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_filebuf<char, std::char_traits<char> >::close()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_fstream<char, std::char_traits<char> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ifstream<char, std::char_traits<char> >::basic_ifstream(char const*, std::_Ios_Openmode)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::append(std::string const&)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `operator new(unsigned long)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `VTT for std::basic_istringstream<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for unsigned int@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::append(char const*)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::domain_error@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::find(char, unsigned long) const@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream::put(char)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for int@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__throw_bad_alloc()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `__cxa_thread_atexit@CXXABI_1.3.7'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for unsigned int*@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::_Rb_tree_increment(std::_Rb_tree_node_base*)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ifstream<char, std::char_traits<char> >::~basic_ifstream()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ios_base::Init::Init()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::condition_variable::condition_variable()@GLIBCXX_3.4.11'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_filebuf<char, std::char_traits<char> >::basic_filebuf()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `VTT for std::basic_istringstream<char, std::char_traits<char>, std::allocator<char> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::domain_error::~domain_error()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::cerr@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::find(char const*, unsigned long, unsigned long) const@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_istringstream<char, std::char_traits<char>, std::allocator<char> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_string<char, std::char_traits<char>, std::allocator<char> >::basic_string(std::allocator<char> const&)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_stringbuf<char, std::char_traits<char>, std::allocator<char> >::str() const@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::invalid_argument@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for void*@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::assign(std::string const&)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ostringstream<char, std::char_traits<char>, std::allocator<char> >::~basic_ostringstream()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::_Rb_tree_rebalance_for_erase(std::_Rb_tree_node_base*, std::_Rb_tree_node_base&)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for unsigned long@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__detail::_List_node_base::_M_hook(std::__detail::_List_node_base*)@GLIBCXX_3.4.15'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__detail::_List_node_base::_M_unhook()@GLIBCXX_3.4.15'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_ostringstream<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_stringbuf<char, std::char_traits<char>, std::allocator<char> >::_M_sync(char*, unsigned long, unsigned long)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_iostream<char, std::char_traits<char> >::~basic_iostream()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::locale::locale(std::locale const&)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_istringstream<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `log2f@GLIBC_2.2.5'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream::operator<<(std::basic_streambuf<char, std::char_traits<char> >*)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_streambuf<wchar_t, std::char_traits<wchar_t> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::exception::~exception()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::_Rep::_S_create(unsigned long, unsigned long, std::allocator<char> const&)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__basic_file<char>::is_open() const@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_istringstream<char, std::char_traits<char>, std::allocator<char> >::~basic_istringstream()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::swap(std::string&)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_ostringstream<char, std::char_traits<char>, std::allocator<char> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_streambuf<char, std::char_traits<char> >::basic_streambuf(std::basic_streambuf<char, std::char_traits<char> > const&)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ios<char, std::char_traits<char> >::init(std::basic_streambuf<char, std::char_traits<char> >*)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__throw_bad_cast()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ios<char, std::char_traits<char> >::clear(std::_Ios_Iostate)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_streambuf<wchar_t, std::char_traits<wchar_t> >::operator=(std::basic_streambuf<wchar_t, std::char_traits<wchar_t> > const&)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for long*@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `operator delete(void*)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream::operator<<(int)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::_Rep::_S_empty_rep_storage@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::_Rep::_M_destroy(std::allocator<char> const&)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_iostream<wchar_t, std::char_traits<wchar_t> >::~basic_iostream()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::runtime_error@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_ofstream<char, std::char_traits<char> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::_Rb_tree_insert_and_rebalance(bool, std::_Rb_tree_node_base*, std::_Rb_tree_node_base*, std::_Rb_tree_node_base&)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_stringstream<char, std::char_traits<char>, std::allocator<char> >::~basic_stringstream()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `VTT for std::basic_stringstream<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream& std::ostream::_M_insert<long>(long)@GLIBCXX_3.4.9'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::istream::get()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for unsigned long long@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ostream<char, std::char_traits<char> >& std::operator<< <std::char_traits<char> >(std::basic_ostream<char, std::char_traits<char> >&, char const*)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::out_of_range::~out_of_range()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::length_error::~length_error()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ostream<char, std::char_traits<char> >& std::__ostream_insert<char, std::char_traits<char> >(std::basic_ostream<char, std::char_traits<char> >&, char const*, long)@GLIBCXX_3.4.9'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::invalid_argument::~invalid_argument()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_string<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >::swap(std::basic_string<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >&)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::cout@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream& std::ostream::_M_insert<unsigned long long>(unsigned long long)@GLIBCXX_3.4.9'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream& std::ostream::_M_insert<void const*>(void const*)@GLIBCXX_3.4.9'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::underflow_error@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_streambuf<char, std::char_traits<char> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for std::out_of_range@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `__cxa_allocate_exception@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_ios<wchar_t, std::char_traits<wchar_t> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for void const*@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ios<wchar_t, std::char_traits<wchar_t> >::init(std::basic_streambuf<wchar_t, std::char_traits<wchar_t> >*)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::reserve(unsigned long)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `__cxa_begin_catch@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for long@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_string<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >::_Rep::_S_empty_rep_storage@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::_M_leak()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_filebuf<char, std::char_traits<char> >::open(char const*, std::_Ios_Openmode)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_stringbuf<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >::_M_sync(wchar_t*, unsigned long, unsigned long)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::istream::getline(char*, long, char)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_istream<char, std::char_traits<char> >& std::getline<char, std::char_traits<char>, std::allocator<char> >(std::basic_istream<char, std::char_traits<char> >&, std::basic_string<char, std::char_traits<char>, std::allocator<char> >&, char)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_stringstream<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::condition_variable::~condition_variable()@GLIBCXX_3.4.11'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_stringbuf<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::insert(unsigned long, char const*, unsigned long)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::assign(char const*, unsigned long)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for unsigned char@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ios_base::ios_base()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__throw_out_of_range(char const*)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::overflow_error::~overflow_error()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__throw_length_error(char const*)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__throw_system_error(int)@GLIBCXX_3.4.11'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ofstream<char, std::char_traits<char> >::close()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream& std::ostream::_M_insert<double>(double)@GLIBCXX_3.4.9'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_streambuf<char, std::char_traits<char> >::operator=(std::basic_streambuf<char, std::char_traits<char> > const&)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for long long@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_string<char, std::char_traits<char>, std::allocator<char> >::basic_string(char const*, unsigned long, std::allocator<char> const&)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ifstream<char, std::char_traits<char> >::close()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `__cxa_guard_release@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `__cxa_throw@CXXABI_1.3'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::underflow_error::~underflow_error()@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::_Rb_tree_decrement(std::_Rb_tree_node_base*)@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::length_error@GLIBCXX_3.4'\n",
+      "/mnt/jeff/anaconda/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_filebuf<char, std::char_traits<char> >::~basic_filebuf()@GLIBCXX_3.4'\n",
+      "collect2: error: ld returned 1 exit status\n",
+      "/mnt/jeff/huggingface/modules/transformers_modules/speech_conformer_encoder.py:2775: FutureWarning: Please specify CheckpointImpl.NO_REENTRANT as CheckpointImpl.REENTRANT will soon be removed as the default and eventually deprecated.\n",
+      "  lambda i: encoder_checkpoint_wrapper(\n"
+     ]
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "b1f758d955e44dce9aac35eed8125134",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "Loading checkpoint shards:   0%|          | 0/5 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "from io import BytesIO\n",
+    "from urllib.request import urlopen\n",
+    "import soundfile\n",
+    "import torch\n",
+    "from datasets import load_dataset, Audio\n",
+    "import numpy as np\n",
+    "from transformers import AutoModel, AutoProcessor, BatchFeature, AutoConfig\n",
+    "from tqdm import tqdm\n",
+    "import json\n",
+    "import os\n",
+    "import time\n",
+    "from datetime import datetime\n",
+    "import sacrebleu\n",
+    "from jiwer import cer, wer\n",
+    "from torch.utils.data import Dataset, DataLoader\n",
+    "import soundfile as sf\n",
+    "import re\n",
+    "from pathlib import Path\n",
+    "\n",
+    "\n",
+    "model_id = \"./\"\n",
+    "\n",
+    "model = AutoModel.from_pretrained(\n",
+    "    model_id, device_map=\"cpu\", trust_remote_code=True,\n",
+    ").eval()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "2d03eb752659479babc4b0095c2cce97",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "Loading checkpoint shards:   0%|          | 0/2 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "from transformers.models.nemotron.modeling_nemotron import NemotronForCausalLM\n",
+    "from transformers import AutoModelForCausalLM\n",
+    "model_id = '/mnt/jeff/InCar/LlamaNemotronOmni/Llama-3.1-Nemotron-Nano-4B-v1.1'\n",
+    "revision = \"main\" #\"v1.0\"\n",
+    "model_org = AutoModelForCausalLM.from_pretrained(\n",
+    "    model_id, device_map=\"cpu\", revision = revision, trust_remote_code=True,\n",
+    ").eval()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "(291, 291)"
+      ]
+     },
+     "execution_count": 3,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "keys = []\n",
+    "for k in model.state_dict():\n",
+    "    if 'language_model' in k and not 'lora' in k:\n",
+    "        keys.append(k)\n",
+    "len(model_org.state_dict()),len(keys)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "new_state_dict = model.state_dict()\n",
+    "for k in keys:\n",
+    "    new_k = k.replace('language_model.model.base_model.','').replace('language_model.','').replace('.base_layer','')\n",
+    "    if not new_k in model_org.state_dict():\n",
+    "        print(k)\n",
+    "    else:\n",
+    "        new_state_dict[k] = model_org.state_dict()[new_k]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "<All keys matched successfully>"
+      ]
+     },
+     "execution_count": 6,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "model.load_state_dict(new_state_dict)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "/mnt/jeff/huggingface/modules/transformers_modules/gemma-3-4b-it-omni/speech_conformer_encoder.py:2775: FutureWarning: Please specify CheckpointImpl.NO_REENTRANT as CheckpointImpl.REENTRANT will soon be removed as the default and eventually deprecated.\n",
+      "  lambda i: encoder_checkpoint_wrapper(\n"
+     ]
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "######################## speech lora #############\n",
+      "######################## text lora #############\n"
+     ]
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "3e79659733a542c685c653b825c77cd6",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "Loading checkpoint shards:   0%|          | 0/3 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Some weights of Gemma3OmniForConditionalGeneration were not initialized from the model checkpoint at /mnt/jeff/InCar/Gemma3omni/gemma-3-4b-it-omni and are newly initialized: ['language_model.model.base_model.model.layers.0.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.0.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.0.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.0.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.0.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.0.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.0.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.0.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.0.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.0.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.0.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.0.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.0.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.0.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.1.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.1.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.1.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.1.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.1.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.1.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.1.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.1.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.1.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.1.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.1.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.1.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.1.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.1.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.10.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.10.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.10.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.10.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.10.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.10.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.10.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.10.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.10.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.10.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.10.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.10.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.10.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.10.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.11.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.11.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.11.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.11.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.11.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.11.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.11.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.11.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.11.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.11.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.11.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.11.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.11.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.11.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.12.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.12.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.12.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.12.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.12.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.12.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.12.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.12.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.12.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.12.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.12.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.12.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.12.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.12.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.13.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.13.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.13.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.13.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.13.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.13.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.13.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.13.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.13.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.13.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.13.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.13.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.13.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.13.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.14.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.14.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.14.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.14.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.14.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.14.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.14.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.14.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.14.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.14.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.14.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.14.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.14.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.14.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.15.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.15.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.15.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.15.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.15.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.15.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.15.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.15.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.15.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.15.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.15.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.15.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.15.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.15.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.16.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.16.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.16.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.16.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.16.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.16.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.16.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.16.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.16.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.16.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.16.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.16.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.16.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.16.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.17.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.17.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.17.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.17.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.17.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.17.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.17.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.17.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.17.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.17.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.17.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.17.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.17.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.17.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.18.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.18.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.18.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.18.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.18.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.18.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.18.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.18.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.18.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.18.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.18.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.18.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.18.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.18.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.19.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.19.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.19.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.19.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.19.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.19.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.19.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.19.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.19.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.19.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.19.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.19.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.19.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.19.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.2.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.2.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.2.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.2.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.2.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.2.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.2.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.2.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.2.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.2.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.2.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.2.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.2.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.2.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.20.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.20.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.20.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.20.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.20.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.20.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.20.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.20.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.20.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.20.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.20.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.20.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.20.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.20.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.21.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.21.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.21.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.21.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.21.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.21.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.21.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.21.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.21.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.21.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.21.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.21.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.21.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.21.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.22.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.22.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.22.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.22.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.22.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.22.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.22.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.22.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.22.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.22.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.22.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.22.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.22.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.22.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.23.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.23.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.23.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.23.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.23.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.23.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.23.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.23.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.23.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.23.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.23.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.23.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.23.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.23.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.24.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.24.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.24.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.24.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.24.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.24.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.24.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.24.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.24.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.24.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.24.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.24.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.24.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.24.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.25.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.25.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.25.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.25.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.25.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.25.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.25.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.25.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.25.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.25.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.25.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.25.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.25.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.25.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.26.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.26.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.26.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.26.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.26.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.26.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.26.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.26.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.26.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.26.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.26.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.26.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.26.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.26.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.27.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.27.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.27.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.27.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.27.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.27.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.27.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.27.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.27.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.27.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.27.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.27.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.27.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.27.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.28.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.28.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.28.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.28.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.28.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.28.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.28.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.28.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.28.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.28.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.28.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.28.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.28.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.28.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.29.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.29.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.29.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.29.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.29.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.29.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.29.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.29.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.29.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.29.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.29.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.29.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.29.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.29.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.3.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.3.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.3.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.3.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.3.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.3.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.3.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.3.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.3.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.3.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.3.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.3.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.3.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.3.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.30.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.30.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.30.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.30.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.30.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.30.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.30.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.30.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.30.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.30.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.30.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.30.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.30.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.30.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.31.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.31.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.31.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.31.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.31.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.31.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.31.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.31.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.31.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.31.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.31.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.31.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.31.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.31.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.32.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.32.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.32.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.32.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.32.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.32.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.32.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.32.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.32.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.32.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.32.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.32.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.32.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.32.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.33.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.33.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.33.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.33.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.33.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.33.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.33.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.33.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.33.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.33.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.33.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.33.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.33.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.33.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.4.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.4.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.4.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.4.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.4.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.4.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.4.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.4.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.4.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.4.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.4.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.4.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.4.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.4.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.5.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.5.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.5.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.5.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.5.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.5.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.5.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.5.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.5.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.5.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.5.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.5.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.5.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.5.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.6.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.6.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.6.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.6.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.6.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.6.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.6.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.6.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.6.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.6.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.6.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.6.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.6.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.6.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.7.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.7.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.7.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.7.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.7.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.7.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.7.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.7.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.7.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.7.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.7.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.7.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.7.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.7.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.8.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.8.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.8.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.8.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.8.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.8.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.8.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.8.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.8.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.8.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.8.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.8.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.8.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.8.self_attn.v_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.9.mlp.down_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.9.mlp.down_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.9.mlp.gate_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.9.mlp.gate_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.9.mlp.up_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.9.mlp.up_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.9.self_attn.k_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.9.self_attn.k_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.9.self_attn.o_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.9.self_attn.o_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.9.self_attn.q_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.9.self_attn.q_proj.lora_B.text.weight', 'language_model.model.base_model.model.layers.9.self_attn.v_proj.lora_A.text.weight', 'language_model.model.base_model.model.layers.9.self_attn.v_proj.lora_B.text.weight']\n",
+      "You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\n"
+     ]
+    }
+   ],
+   "source": [
+    "model_id = \"/mnt/jeff/InCar/Gemma3omni/gemma-3-4b-it-omni\"\n",
+    "revision = \"main\" #\"v1.0\"\n",
+    "\n",
+    "model2 = AutoModel.from_pretrained(\n",
+    "    model_id, device_map=\"cpu\", revision = revision, trust_remote_code=True\n",
+    ").eval()\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "model.audio_tower = model2.audio_tower"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 10,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "tmp = model.audio_projector.state_dict()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 11,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "0.weight tensor(0)\n",
+      "0.bias tensor(0)\n",
+      "2.weight tensor(0)\n",
+      "2.bias tensor(0)\n"
+     ]
+    }
+   ],
+   "source": [
+    "for k in tmp:\n",
+    "    print(k,torch.sum(torch.isnan(tmp[k])))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "tmp['0.bias'] = torch.zeros(tmp['0.bias'].shape)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "<All keys matched successfully>"
+      ]
+     },
+     "execution_count": 6,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "model.audio_projector.load_state_dict(tmp)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "######################## speech lora #############\n"
+     ]
+    }
+   ],
+   "source": [
+    "model.init_lora()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "model.save_pretrained('../init_nemotron_omni')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "base",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.12.7"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}

config.json

@@ -0,0 +1,98 @@
+{
+  "architectures": [
+    "NemotronOmniForConditionalGeneration"
+  ],
+  "initializer_range": 0.02,
+  
+  "text_config": {
+    "attention_bias": false,
+    "model_type": "llama_text",
+    "attention_dropout": 0.0,
+    "bos_token_id": 128000,
+    "eos_token_id": 128001,
+    "head_dim": 128,
+    "hidden_act": "silu",
+    "hidden_size": 3072,
+    "intermediate_size": 9216,
+    "max_position_embeddings": 131072,
+    "mlp_bias": false,
+    "num_attention_heads": 32,
+    "num_hidden_layers": 32,
+    "num_key_value_heads": 8,
+    "pretraining_tp": 1,
+    "rms_norm_eps": 1e-05,
+    "rope_scaling": {
+      "factor": 4.0,
+      "low_freq_factor": 1.0,
+      "high_freq_factor": 4.0,
+      "original_max_position_embeddings": 8192,
+      "rope_type": "llama3"
+    },
+    "rope_theta": 3565775107.2609234,
+    "tie_word_embeddings": false,
+    "torch_dtype": "bfloat16",
+    "transformers_version": "4.47.1",
+    "use_cache": true,
+    "vocab_size": 128256
+  },
+  
+  "audio_config": {
+    "activation": "swish",
+    "activation_checkpointing": {
+      "interval": 1,
+      "module": "transformer",
+      "offload": false
+    },
+    "attention_dim": 1024,
+    "attention_heads": 16,
+    "batch_norm": false,
+    "bias_in_glu": true,
+    "causal": true,
+    "chunk_size": -1,
+    "cnn_layer_norm": true,
+    "conv_activation": "swish",
+    "conv_glu_type": "swish",
+    "depthwise_multiplier": 1,
+    "depthwise_seperable_out_channel": 1024,
+    "dropout_rate": 0.0,
+    "encoder_embedding_config": {
+      "input_size": 80
+    },
+    "ext_pw_kernel_size": 1,
+    "ext_pw_out_channel": 1024,
+    "input_layer": "nemo_conv",
+    "input_size": 80,
+    "kernel_size": 3,
+    "left_chunk": 18,
+    "linear_units": 1536,
+    "nemo_conv_settings": {
+      "conv_channels": 1024
+    },
+    "num_blocks": 24,
+    "relative_attention_bias_args": {
+      "t5_bias_max_distance": 500,
+      "type": "t5"
+    },
+    "time_reduction": 8
+  },
+  "speech_lora": {
+    "dp": 0.01,
+    "layer": "((layers.*self_attn\\.(q|k|v|o)_proj)|(layers.*mlp\\.(gate|up|down)_proj))",
+    "lora_alpha": 320,
+    "r": 320,
+    "use_rslora": true
+  },
+  "auto_map": {
+    "AutoConfig": "configuration_nemotronOmni.NemotronOmniConfig",
+    "AutoModel": "modeling_NemotronOmni.NemotronOmniForConditionalGeneration"
+  },
+  "audio_token_index": 128255,
+  "boa_token_index": 128253,
+  "eoa_token_index": 128254,
+  "eos_token_id": [
+    128001,
+    128008,
+    128009
+  ],
+  "model_type": "NemotronOmni"
+}

configuration_nemotronOmni.py

@@ -0,0 +1,151 @@
+from typing import Optional
+
+from transformers import AutoConfig, LlamaConfig
+from transformers.configuration_utils import PretrainedConfig
+from transformers.modeling_rope_utils import rope_config_validation
+from transformers.utils import logging
+from transformers.models.siglip import SiglipVisionConfig
+
+
+logger = logging.get_logger(__name__)
+
+class AudioConfig(PretrainedConfig):
+    model_type = "nemotron_audio"
+    
+    def __init__(
+        self,
+        input_size=80,
+        attention_dim=1024,
+        attention_heads=16,
+        num_blocks=24,
+        linear_units=1536,
+        dropout_rate=0.0,
+        kernel_size=3,
+        ext_pw_kernel_size=1,
+        ext_pw_out_channel=1024,
+        depthwise_seperable_out_channel=1024,
+        depthwise_multiplier=1,
+        activation="swish",
+        conv_activation="swish",
+        conv_glu_type="swish",
+        bias_in_glu=True,
+        causal=True,
+        batch_norm=False,
+        cnn_layer_norm=True,
+        time_reduction=8,
+        input_layer="nemo_conv",
+        nemo_conv_settings=None,
+        chunk_size=-1,
+        left_chunk=18,
+        relative_attention_bias_args=None,
+        activation_checkpointing=None,
+        encoder_embedding_config=None,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+        
+        self.input_size = input_size
+        self.attention_dim = attention_dim
+        self.attention_heads = attention_heads
+        self.num_blocks = num_blocks
+        self.linear_units = linear_units
+        self.dropout_rate = dropout_rate
+        self.kernel_size = kernel_size
+        self.ext_pw_kernel_size = ext_pw_kernel_size
+        self.ext_pw_out_channel = ext_pw_out_channel
+        self.depthwise_seperable_out_channel = depthwise_seperable_out_channel
+        self.depthwise_multiplier = depthwise_multiplier
+        self.activation = activation
+        self.conv_activation = conv_activation
+        self.conv_glu_type = conv_glu_type
+        self.bias_in_glu = bias_in_glu
+        self.causal = causal
+        self.batch_norm = batch_norm
+        self.cnn_layer_norm = cnn_layer_norm
+        self.time_reduction = time_reduction
+        self.input_layer = input_layer
+        
+        if nemo_conv_settings is None:
+            self.nemo_conv_settings = {"conv_channels": 1024}
+        else:
+            self.nemo_conv_settings = nemo_conv_settings
+            
+        self.chunk_size = chunk_size
+        self.left_chunk = left_chunk
+        
+        if relative_attention_bias_args is None:
+            self.relative_attention_bias_args = {"type": "t5", "t5_bias_max_distance": 500}
+        else:
+            self.relative_attention_bias_args = relative_attention_bias_args
+            
+        if activation_checkpointing is None:
+            self.activation_checkpointing = {"interval": 1, "module": "transformer", "offload": False}
+        else:
+            self.activation_checkpointing = activation_checkpointing
+            
+        if encoder_embedding_config is None:
+            self.encoder_embedding_config = {"input_size": input_size}
+        else:
+            self.encoder_embedding_config = encoder_embedding_config
+    
+
+class NemotronOmniConfig(PretrainedConfig):
+
+    model_type = "NemotronOmni"
+    sub_configs = {
+        "text_config": LlamaConfig,
+        "audio_config": AudioConfig,
+    }
+
+    def __init__(
+        self,
+        text_config: Optional[LlamaConfig] = None,
+        vision_config: Optional[SiglipVisionConfig] = None,
+        audio_config: Optional[AudioConfig] = None,
+        # mm_tokens_per_image: int = 256,
+        # boi_token_index: int = 255_999,
+        # eoi_token_index: int = 256_000,
+        boa_token_index: int = 128253,
+        eoa_token_index: int = 128254,
+        # image_token_index: int = 262_144,
+        audio_token_index: int = 128255,
+        initializer_range: float = 0.02,
+        **kwargs,
+    ):
+        if text_config is None:
+            text_config = LlamaConfig()
+            logger.info("text_config is None, using default LlamaConfig vision config.")
+        elif isinstance(text_config, dict):
+            text_config = LlamaConfig(**text_config)
+
+        # if isinstance(vision_config, dict):
+        #     vision_config = SiglipVisionConfig(**vision_config)
+        # else:
+        #     vision_config = SiglipVisionConfig()
+        #     logger.info(
+        #         "vision_config is None or incompatible with Gemma3VisionConfig intialization. Gemma3 will be limited "
+        #         "to text tasks."
+        #     )
+            
+        if isinstance(audio_config, dict):
+            audio_config = AudioConfig(**audio_config)
+        else:
+            audio_config = AudioConfig()
+            logger.info(
+                "audio_config is None or incompatible with Gemma3AudioConfig intialization. Gemma3 will be limited "
+                "to text tasks."
+            )
+
+        self.text_config = text_config
+        # self.vision_config = vision_config
+        self.audio_config = audio_config
+        # self.mm_tokens_per_image = mm_tokens_per_image
+        # self.boi_token_index = boi_token_index
+        # self.eoi_token_index = eoi_token_index
+        self.boa_token_index = boa_token_index
+        self.eoa_token_index = eoa_token_index
+        # self.image_token_index = image_token_index
+        self.audio_token_index = audio_token_index
+        self.initializer_range = initializer_range
+
+        super().__init__(**kwargs)

eval.py

@@ -0,0 +1,311 @@
+from io import BytesIO
+from urllib.request import urlopen
+import soundfile
+import torch
+from datasets import load_dataset, Audio
+import numpy as np
+from transformers import AutoModel, AutoProcessor, BatchFeature
+from tqdm import tqdm
+import json
+import os
+import time
+from datetime import datetime
+from whisper_normalizer.english import EnglishTextNormalizer
+from whisper_normalizer.basic import BasicTextNormalizer
+import sacrebleu
+from jiwer import cer, wer
+from torch.utils.data import Dataset, DataLoader
+import soundfile as sf
+import re
+from pathlib import Path
+import opencc
+from ASRDataset import *
+converter = opencc.OpenCC('s2tw.json')
+normalizer = {
+    "en_us" : EnglishTextNormalizer(),
+    "other" : BasicTextNormalizer()
+}
+
+model_id = "/mnt/jeff/InCar/LlamaNemotronOmni/test_nemotron_omni"
+revision = "main" #"v1.0"
+
+model = AutoModel.from_pretrained(
+    model_id, device_map="cuda", revision = revision, trust_remote_code=True
+).eval()
+
+processor = AutoProcessor.from_pretrained(
+    model_id, revision = revision, trust_remote_code=True
+)
+if 'LlamaNemotronOmni' in model_id:
+    processor.tokenizer.pad_token_id = processor.tokenizer.eos_token_id
+
+results_dir = f"evaluation_results_{datetime.now().strftime('%Y%m%d_%H%M%S')}"
+os.makedirs(results_dir, exist_ok=True)
+
+
+INSTRUCTION = {
+    "ast": "Translate the audio to {0}.",
+    "asr": "Transcribe the audio clip into text.",
+}
+
+def covost_collate_fn_test(batch):
+    input_ids_list = []
+    input_audio_embeds_list = []
+    audio_embed_sizes_list = []
+    audio_attention_mask_list = []
+    input_modes_list = []
+    answer_list = []
+    for inputs in batch:
+        input_ids_list.append(inputs['input_ids'][0])
+        input_audio_embeds_list.append(inputs['input_audio_embeds'])
+        audio_embed_sizes_list.append(inputs['audio_embed_sizes'])
+        audio_attention_mask_list.append(
+            inputs['input_audio_embeds'].new_full((inputs['input_audio_embeds'].size(1),), True, dtype=torch.bool)
+        )
+        input_modes_list.append(inputs['input_modes'])
+        answer_list.append(inputs['answer'])
+
+    try:
+        input_ids = pad_sequence(input_ids_list, padding_side='left', padding_value=0)
+        audio_attention_mask = (
+            pad_sequence(audio_attention_mask_list, padding_side='right', padding_value=False)
+            if len(audio_attention_mask_list) > 1
+            else None
+        )
+    except Exception as e:
+        print(e)
+        print(input_ids_list)
+        print(audio_attention_mask)
+        raise
+    attention_mask = (input_ids != 0).long()
+    input_audio_embeds = cat_with_pad(input_audio_embeds_list, dim=0)
+    audio_embed_sizes = torch.cat(audio_embed_sizes_list)
+    input_modes = torch.cat(input_modes_list)
+
+    return BatchFeature(
+        {
+            'input_ids': input_ids,
+            'attention_mask': attention_mask,
+            'input_audio_embeds': input_audio_embeds,
+            'audio_embed_sizes': audio_embed_sizes,
+            'audio_attention_mask': audio_attention_mask,
+            'input_modes': input_modes,
+            'answer': answer_list,
+        }
+    )
+
+def save_results(results, dataset_name, task, source_lang, target_lang=None, sample_idx=None):
+    filename = f"{task}_{dataset_name}_{source_lang}"
+    if target_lang:
+        filename += f"_to_{target_lang}"
+    if sample_idx is not None:
+        filename += f"_sample_{sample_idx}"
+    
+    filepath = os.path.join(results_dir, f"{filename}.json")
+    
+    results["timestamp"] = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+    
+    with open(filepath, 'w', encoding='utf-8') as f:
+        json.dump(results, f, ensure_ascii=False, indent=2)
+    
+    return filepath
+
+def evaluate_task(dataset, source_lang, target_lang, num_samples=-1, batch_size = 4, is_asr=True):
+    task_type = "asr" if is_asr else "translation"
+    eval_lang = source_lang if is_asr else target_lang
+    if eval_lang in normalizer:
+        eval_normalizer = normalizer[eval_lang]
+    else:
+        eval_normalizer = normalizer['other']
+    sample_results = []
+    
+    if num_samples > 0 and num_samples < len(dataset):
+        indices = np.random.choice(len(dataset), num_samples, replace=False)
+        dataset = dataset.select(indices)
+    
+    dataloader = DataLoader(dataset, batch_size=batch_size, shuffle=False, collate_fn=covost_collate_fn_test)
+        
+    evaluated_samples = {} 
+    
+    for batch_idx, batch in enumerate(tqdm(dataloader)):
+        batch_references = batch.pop("answer")
+
+        if torch.cuda.is_available():
+            try:
+                batch = {k: v.to("cuda") for k, v in batch.items()}
+            except:
+                print('error')
+                break
+        
+        with torch.inference_mode():
+            generate_ids = model.generate(**batch, 
+            max_new_tokens=256,
+            #temperature = 1.0, top_p = 0.95, top_k = 64, do_sample=True
+            )
+            
+            input_lengths = batch['input_ids'].shape[1]
+            generate_ids = generate_ids[:, input_lengths:]
+        
+            batch_predictions = processor.batch_decode(
+                generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
+            )
+        
+        for i, (reference, prediction) in enumerate(zip(batch_references, batch_predictions)):
+            idx = batch_idx * batch_size + i
+            sample_result = {
+                "id": idx,
+                "reference": reference,
+                "prediction": converter.convert(prediction)
+            }
+            sample_results.append(sample_result)
+
+        if (batch_idx + 1) % 10 == 0:
+            temp_results = []
+            
+            for item in sample_results:
+                sample_id = item["id"]
+                
+                if sample_id in evaluated_samples:
+                    temp_item = item.copy()
+                    temp_item.update(evaluated_samples[sample_id])
+                    temp_results.append(temp_item)
+                else:
+                    temp_item = item.copy()
+                    try:
+                        ref = eval_normalizer(item["reference"])
+                        pred = eval_normalizer(item["prediction"])
+
+                        # BLEU, WER/CER
+                        utt_bleu = sacrebleu.sentence_bleu(pred, [ref]).score
+                        utt_cer = round(cer(re.sub(r"\s+", "", ref), re.sub(r"\s+", "", pred)) * 100, 2)
+                        utt_wer = round(wer(ref, pred) * 100, 2)
+
+                        metrics = {
+                            "bleu": utt_bleu,
+                            "cer": min(100,utt_cer),
+                            "wer": utt_wer
+                        }
+                        
+                        evaluated_samples[sample_id] = metrics
+                        temp_item.update(metrics)
+                    except Exception as e:
+                        print(f"Error evaluating sample {sample_id}: {e}")
+                        metrics = {
+                            "bleu": 0,
+                            "cer": 100,
+                            "wer": 100,
+                            "error": str(e)
+                        }
+                        evaluated_samples[sample_id] = metrics
+                        temp_item.update(metrics)
+                
+                    temp_results.append(temp_item)
+
+            partial_results = {
+                "task": task_type,
+                "source_lang": source_lang,
+                "target_lang": target_lang,
+                "num_samples": len(temp_results),
+                "sample_results": temp_results
+            }
+            save_results(partial_results, dataset.name, task_type, source_lang, target_lang)
+
+    for item in sample_results:
+        ref = eval_normalizer(item["reference"])
+        pred = eval_normalizer(item["prediction"])
+
+        utt_bleu = sacrebleu.sentence_bleu(pred, [ref]).score
+        utt_cer = round(cer(re.sub(r"\s+", "", ref), re.sub(r"\s+", "", pred)) * 100, 2)
+        utt_wer = round(wer(ref, pred) * 100, 2)
+
+        item.update({
+            "bleu": utt_bleu,
+            "cer": min(100,utt_cer),
+            "wer": utt_wer
+        })
+
+    avg_bleu = sum(item["bleu"] for item in sample_results) / len(sample_results)
+    avg_cer = sum(item["cer"] for item in sample_results) / len(sample_results)
+    avg_wer = sum(item["wer"] for item in sample_results) / len(sample_results)
+    
+    results = {
+        "dataset": dataset.name,
+        "task": task_type,
+        "source_lang": source_lang,
+        "target_lang": target_lang,
+        "num_samples": len(sample_results),
+        "metrics": {
+            "bleu": avg_bleu,
+            "cer": avg_cer,
+            "wer": avg_wer
+        },
+        "sample_results": sample_results
+    }
+    
+    save_results(results, dataset.name, task_type, source_lang, target_lang)
+    return results
+
+
+if __name__ == "__main__":
+
+    source_languages = [
+        ("en_us", "English"),
+    ]
+    
+    target_languages = [
+        ("zh-TW", "zh-TW"),
+    ]
+
+    num_samples = -1
+    batch_size = 32
+
+    for source_lang, target_lang in zip(source_languages, target_languages):
+        print(f"\n===== {source_lang[0]} ASR =====")
+        
+        split = "test"
+
+        datasets = []
+
+        
+        
+        commonvoice_speech_tw = CommonVoiceDataset(
+            processor=processor,
+            source_lang="zh-TW",
+            split=split
+        )
+        datasets.append(commonvoice_speech_tw)
+        fleurs = FleursDataset(
+            processor=processor,
+            split=split,
+            source_lang="en_us",  # English
+            mode="asr"
+        )
+        datasets.append(fleurs)
+
+        # Libri Speech Clean ASR mode (English -> English text)
+        # libri_speech_clean = LibriSpeechDataset(
+        #     processor=processor,
+        #     subset="clean",
+        #     split=split
+        # )
+        # datasets.append(libri_speech_clean)
+
+        # # Libri Speech Other ASR mode (English -> English text)
+        # libri_speech_other = LibriSpeechDataset(
+        #     processor=processor,
+        #     subset="other",
+        #     split=split
+        # )
+        # datasets.append(libri_speech_other)
+
+        # Fleurs ASR mode (English -> English text)
+        
+
+        for dataset in datasets:
+            # ASR
+            asr_results = evaluate_task(dataset, source_lang[0], target_lang[0], num_samples, batch_size=batch_size, is_asr = True)
+
+            print(f"\n=== {asr_results.get('dataset', 'Dataset')} | {source_lang[0]} ASR===")
+            print(f"BLEU: {asr_results.get('metrics', {}).get('bleu', 'N/A')}")
+            print(f"WER: {asr_results.get('metrics', {}).get('wer', 'N/A')}")
+            print(f"CER: {asr_results.get('metrics', {}).get('cer', 'N/A')}")

generation_config.json

@@ -0,0 +1,10 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 128000,
+  "eos_token_id": [
+    128001,
+    128008,
+    128009
+  ],
+  "transformers_version": "4.47.1"
+}

image_processing.py

@@ -0,0 +1,112 @@
+from typing import List, Optional, Union
+
+from PIL import Image
+import torch
+from transformers.image_processing_base import BatchFeature
+from transformers.image_processing_utils_fast import (BaseImageProcessorFast,
+                                                      divide_to_patches)
+from transformers.image_utils import (ChannelDimension, SizeDict,
+                                      get_image_size, make_list_of_images,
+                                      get_image_type, ImageInput, ImageType)
+from transformers.utils import TensorType
+
+
+def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
+    best_factor = float('-inf')
+    best_ratio = (1, 1)
+    area = width * height
+    for ratio in target_ratios:
+        target_aspect_ratio = ratio[0] / ratio[1]
+        factor_based_on_area_n_ratio = min(
+            (ratio[0]*ratio[1]*image_size*image_size)/ area, 0.6
+            )* min(
+                target_aspect_ratio/aspect_ratio, aspect_ratio/target_aspect_ratio)
+        if factor_based_on_area_n_ratio > best_factor:
+            best_factor = factor_based_on_area_n_ratio
+            best_ratio = ratio
+    return best_ratio
+
+
+class LlamaNemotronNanoVLImageProcessor(BaseImageProcessorFast):
+    model_input_names = ["pixel_values"]
+
+    def __init__(self, image_size=512, max_num_tiles=12, use_thumbnail=True, **kwargs):
+        super().__init__(**kwargs)
+        self.image_size = image_size
+        self.max_num_tiles = max_num_tiles
+        self.use_thumbnail = use_thumbnail
+
+    # Based on https://github.com/OpenGVLab/InternVL/blob/c62fa4f7c850165d7386bdc48ac6bc5a6fab0864/internvl_chat/internvl/train/dataset.py#L702
+    def dynamic_preprocess(self, image, image_size=448, max_num_tiles=12, use_thumbnail=False):
+        orig_height, orig_width = get_image_size(image, channel_dim=ChannelDimension.FIRST)
+        aspect_ratio = orig_width / orig_height
+
+        # calculate the existing image aspect ratio
+        target_ratios = set(
+            (i, j) for n in range(1, max_num_tiles + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
+            i * j <= max_num_tiles and i * j >= 1)
+        target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
+
+        # find the closest aspect ratio to the target
+        target_aspect_ratio = find_closest_aspect_ratio(
+            aspect_ratio, target_ratios, orig_width, orig_height, image_size)
+
+        # calculate the target width and height
+        target_width = image_size * target_aspect_ratio[0]
+        target_height = image_size * target_aspect_ratio[1]
+
+        resized_img = self.resize(image, SizeDict(height=target_height, width=target_width))
+        patches = divide_to_patches(resized_img, image_size)
+        if use_thumbnail and len(patches) != 1:
+            patches.append(self.resize(image, SizeDict(height=image_size, width=image_size)))
+
+        return patches
+
+    def _process_image(
+        self,
+        image: ImageInput,
+        **kwargs,
+    ) -> torch.Tensor:
+        image_type = get_image_type(image)
+        if image_type not in [ImageType.PIL]:
+            raise ValueError(f"Unsupported input image type {image_type}. Only PIL images supported")
+        image = image.resize((image.width * 2, image.height * 2), Image.BILINEAR)
+        return super()._process_image(image, **kwargs)
+
+    def _preprocess(
+        self,
+        images: List[torch.Tensor],
+        image_size: int = None,
+        max_num_tiles: int = None,
+        use_thumbnail: bool = None,
+        do_rescale: bool = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        **kwargs,
+    ) -> List[torch.Tensor]:
+        image_size = image_size if image_size is not None else self.image_size
+        max_num_tiles = max_num_tiles if max_num_tiles is not None else self.max_num_tiles
+        use_thumbnail = use_thumbnail if use_thumbnail is not None else self.use_thumbnail
+        do_rescale = do_rescale if do_rescale is not None else self.do_rescale
+
+        images = make_list_of_images(images)
+
+        all_patches = []
+        num_patches = []
+        for image in images:
+            patches = self.dynamic_preprocess(
+                image, image_size, max_num_tiles, use_thumbnail
+            )
+            all_patches.extend(patches)
+            num_patches.append(len(patches))
+
+        pixel_values = torch.stack(all_patches, dim=0)
+        pixel_values = self.rescale_and_normalize(
+            pixel_values,
+            do_rescale,
+            self.rescale_factor,
+            do_normalize=self.do_normalize,
+            image_mean=self.image_mean,
+            image_std=self.image_std
+        )
+
+        return BatchFeature(data={"pixel_values": pixel_values, "num_patches": num_patches}, tensor_type=return_tensors)

model-00001-of-00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a92b30f1db46c461b3cc8925efd5d4dc6aec0d09f8fee187401e883bf6f0ea9
+size 4993077672

model-00002-of-00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:71bdd0cd6cae461ba7480a092286f1ccd0ac633f8fd54df00bf3951945722886
+size 4968834432

model-00003-of-00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c752aa51ccfb3fa23c533e8e90abe1ae910e0f1ff96143f00e7618414af1ab8
+size 1187810344

modeling_NemotronOmni.py

@@ -0,0 +1,447 @@
+import copy
+from collections.abc import Callable
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from transformers.cache_utils import Cache, HybridCache, StaticCache
+from transformers.generation import GenerationMixin
+from transformers.utils.deprecation import deprecate_kwarg
+from transformers import AutoModel, AutoModelForCausalLM
+
+from transformers.models.nemotron.modeling_nemotron import NemotronPreTrainedModel
+from transformers.modeling_utils import PreTrainedModel
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers import AutoConfig, AutoModelForCausalLM
+from transformers.utils import (
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_torchdynamo_compiling,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_nemotronOmni import NemotronOmniConfig
+from .speech_conformer_encoder import ConformerEncoder
+from enum import Enum
+class InputMode(Enum):
+    LANGUAGE = 0
+    VISION = 1
+    SPEECH = 2
+    VISION_SPEECH = 3
+logger = logging.get_logger(__name__)
+_CONFIG_FOR_DOC = "NemotronOmniConfig"
+
+class NemotronOmniForConditionalGeneration(PreTrainedModel, GenerationMixin):
+    config_class = NemotronOmniConfig
+    use_gradient_checkpointing=False
+    def __init__(self, config: NemotronOmniConfig):
+        super().__init__(config)
+        audio_config = config.audio_config.to_diff_dict()
+        for item in ['transformers_version', 'model_type', 'torch_dtype']:
+            if item in audio_config:
+                audio_config.pop(item)
+        self.audio_tower = ConformerEncoder(**audio_config)
+        self.audio_tower.post_init({})
+        self.audio_tower = self.audio_tower.to(dtype=self.dtype)
+        self.audio_projector = nn.Sequential(
+            nn.Linear(in_features=config.audio_config.attention_dim, out_features=config.text_config.hidden_size, bias=True),
+            nn.GELU(approximate='none'),
+            nn.Linear(in_features=config.text_config.hidden_size, out_features=config.text_config.hidden_size, bias=True)
+        ).to(dtype=self.dtype)
+
+        self.vocab_size = config.text_config.vocab_size
+
+        language_model = AutoModelForCausalLM.from_config(config=config.text_config)
+
+        if language_model._tied_weights_keys is not None:
+            self._tied_weights_keys = [f"language_model.{k}" for k in language_model._tied_weights_keys]
+        self.language_model = language_model
+
+        self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1
+        self.init_lora()
+        self.post_init()
+        
+        
+    def init_lora(self):
+        from peft import LoraConfig, get_peft_model
+        import warnings
+        print('######################## speech lora #############')
+        speech_lora_config = LoraConfig(
+            r=self.config.speech_lora['r'],
+            lora_alpha=self.config.speech_lora['lora_alpha'],
+            target_modules=self.config.speech_lora['layer'],
+            use_rslora=self.config.speech_lora['use_rslora'],
+            lora_dropout=self.config.speech_lora['dp'],
+            task_type="CAUSAL_LM",
+        )
+        self.language_model.model = get_peft_model(self.language_model.model, speech_lora_config, adapter_name="speech")
+        # print('######################## text lora #############')
+        # text_lora_config = LoraConfig(
+        #     r=self.config.text_lora['r'],
+        #     lora_alpha=self.config.text_lora['lora_alpha'],
+        #     target_modules=self.config.text_lora['layer'],
+        #     use_rslora=self.config.text_lora['use_rslora'],
+        #     lora_dropout=self.config.text_lora['dp'],
+        #     task_type="CAUSAL_LM",
+        # )
+        # self.language_model.model.base_model.active_adapter.append("text")
+        # self.language_model.model.add_adapter("text", text_lora_config)
+        
+    def set_lora_adapter(self, adapter_name) -> None:
+        from peft.tuners.lora.layer import LoraLayer
+        for module in self.modules():
+            if isinstance(module, LoraLayer):
+                if module.merged:
+                    warnings.warn("Adapter cannot be set when the model is merged. Unmerging the model first.")
+                    module.unmerge()
+                module.set_adapter(adapter_name)
+                module._disable_adapters = False
+
+    def unset_lora_adapter(self) -> None:
+        # Ref: peft/tuners/tuners_utils.py - enable_adapters()
+        # Ref: peft/tuners/lora/layer.py
+        from peft.tuners.lora.layer import LoraLayer
+        for module in self.modules():
+            if isinstance(module, LoraLayer):
+                # disable grads on all adapter layers
+                # TODO weijian: may use enable_adapters() instead
+                for layer_name in module.adapter_layer_names:
+                    layer = getattr(module, layer_name)
+                    layer.requires_grad_(False)
+                module._disable_adapters = True
+    def get_input_embeddings(self):
+        return self.language_model.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.language_model.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.language_model.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.language_model.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.language_model.model = decoder
+
+    def get_decoder(self):
+        return self.language_model.model
+
+    def _update_causal_mask(
+        self,
+        attention_mask,
+        token_type_ids,
+        past_key_values,
+        cache_position,
+        input_tensor,
+        is_training: bool = False,
+    ):
+        if self.config.text_config._attn_implementation == "flash_attention_2":
+            return attention_mask
+
+        if attention_mask is not None and attention_mask.dim() == 4:
+            # In this case we assume that the mask comes already in inverted
+            # form and requires no inversion or slicing.
+            return attention_mask
+
+        using_static_cache = isinstance(past_key_values, StaticCache)
+        min_dtype = torch.finfo(self.dtype).min
+        inputs_lead_dim, sequence_length = input_tensor.shape[:2]
+        if using_static_cache:
+            target_length = past_key_values.get_max_cache_shape()
+        elif isinstance(past_key_values, HybridCache):
+            target_length = past_key_values.get_max_cache_shape()
+        else:
+            target_length = (
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else cache_position[0] + sequence_length + 1
+            )
+
+        if attention_mask is not None and attention_mask.dim() == 4:
+            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+            return attention_mask
+
+        causal_mask = torch.full(
+            (sequence_length, target_length), fill_value=min_dtype, dtype=self.dtype, device=cache_position.device
+        )
+
+        # Causal diagonal mask only if training, otherwise attend to the whole prefix. Training-specific attn for prefix is handled below
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+
+        causal_mask *= torch.arange(target_length, device=cache_position.device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(inputs_lead_dim, 1, -1, -1)
+
+        # Apply bidirectional mask on images if token type ids are provided
+        if token_type_ids is not None and sequence_length != 1:
+            token_type_mask = token_type_ids.unsqueeze(1) == token_type_ids.unsqueeze(2)
+            token_type_mask[token_type_ids == 0] = False  # if text token do not change anything
+            token_type_mask = token_type_mask.unsqueeze(1).to(causal_mask.device, dtype=torch.bool)
+            causal_mask = causal_mask.clone()
+            causal_mask[:, :, :, :sequence_length] = causal_mask[:, :, :, :sequence_length].masked_fill(
+                token_type_mask, 0.0
+            )
+
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+
+            # Then apply padding mask (will mask pad tokens)
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(causal_mask.device)
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+        return causal_mask
+        
+    def get_audio_features(self, input_audio_embeds: torch.FloatTensor, audio_attention_mask: torch.FloatTensor, audio_embed_sizes: torch.FloatTensor):
+        """
+        Projects the last hidden state from the audio model into language model space.
+        
+        Args:
+            audio_inputs (`torch.FloatTensor]` of shape `(batch_size, sequence_length, feature_dim)`)
+                The tensors corresponding to the input audio features.
+                
+        Returns:
+            audio_features (`torch.Tensor`): Audio feature tensor of shape `(batch_size, audio_length, embed_dim)`).
+        """
+        
+        audio_features, masks = self.audio_tower(input_audio_embeds, audio_attention_mask)
+        audio_outputs = self.audio_projector(audio_features)
+        return audio_outputs
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        input_audio_embeds: torch.FloatTensor = None,
+        audio_embed_sizes: torch.FloatTensor = None,
+        audio_attention_mask: torch.FloatTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        input_modes: torch.LongTensor = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[List[torch.FloatTensor], Cache]] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **lm_kwargs,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if isinstance(input_modes, torch.Tensor):
+            # len(input_mode) == num_beams in beam search, and all elements of input_mode should have the same value
+            input_modes = input_modes.unique()
+            if len(input_modes) != 1:
+                raise ValueError("Elements of input_modes should have the same value")
+
+        input_mode = InputMode(input_modes.item())
+
+        if input_mode in [InputMode.VISION_SPEECH, InputMode.VISION]:
+            self.unset_lora_adapter()
+            #self.set_lora_adapter('vision')
+            #audio_projection_mode = 'vision'
+        elif input_mode == InputMode.SPEECH:
+            self.unset_lora_adapter()
+            self.set_lora_adapter('speech')
+            #audio_projection_mode = 'speech'
+        elif input_mode == InputMode.LANGUAGE:
+            self.unset_lora_adapter()
+            # self.set_lora_adapter('text')
+        else:
+            raise ValueError(f"Invalid input_mode: {input_mode}")
+
+        is_training = token_type_ids is not None and labels is not None
+
+        # Replace image id woth PAD if the image token if OOV, to avoid index-errors
+        if input_ids is not None or self.config.audio_token_index >= self.vocab_size:
+            special_audio_mask = input_ids == self.config.audio_token_index
+            llm_input_ids = input_ids.clone()
+            llm_input_ids[special_audio_mask] = 0
+        else:
+            llm_input_ids = input_ids
+
+        if inputs_embeds is None:
+            inputs_embeds = self.language_model.model.embed_tokens(input_ids)
+        inputs_embeds = inputs_embeds.to(dtype=self.dtype)
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0) + 1  # Gemma3 positions are 1-indexed
+        # Merge text and audios
+        if input_audio_embeds is not None:
+            input_audio_embeds=input_audio_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
+            if audio_attention_mask is not None:
+                audio_attention_mask=audio_attention_mask.to(inputs_embeds.device, inputs_embeds.dtype)
+            audio_features = self.get_audio_features(input_audio_embeds, audio_attention_mask, audio_embed_sizes)
+            if input_ids is None:
+                special_audio_mask = inputs_embeds == self.get_input_embeddings()(
+                    torch.tensor(self.config.audio_token_index, dtype=torch.long, device=inputs_embeds.device)
+                )
+            else:
+                special_audio_mask = (input_ids == self.config.audio_token_index).unsqueeze(-1)
+                special_audio_mask = special_audio_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
+            masked_audio_features = []
+            for i, size in enumerate(audio_embed_sizes):
+                masked_audio_features.append(audio_features[i, :size, :])
+            masked_audio_features = torch.cat(masked_audio_features, dim=0)
+            if not is_torchdynamo_compiling() and inputs_embeds[special_audio_mask].numel() != masked_audio_features.numel():
+                audio_tokens_in_text = (special_audio_mask).sum(dim=1).sum(dim=0)[0]
+                masked_audio_size = audio_embed_sizes#.sum()[0]
+                raise ValueError(
+                    f"Number of audio does not match number of special audio tokens in the input text. "
+                    f"Got {audio_tokens_in_text} audio tokens in the text but {masked_audio_size} "
+                    "tokens from audio embeddings. "
+                    f"{masked_audio_features.numel()} \n"
+                    f"{inputs_embeds[special_audio_mask].numel()} \n"
+                    f"{audio_features} \n"
+                    f"{inputs_embeds[special_audio_mask]} \n"
+                    f"{special_audio_mask} \n"
+                )
+            masked_audio_features = masked_audio_features.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(special_audio_mask, masked_audio_features)
+        # mask out pad-token-ids in labels for BC
+        if labels is not None and self.pad_token_id in labels:
+            logger.warning_once(
+                "`labels` contains `pad_token_id` which will be masked with `config.ignore_index`. "
+                "You have to mask out `pad_token_id` when preparing `labels`, this behavior will be removed in v.4.46.",
+            )
+            labels = torch.where(input_ids == self.pad_token_id, self.config.ignore_index, labels)
+
+        causal_mask = self._update_causal_mask(
+            attention_mask, token_type_ids, past_key_values, cache_position, inputs_embeds, is_training
+        )
+        
+        outputs = self.language_model(
+            attention_mask=causal_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+            logits_to_keep=logits_to_keep,
+            **lm_kwargs,
+        )
+        logits = outputs.logits
+        loss = None
+        # print('#############################')
+        # print(logits)
+        if labels is not None:
+            # Upcast to float if we need to compute the loss to avoid potential precision issues
+            logits = logits.float()
+            shift_logits = logits[..., :-1, :]
+            shift_labels = labels[..., 1:]
+            if attention_mask is not None:
+                # we use the input attention mask to shift the logits and labels, because it is 2D.
+                # we also crop attn mask in case it is longer, which happens in PrefixTuning with peft
+                shift_attention_mask = attention_mask[:, -shift_logits.shape[1] :].to(logits.device)
+                shift_logits = shift_logits[shift_attention_mask.to(logits.device) != 0].contiguous()
+                shift_labels = shift_labels[shift_attention_mask.to(shift_labels.device) != 0].contiguous()
+            else:
+                shift_logits = shift_logits.contiguous()
+                shift_labels = shift_labels.contiguous()
+            # Flatten the tokens
+            loss_fct = nn.CrossEntropyLoss()
+
+            flat_logits = shift_logits.view(-1, self.config.text_config.vocab_size)
+            flat_labels = shift_labels.view(-1).to(shift_logits.device)
+            loss = loss_fct(flat_logits, flat_labels)
+            # print('flat logits',flat_logits)
+            # print(flat_labels)
+            # print(loss)
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            # image_hidden_states=image_features if pixel_values is not None else None,
+            # audio_hidden_states=audio_features if input_audio_embeds is not None else None,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        input_modes=None,
+        inputs_embeds=None,
+        cache_position=None,
+        position_ids=None,
+        pixel_values=None,
+        input_audio_embeds=None,
+        audio_embed_sizes=None,
+        audio_attention_mask=None,
+        attention_mask=None,
+        token_type_ids=None,
+        use_cache=True,
+        logits_to_keep=None,
+        labels=None,
+        **kwargs,
+    ):
+        # Overwritten -- custom `position_ids` and `pixel_values` handling
+        model_inputs = self.language_model.prepare_inputs_for_generation(
+            input_ids,
+            past_key_values=past_key_values,
+            input_modes=input_modes,
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            cache_position=cache_position,
+            use_cache=use_cache,
+            logits_to_keep=logits_to_keep,
+            token_type_ids=token_type_ids,
+            **kwargs,
+        )
+
+        # position_ids in Gemma3 are 1-indexed
+        if model_inputs.get("position_ids") is not None:
+            model_inputs["position_ids"] += 1
+        # If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore
+        # Otherwise we need pixel values to be passed to model. NOTE: use_cache=False needs pixel_values always
+        if cache_position[0] == 0:
+            model_inputs["pixel_values"] = pixel_values
+            model_inputs["input_audio_embeds"] = input_audio_embeds
+            model_inputs["audio_embed_sizes"] = audio_embed_sizes
+            model_inputs["audio_attention_mask"] = audio_attention_mask
+        model_inputs["input_modes"] = input_modes
+        is_training = token_type_ids is not None and labels is not None
+        if cache_position[0] == 0 and isinstance(past_key_values, HybridCache):
+            input_tensor = inputs_embeds if inputs_embeds is not None else input_ids
+            causal_mask = self._update_causal_mask(
+                attention_mask, token_type_ids, past_key_values, cache_position, input_tensor, is_training
+            )
+            model_inputs["attention_mask"] = causal_mask
+
+        return model_inputs
+
+    def tie_weights(self):
+        return self.language_model.tie_weights()
+

preprocessing_NemotronOmni.py

@@ -0,0 +1,369 @@
+import re
+from typing import List, Optional, Union, Tuple
+from math import ceil
+
+import numpy as np
+import torch
+import scipy
+from torch.nn.utils.rnn import pad_sequence
+
+from enum import Enum
+
+from transformers import AutoFeatureExtractor
+from transformers.feature_extraction_utils import BatchFeature
+from transformers.feature_extraction_sequence_utils import SequenceFeatureExtractor
+from transformers.image_utils import ImageInput, make_nested_list_of_images
+from transformers.processing_utils import ImagesKwargs, ProcessingKwargs, ProcessorMixin, Unpack, AudioKwargs
+from transformers.tokenization_utils_base import PreTokenizedInput, TextInput
+from transformers.utils import to_py_obj, TensorType
+from transformers.audio_utils import AudioInput
+
+
+
+
+def speechlib_mel(sample_rate, n_fft, n_mels, fmin=None, fmax=None):
+    """Create a Mel filter-bank the same as SpeechLib FbankFC.
+
+    Args:
+        sample_rate (int): Sample rate in Hz. number > 0 [scalar]
+        n_fft (int): FFT size. int > 0 [scalar]
+        n_mel (int): Mel filter size. int > 0 [scalar]
+        fmin (float): lowest frequency (in Hz). If None use 0.0.
+            float >= 0 [scalar]
+        fmax: highest frequency (in Hz). If None use sample_rate / 2.
+            float >= 0 [scalar]
+
+    Returns
+        out (numpy.ndarray): Mel transform matrix
+            [shape=(n_mels, 1 + n_fft/2)]
+    """
+
+    bank_width = int(n_fft // 2 + 1)
+    if fmax is None:
+        fmax = sample_rate / 2
+    if fmin is None:
+        fmin = 0
+    assert fmin >= 0, "fmin cannot be negtive"
+    assert fmin < fmax <= sample_rate / 2, "fmax must be between (fmin, samplerate / 2]"
+
+    def mel(f):
+        return 1127.0 * np.log(1.0 + f / 700.0)
+
+    def bin2mel(fft_bin):
+        return 1127.0 * np.log(1.0 + fft_bin * sample_rate / (n_fft * 700.0))
+
+    def f2bin(f):
+        return int((f * n_fft / sample_rate) + 0.5)
+
+    # Spec 1: FFT bin range [f2bin(fmin) + 1, f2bin(fmax) - 1]
+    klo = f2bin(fmin) + 1
+    khi = f2bin(fmax)
+
+    khi = max(khi, klo)
+
+    # Spec 2: SpeechLib uses trianges in Mel space
+    mlo = mel(fmin)
+    mhi = mel(fmax)
+    m_centers = np.linspace(mlo, mhi, n_mels + 2)
+    ms = (mhi - mlo) / (n_mels + 1)
+
+    matrix = np.zeros((n_mels, bank_width), dtype=np.float32)
+    for m in range(0, n_mels):
+        left = m_centers[m]
+        center = m_centers[m + 1]
+        right = m_centers[m + 2]
+        for fft_bin in range(klo, khi):
+            mbin = bin2mel(fft_bin)
+            if left < mbin < right:
+                matrix[m, fft_bin] = 1.0 - abs(center - mbin) / ms
+
+    return matrix
+
+
+class NemotronAudioFeatureExtractor(SequenceFeatureExtractor):
+    model_input_names = ["input_audio_embeds", "audio_embed_sizes", "audio_attention_mask"]
+
+    def __init__(self, audio_compression_rate=8, 
+                 audio_downsample_rate=1, 
+                 audio_feat_stride=1,
+                 feature_size = 80,
+                 sampling_rate = 16000,
+                 padding_value = 0.0,
+                 **kwargs):
+        
+        super().__init__(feature_size=feature_size, 
+                         sampling_rate=sampling_rate, 
+                         padding_value=padding_value, **kwargs)
+
+        self.compression_rate = audio_compression_rate
+        self.qformer_compression_rate = audio_downsample_rate
+        self.feat_stride = audio_feat_stride
+
+        self._eightk_method = "fillzero"
+        self._mel = speechlib_mel(self.sampling_rate, 512, self.feature_size, fmin=None, fmax=self.sampling_rate//2-self.feature_size-230).T
+
+        self._hamming400 = np.hamming(400)  # for 16k audio
+        self._hamming200 = np.hamming(200)  # for 8k audio
+
+    def duration_to_frames(self, duration):
+        """duration in s, estimated frames"""
+        frame_rate = 10
+
+        num_frames = duration * 1000 // frame_rate
+        return num_frames
+
+    def __call__(
+        self,
+        audios: List[AudioInput],
+        sampling_rate = 16000,
+        return_attention_mask=True,
+        padding="max_length",
+        return_tensors: Optional[Union[str, TensorType]] = None,
+    ):
+        # Ref: https://github.com/huggingface/transformers/blob/v4.47.0/src/transformers/models/audio_spectrogram_transformer/feature_extraction_audio_spectrogram_transformer.py#L161
+        returned_input_audio_embeds = []
+        returned_audio_embed_sizes = []
+        audio_frames_list = []
+
+        for audio_data in audios:
+            audio_embeds = self._extract_features(audio_data, sampling_rate)
+            audio_frames = len(audio_embeds) * self.feat_stride
+            audio_embed_size = self._compute_audio_embed_size(audio_frames)
+
+            returned_input_audio_embeds.append(torch.tensor(audio_embeds))
+            returned_audio_embed_sizes.append(torch.tensor(audio_embed_size).long())
+            audio_frames_list.append(audio_frames)
+
+        returned_input_audio_embeds = pad_sequence(
+            returned_input_audio_embeds, batch_first=True
+        )
+        returned_audio_embed_sizes = torch.stack(returned_audio_embed_sizes, dim=0)
+        audio_frames = torch.tensor(audio_frames_list)
+        returned_audio_attention_mask = torch.arange(0, audio_frames.max()).unsqueeze(0) < audio_frames.unsqueeze(1) if len(audios) > 1 else None
+
+        data = {
+            "input_audio_embeds": returned_input_audio_embeds,
+            "audio_embed_sizes": returned_audio_embed_sizes,
+        }
+        if returned_audio_attention_mask is not None and return_attention_mask:
+            data["audio_attention_mask"] = returned_audio_attention_mask
+
+        return BatchFeature(data=data, tensor_type=return_tensors)
+
+    def _extract_spectrogram(self, wav, fs):
+        """Extract spectrogram features from waveform.
+        Args:
+            wav (1D array): waveform of the input
+            fs (int): sampling rate of the waveform, 16000 or 8000.
+                If fs=8000, the waveform will be resampled to 16000Hz.
+        Output:
+            log_fbank (2D array): a TxD matrix of log Mel filterbank features.
+                D=80, and T is the number of frames.
+        """
+        if wav.ndim > 1:
+            wav = np.squeeze(wav)
+
+        # by default, we extract the mean if stereo
+        if len(wav.shape) == 2:
+            wav = wav.mean(1)
+
+        # Resample to 16000 or 8000 if needed
+        if fs > 16000:
+            wav = scipy.signal.resample_poly(wav, 1, fs // 16000)
+            fs = 16000
+        elif 8000 < fs < 16000:
+            wav = scipy.signal.resample_poly(wav, 1, fs // 8000)
+            fs = 8000
+        elif fs < 8000:
+            raise RuntimeError(f"Unsupported sample rate {fs}")
+
+        if fs == 8000:
+            if self._eightk_method == "resample":
+                # Input audio is 8 kHz. Convert to 16 kHz before feature
+                # extraction
+                wav = scipy.signal.resample_poly(wav, 2, 1)
+                fs = 16000
+            # Do nothing here for fillzero method
+        elif fs != 16000:
+            # Input audio is not a supported sample rate.
+            raise RuntimeError(f"Input data using an unsupported sample rate: {fs}")
+
+        preemphasis = 0.97
+
+        if fs == 8000:
+            n_fft = 256
+            win_length = 200
+            hop_length = 80
+            fft_window = self._hamming200
+        elif fs == 16000:
+            n_fft = 512
+            win_length = 400
+            hop_length = 160
+            fft_window = self._hamming400
+
+        # Spec 1: SpeechLib cut remaining sample insufficient for a hop
+        n_batch = (wav.shape[0] - win_length) // hop_length + 1
+        # Here we don't use stride_tricks since the input array may not satisfy
+        # memory layout requirement and we need writeable output
+        # Here we only use list of views before copy to desination
+        # so it is more efficient than broadcasting
+        y_frames = np.array(
+            [wav[_stride : _stride + win_length] for _stride in range(0, hop_length * n_batch, hop_length)],
+            dtype=np.float32,
+        )
+
+        # Spec 2: SpeechLib applies preemphasis within each batch
+        y_frames_prev = np.roll(y_frames, 1, axis=1)
+        y_frames_prev[:, 0] = y_frames_prev[:, 1]
+        y_frames = (y_frames - preemphasis * y_frames_prev) * 32768
+
+        S = np.fft.rfft(fft_window * y_frames, n=n_fft, axis=1).astype(np.complex64)
+
+        if fs == 8000:
+            # Need to pad the output to look like 16 kHz data but with zeros in
+            # the 4 to 8 kHz bins.
+            frames, bins = S.shape
+            padarray = np.zeros((frames, bins))
+            S = np.concatenate((S[:, 0:-1], padarray), axis=1)  # Nyquist bin gets set to zero
+
+        spec = np.abs(S).astype(np.float32)
+        return spec
+
+    def _extract_features(self, wav, fs):
+        """Extract log filterbank features from waveform.
+        Args:
+            wav (1D array): waveform of the input
+            fs (int): sampling rate of the waveform, 16000 or 8000.
+                If fs=8000, the waveform will be resampled to 16000Hz.
+        Output:
+            log_fbank (2D array): a TxD matrix of log Mel filterbank features.
+                D=80, and T is the number of frames.
+        """
+        spec = self._extract_spectrogram(wav, fs)
+        spec_power = spec**2
+
+        fbank_power = np.clip(spec_power.dot(self._mel), 1.0, None)
+        log_fbank = np.log(fbank_power).astype(np.float32)
+
+        return log_fbank
+
+    def _compute_audio_embed_size(self, audio_frames):
+        integer = audio_frames // self.compression_rate
+        remainder = audio_frames % self.compression_rate
+
+        result = integer if remainder == 0 else integer + 1
+
+        integer = result // self.qformer_compression_rate
+        remainder = result % self.qformer_compression_rate
+        result = integer if remainder == 0 else integer + 1  # qformer compression
+
+        return result
+
+class NemotronOmniProcessor(ProcessorMixin):
+    attributes = ["image_processor", "feature_extractor", "tokenizer"]
+    valid_kwargs = ["chat_template", "image_seq_length"]
+    image_processor_class = "AutoImageProcessor"
+    feature_extractor_class = "NemotronAudioFeatureExtractor"
+    tokenizer_class = "AutoTokenizer"
+
+    def __init__(
+        self,
+        image_processor,
+        feature_extractor,
+        tokenizer,
+        chat_template=None,
+        image_seq_length: int = 256,
+        **kwargs,
+    ):
+        self.image_seq_length = image_seq_length
+        self.image_token_id = -999#tokenizer.image_token_id
+        self.boi_token = ''#tokenizer.boi_token
+        self.image_token = ''#tokenizer.image_token
+        self.eoi_token=''
+        image_tokens_expanded = "".join([self.image_token] * image_seq_length)
+        self.full_image_sequence = f"\n\n{self.boi_token}{image_tokens_expanded}{self.eoi_token}\n\n"
+        
+        self.audio_token_id = 128255
+        self.boa_token = "<start_of_audio>"
+        self.eoa_token = "<end_of_audio>"
+        self.audio_token = "<audio_soft_token>"
+        
+        super().__init__(
+            image_processor=image_processor,
+            feature_extractor=feature_extractor,
+            tokenizer=tokenizer,
+            chat_template=chat_template,
+            **kwargs,
+        )
+
+    def __call__(
+        self,
+        images: ImageInput = None,
+        text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
+        videos=None,
+        audio: List[AudioInput] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+    ) -> BatchFeature:
+        if text is None and images is None:
+            raise ValueError("Provide at least one of `text` or `audio`.")
+
+
+        if isinstance(text, str):
+            text = [text]
+        elif not isinstance(text, list) and not isinstance(text[0], str):
+            raise ValueError("Invalid input text. Please provide a string, or a list of strings")
+
+    
+        audio_inputs = {}
+        if audio is not None:
+            full_audio_sequences = []
+            audio_inputs = self.feature_extractor(audio)
+            for i, embed_size in enumerate(audio_inputs.audio_embed_sizes):
+                audio_tokens_expanded = "".join([self.audio_token] * embed_size)
+                full_audio_sequence = f"\n\n{self.boa_token}{audio_tokens_expanded}{self.eoa_token}\n\n"
+                full_audio_sequences.append(full_audio_sequence)
+            
+            text = [prompt.replace(self.boa_token, audio_sequences) for (prompt, audio_sequences) in zip(text, full_audio_sequences)]
+
+        text_inputs = self.tokenizer(text=text,  return_tensors="np")
+
+        # Add token type ids manually, as tokenizer can't do arbitrary position token types
+        array_ids = text_inputs["input_ids"]
+        mm_token_type_ids = np.zeros_like(text_inputs["input_ids"])
+        mm_token_type_ids[array_ids == self.image_token_id] = 1
+        mm_token_type_ids[array_ids == self.audio_token_id] = 2
+
+        has_vision_ids = np.any(mm_token_type_ids == 1, axis=1)
+        has_audio_ids = np.any(mm_token_type_ids == 2, axis=1)
+
+        input_modes = (has_audio_ids << 1) | has_vision_ids
+
+        text_inputs = {k: v.tolist() for k, v in text_inputs.items()}  # in case user requested list inputs
+        text_inputs["token_type_ids"] = mm_token_type_ids.tolist()
+        text_inputs["input_modes"] = input_modes.tolist()
+        
+        return BatchFeature(data={**text_inputs, **audio_inputs}, tensor_type=return_tensors)
+
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.batch_decode with CLIP->Gemma
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to GemmaTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.decode with CLIP->Gemma
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to GemmaTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    @property
+    def model_input_names(self):
+        tokenizer_input_names = self.tokenizer.model_input_names + ["token_type_ids"]
+        image_processor_input_names = self.image_processor.model_input_names
+        return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
+
+AutoFeatureExtractor.register("NemotronAudioFeatureExtractor", NemotronAudioFeatureExtractor)

preprocessor_config.json

@@ -0,0 +1,29 @@
+{
+  "image_processor_type": "LlamaNemotronNanoVLImageProcessor",
+  "auto_map": {
+    "AutoImageProcessor": "image_processing.LlamaNemotronNanoVLImageProcessor"
+  },
+  "audio_compression_rate": 8,
+  "audio_downsample_rate": 1,
+  "audio_feat_stride": 1,
+  "compression_rate": 8,
+  "do_convert_rgb": null,
+  "do_normalize": true,
+  "do_pan_and_scan": null,
+  "do_rescale": true,
+  "do_resize": true,
+  "feat_stride": 1,
+  "feature_extractor_type": "NemotronAudioFeatureExtractor",
+  "processor_class": "NemotronOmniProcessor",
+  "feature_size": 80,
+  "padding_side": "right",
+  "padding_value": 0.0,
+  "pan_and_scan_max_num_crops": null,
+  "pan_and_scan_min_crop_size": null,
+  "pan_and_scan_min_ratio_to_activate": null,
+  "qformer_compression_rate": 1,
+  "resample": 2,
+  "rescale_factor": 0.00392156862745098,
+  "return_attention_mask": true,
+  "sampling_rate": 16000
+}

processor_config.json

@@ -0,0 +1,6 @@
+{
+  "auto_map": {
+    "AutoProcessor": "preprocessing_NemotronOmni.NemotronOmniProcessor"
+  },
+  "processor_class": "NemotronOmniProcessor"
+}

special_tokens_map.json

@@ -0,0 +1,37 @@
+{
+  "bos_token": {
+    "content": "<|begin_of_text|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|eot_id|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "boa_token": {
+    "content": "<start_of_audio>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eoa_token": {
+    "content": "<end_of_audio>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "audio_token": {
+    "content": "<audio_soft_token>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd2ec0ed10737c65e5a71158bdd29bc0f8c391462e7c78562f48c7e173d265d4
+size 17209878

tokenizer_config.json

@@ -0,0 +1,2063 @@
+{
+  "added_tokens_decoder": {
+    "128000": {
+      "content": "<|begin_of_text|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128001": {
+      "content": "<|end_of_text|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128002": {
+      "content": "<|reserved_special_token_0|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128003": {
+      "content": "<|reserved_special_token_1|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128004": {
+      "content": "<|finetune_right_pad_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128005": {
+      "content": "<|reserved_special_token_2|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128006": {
+      "content": "<|start_header_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128007": {
+      "content": "<|end_header_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128008": {
+      "content": "<|eom_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128009": {
+      "content": "<|eot_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128010": {
+      "content": "<|python_tag|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128011": {
+      "content": "<|reserved_special_token_3|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128012": {
+      "content": "<|reserved_special_token_4|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128013": {
+      "content": "<|reserved_special_token_5|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128014": {
+      "content": "<|reserved_special_token_6|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128015": {
+      "content": "<|reserved_special_token_7|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128016": {
+      "content": "<|reserved_special_token_8|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128017": {
+      "content": "<|reserved_special_token_9|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128018": {
+      "content": "<|reserved_special_token_10|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128019": {
+      "content": "<|reserved_special_token_11|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128020": {
+      "content": "<|reserved_special_token_12|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128021": {
+      "content": "<|reserved_special_token_13|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128022": {
+      "content": "<|reserved_special_token_14|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128023": {
+      "content": "<|reserved_special_token_15|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128024": {
+      "content": "<|reserved_special_token_16|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128025": {
+      "content": "<|reserved_special_token_17|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128026": {
+      "content": "<|reserved_special_token_18|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128027": {
+      "content": "<|reserved_special_token_19|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128028": {
+      "content": "<|reserved_special_token_20|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128029": {
+      "content": "<|reserved_special_token_21|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128030": {
+      "content": "<|reserved_special_token_22|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128031": {
+      "content": "<|reserved_special_token_23|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128032": {
+      "content": "<|reserved_special_token_24|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128033": {
+      "content": "<|reserved_special_token_25|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128034": {
+      "content": "<|reserved_special_token_26|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128035": {
+      "content": "<|reserved_special_token_27|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128036": {
+      "content": "<|reserved_special_token_28|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128037": {
+      "content": "<|reserved_special_token_29|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128038": {
+      "content": "<|reserved_special_token_30|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128039": {
+      "content": "<|reserved_special_token_31|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128040": {
+      "content": "<|reserved_special_token_32|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128041": {
+      "content": "<|reserved_special_token_33|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128042": {
+      "content": "<|reserved_special_token_34|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128043": {
+      "content": "<|reserved_special_token_35|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128044": {
+      "content": "<|reserved_special_token_36|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128045": {
+      "content": "<|reserved_special_token_37|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128046": {
+      "content": "<|reserved_special_token_38|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128047": {
+      "content": "<|reserved_special_token_39|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128048": {
+      "content": "<|reserved_special_token_40|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128049": {
+      "content": "<|reserved_special_token_41|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128050": {
+      "content": "<|reserved_special_token_42|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128051": {
+      "content": "<|reserved_special_token_43|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128052": {
+      "content": "<|reserved_special_token_44|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128053": {
+      "content": "<|reserved_special_token_45|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128054": {
+      "content": "<|reserved_special_token_46|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128055": {
+      "content": "<|reserved_special_token_47|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128056": {
+      "content": "<|reserved_special_token_48|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128057": {
+      "content": "<|reserved_special_token_49|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128058": {
+      "content": "<|reserved_special_token_50|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128059": {
+      "content": "<|reserved_special_token_51|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128060": {
+      "content": "<|reserved_special_token_52|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128061": {
+      "content": "<|reserved_special_token_53|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128062": {
+      "content": "<|reserved_special_token_54|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128063": {
+      "content": "<|reserved_special_token_55|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128064": {
+      "content": "<|reserved_special_token_56|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128065": {
+      "content": "<|reserved_special_token_57|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128066": {
+      "content": "<|reserved_special_token_58|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128067": {
+      "content": "<|reserved_special_token_59|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128068": {
+      "content": "<|reserved_special_token_60|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128069": {
+      "content": "<|reserved_special_token_61|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128070": {
+      "content": "<|reserved_special_token_62|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128071": {
+      "content": "<|reserved_special_token_63|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128072": {
+      "content": "<|reserved_special_token_64|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128073": {
+      "content": "<|reserved_special_token_65|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128074": {
+      "content": "<|reserved_special_token_66|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128075": {
+      "content": "<|reserved_special_token_67|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128076": {
+      "content": "<|reserved_special_token_68|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128077": {
+      "content": "<|reserved_special_token_69|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128078": {
+      "content": "<|reserved_special_token_70|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128079": {
+      "content": "<|reserved_special_token_71|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128080": {
+      "content": "<|reserved_special_token_72|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128081": {
+      "content": "<|reserved_special_token_73|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128082": {
+      "content": "<|reserved_special_token_74|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128083": {
+      "content": "<|reserved_special_token_75|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128084": {
+      "content": "<|reserved_special_token_76|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128085": {
+      "content": "<|reserved_special_token_77|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128086": {
+      "content": "<|reserved_special_token_78|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128087": {
+      "content": "<|reserved_special_token_79|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128088": {
+      "content": "<|reserved_special_token_80|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128089": {
+      "content": "<|reserved_special_token_81|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128090": {
+      "content": "<|reserved_special_token_82|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128091": {
+      "content": "<|reserved_special_token_83|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128092": {
+      "content": "<|reserved_special_token_84|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128093": {
+      "content": "<|reserved_special_token_85|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128094": {
+      "content": "<|reserved_special_token_86|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128095": {
+      "content": "<|reserved_special_token_87|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128096": {
+      "content": "<|reserved_special_token_88|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128097": {
+      "content": "<|reserved_special_token_89|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128098": {
+      "content": "<|reserved_special_token_90|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128099": {
+      "content": "<|reserved_special_token_91|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128100": {
+      "content": "<|reserved_special_token_92|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128101": {
+      "content": "<|reserved_special_token_93|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128102": {
+      "content": "<|reserved_special_token_94|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128103": {
+      "content": "<|reserved_special_token_95|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128104": {
+      "content": "<|reserved_special_token_96|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128105": {
+      "content": "<|reserved_special_token_97|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128106": {
+      "content": "<|reserved_special_token_98|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128107": {
+      "content": "<|reserved_special_token_99|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128108": {
+      "content": "<|reserved_special_token_100|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128109": {
+      "content": "<|reserved_special_token_101|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128110": {
+      "content": "<|reserved_special_token_102|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128111": {
+      "content": "<|reserved_special_token_103|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128112": {
+      "content": "<|reserved_special_token_104|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128113": {
+      "content": "<|reserved_special_token_105|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128114": {
+      "content": "<|reserved_special_token_106|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128115": {
+      "content": "<|reserved_special_token_107|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128116": {
+      "content": "<|reserved_special_token_108|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128117": {
+      "content": "<|reserved_special_token_109|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128118": {
+      "content": "<|reserved_special_token_110|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128119": {
+      "content": "<|reserved_special_token_111|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128120": {
+      "content": "<|reserved_special_token_112|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128121": {
+      "content": "<|reserved_special_token_113|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128122": {
+      "content": "<|reserved_special_token_114|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128123": {
+      "content": "<|reserved_special_token_115|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128124": {
+      "content": "<|reserved_special_token_116|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128125": {
+      "content": "<|reserved_special_token_117|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128126": {
+      "content": "<|reserved_special_token_118|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128127": {
+      "content": "<|reserved_special_token_119|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128128": {
+      "content": "<|reserved_special_token_120|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128129": {
+      "content": "<|reserved_special_token_121|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128130": {
+      "content": "<|reserved_special_token_122|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128131": {
+      "content": "<|reserved_special_token_123|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128132": {
+      "content": "<|reserved_special_token_124|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128133": {
+      "content": "<|reserved_special_token_125|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128134": {
+      "content": "<|reserved_special_token_126|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128135": {
+      "content": "<|reserved_special_token_127|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128136": {
+      "content": "<|reserved_special_token_128|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128137": {
+      "content": "<|reserved_special_token_129|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128138": {
+      "content": "<|reserved_special_token_130|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128139": {
+      "content": "<|reserved_special_token_131|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128140": {
+      "content": "<|reserved_special_token_132|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128141": {
+      "content": "<|reserved_special_token_133|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128142": {
+      "content": "<|reserved_special_token_134|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128143": {
+      "content": "<|reserved_special_token_135|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128144": {
+      "content": "<|reserved_special_token_136|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128145": {
+      "content": "<|reserved_special_token_137|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128146": {
+      "content": "<|reserved_special_token_138|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128147": {
+      "content": "<|reserved_special_token_139|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128148": {
+      "content": "<|reserved_special_token_140|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128149": {
+      "content": "<|reserved_special_token_141|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128150": {
+      "content": "<|reserved_special_token_142|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128151": {
+      "content": "<|reserved_special_token_143|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128152": {
+      "content": "<|reserved_special_token_144|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128153": {
+      "content": "<|reserved_special_token_145|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128154": {
+      "content": "<|reserved_special_token_146|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128155": {
+      "content": "<|reserved_special_token_147|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128156": {
+      "content": "<|reserved_special_token_148|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128157": {
+      "content": "<|reserved_special_token_149|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128158": {
+      "content": "<|reserved_special_token_150|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128159": {
+      "content": "<|reserved_special_token_151|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128160": {
+      "content": "<|reserved_special_token_152|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128161": {
+      "content": "<|reserved_special_token_153|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128162": {
+      "content": "<|reserved_special_token_154|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128163": {
+      "content": "<|reserved_special_token_155|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128164": {
+      "content": "<|reserved_special_token_156|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128165": {
+      "content": "<|reserved_special_token_157|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128166": {
+      "content": "<|reserved_special_token_158|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128167": {
+      "content": "<|reserved_special_token_159|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128168": {
+      "content": "<|reserved_special_token_160|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128169": {
+      "content": "<|reserved_special_token_161|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128170": {
+      "content": "<|reserved_special_token_162|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128171": {
+      "content": "<|reserved_special_token_163|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128172": {
+      "content": "<|reserved_special_token_164|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128173": {
+      "content": "<|reserved_special_token_165|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128174": {
+      "content": "<|reserved_special_token_166|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128175": {
+      "content": "<|reserved_special_token_167|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128176": {
+      "content": "<|reserved_special_token_168|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128177": {
+      "content": "<|reserved_special_token_169|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128178": {
+      "content": "<|reserved_special_token_170|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128179": {
+      "content": "<|reserved_special_token_171|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128180": {
+      "content": "<|reserved_special_token_172|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128181": {
+      "content": "<|reserved_special_token_173|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128182": {
+      "content": "<|reserved_special_token_174|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128183": {
+      "content": "<|reserved_special_token_175|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128184": {
+      "content": "<|reserved_special_token_176|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128185": {
+      "content": "<|reserved_special_token_177|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128186": {
+      "content": "<|reserved_special_token_178|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128187": {
+      "content": "<|reserved_special_token_179|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128188": {
+      "content": "<|reserved_special_token_180|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128189": {
+      "content": "<|reserved_special_token_181|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128190": {
+      "content": "<|reserved_special_token_182|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128191": {
+      "content": "<|reserved_special_token_183|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128192": {
+      "content": "<|reserved_special_token_184|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128193": {
+      "content": "<|reserved_special_token_185|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128194": {
+      "content": "<|reserved_special_token_186|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128195": {
+      "content": "<|reserved_special_token_187|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128196": {
+      "content": "<|reserved_special_token_188|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128197": {
+      "content": "<|reserved_special_token_189|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128198": {
+      "content": "<|reserved_special_token_190|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128199": {
+      "content": "<|reserved_special_token_191|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128200": {
+      "content": "<|reserved_special_token_192|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128201": {
+      "content": "<|reserved_special_token_193|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128202": {
+      "content": "<|reserved_special_token_194|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128203": {
+      "content": "<|reserved_special_token_195|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128204": {
+      "content": "<|reserved_special_token_196|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128205": {
+      "content": "<|reserved_special_token_197|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128206": {
+      "content": "<|reserved_special_token_198|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128207": {
+      "content": "<|reserved_special_token_199|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128208": {
+      "content": "<|reserved_special_token_200|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128209": {
+      "content": "<|reserved_special_token_201|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128210": {
+      "content": "<|reserved_special_token_202|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128211": {
+      "content": "<|reserved_special_token_203|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128212": {
+      "content": "<|reserved_special_token_204|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128213": {
+      "content": "<|reserved_special_token_205|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128214": {
+      "content": "<|reserved_special_token_206|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128215": {
+      "content": "<|reserved_special_token_207|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128216": {
+      "content": "<|reserved_special_token_208|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128217": {
+      "content": "<|reserved_special_token_209|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128218": {
+      "content": "<|reserved_special_token_210|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128219": {
+      "content": "<|reserved_special_token_211|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128220": {
+      "content": "<|reserved_special_token_212|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128221": {
+      "content": "<|reserved_special_token_213|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128222": {
+      "content": "<|reserved_special_token_214|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128223": {
+      "content": "<|reserved_special_token_215|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128224": {
+      "content": "<|reserved_special_token_216|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128225": {
+      "content": "<|reserved_special_token_217|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128226": {
+      "content": "<|reserved_special_token_218|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128227": {
+      "content": "<|reserved_special_token_219|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128228": {
+      "content": "<|reserved_special_token_220|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128229": {
+      "content": "<|reserved_special_token_221|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128230": {
+      "content": "<|reserved_special_token_222|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128231": {
+      "content": "<|reserved_special_token_223|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128232": {
+      "content": "<|reserved_special_token_224|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128233": {
+      "content": "<|reserved_special_token_225|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128234": {
+      "content": "<|reserved_special_token_226|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128235": {
+      "content": "<|reserved_special_token_227|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128236": {
+      "content": "<|reserved_special_token_228|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128237": {
+      "content": "<|reserved_special_token_229|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128238": {
+      "content": "<|reserved_special_token_230|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128239": {
+      "content": "<|reserved_special_token_231|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128240": {
+      "content": "<|reserved_special_token_232|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128241": {
+      "content": "<|reserved_special_token_233|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128242": {
+      "content": "<|reserved_special_token_234|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128243": {
+      "content": "<|reserved_special_token_235|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128244": {
+      "content": "<|reserved_special_token_236|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128245": {
+      "content": "<|reserved_special_token_237|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128246": {
+      "content": "<|reserved_special_token_238|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128247": {
+      "content": "<|reserved_special_token_239|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128248": {
+      "content": "<|reserved_special_token_240|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128249": {
+      "content": "<|reserved_special_token_241|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128250": {
+      "content": "<|reserved_special_token_242|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128251": {
+      "content": "<|reserved_special_token_243|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128252": {
+      "content": "<|reserved_special_token_244|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128253": {
+      "content": "<start_of_audio>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128254": {
+      "content": "<end_of_audio>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128255": {
+      "content": "<audio_soft_token>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<|begin_of_text|>",
+  "chat_template": "{%- if tools %}{{- '<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n' -}}{%- if messages[0].role == 'system' and messages[0].content != '' -%}{{- messages[0].content + '\n\n' -}}{%- else -%}{{- 'detailed thinking off\n\n' -}}{%- endif -%}{{- '<AVAILABLE_TOOLS>[' -}}{%- for tool in tools -%}{{- (tool.function if tool.function is defined else tool) | tojson -}}{{- ', ' if not loop.last else '' -}}{%- endfor -%}{{- ']</AVAILABLE_TOOLS>' -}}{{- '<|eot_id|>' -}}{%- else %}{{- '<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n' -}}{%- if messages[0].role == 'system' and messages[0].content != '' -%}{{- messages[0].content -}}{%- else -%}{{- 'detailed thinking off' -}}{%- endif %}{{- '<|eot_id|>' -}}{%- endif %}{%- for message in messages -%}{%- if (message.role == 'user') -%}{{- '<|start_header_id|>user<|end_header_id|>\n\n' + message.content + '<|eot_id|>' -}}{%- elif message.role == 'assistant' -%}{%- set content = message.content -%}{{- '<|start_header_id|>assistant<|end_header_id|>\n\n' + content -}}{%- if message.tool_calls -%}{{- '<TOOLCALL>[' -}}{%- for tool_call in message.tool_calls -%}{%- if tool_call.function -%}{%- set tool_call = tool_call.function -%}{%- endif -%}{{- '{\"name\": \"' }}{{- tool_call.name }}{{- '\", \"arguments\": ' -}}{%- if tool_call.arguments is string -%}{{- tool_call.arguments -}}{%- else -%}{{- tool_call.arguments | tojson -}}{%- endif -%}{{- ', ' if not loop.last else '' -}}{%- endfor -%}{{- ']</TOOLCALL>' -}}{%- endif %}{{- '<|eot_id|>' -}}{%- elif message.role == 'tool' -%}{%- if loop.first or (messages[loop.index0 - 1].role != 'tool') -%}{{- '<|start_header_id|>user<|end_header_id|>\n\n' }}{{- '<TOOL_RESPONSE>[' -}}{%- endif -%}{{- message.content -}}{{- ', ' if not loop.last and (messages[loop.index0 + 1].role == 'tool') else '' -}}{%- if loop.last or (messages[loop.index0 + 1].role != 'tool') -%}{{- ']</TOOL_RESPONSE>' -}}{{- '<|eot_id|>' -}}{%- endif %}{%- endif %}{%- endfor %}{%- if add_generation_prompt %}{{- '<|start_header_id|>assistant<|end_header_id|>\n\n' -}}{%- endif %}",
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "<|eot_id|>",
+  "extra_special_tokens": {},
+  "model_input_names": [
+    "input_ids",
+    "attention_mask"
+  ],
+  "model_max_length": 131072,
+  "tokenizer_class": "PreTrainedTokenizerFast"
+}

training.py

@@ -0,0 +1,398 @@
+import datasets
+datasets.config.DOWNLOADED_DATASETS_PATH = "/mnt/jeff/huggingface/data"
+import os
+os.environ['HF_HOME'] = '/mnt/jeff/huggingface'
+
+import argparse
+import json
+import os
+from pathlib import Path
+
+import numpy as np
+import torch
+import sacrebleu
+
+from datasets import load_dataset
+from torch.utils.data import Dataset, ConcatDataset
+from tqdm import tqdm
+from transformers import (
+    AutoProcessor,
+    AutoModel,
+    BatchFeature,
+    Trainer,
+    TrainingArguments,
+    StoppingCriteria,
+    StoppingCriteriaList,
+)
+from collections import defaultdict
+
+import soundfile as sf
+from datasets import Audio
+import random
+from ASRDataset import *
+
+
+def count_parameters_by_module(model):
+    # dictionary for parameters number by modules
+    module_params = defaultdict(lambda: {"total": 0, "trainable": 0})
+    
+    # all params
+    total_params = 0
+    total_trainable_params = 0
+    
+    # Check Embedding Token masks
+    embedding_masks = {}
+    for name, param in model.named_parameters():
+        if 'embed_tokens.weight' in name and hasattr(param, '_backward_hooks') and param._backward_hooks:
+            # check if params has embedding_grad_mask_hook
+            for hook_id, hook_fn in param._backward_hooks.items():
+                if hook_fn.__code__.co_name == 'embedding_grad_mask_hook':
+                    # Accessing mask variables in the closure of hook functions
+                    for cell in hook_fn.__closure__ or []:
+                        if isinstance(cell.cell_contents, torch.Tensor) and cell.cell_contents.dtype == torch.bool:
+                            # check mask tensor
+                            embedding_masks[name] = ~cell.cell_contents  # True : Trainable
+                 
+    # Count params by modules
+    for name, param in model.named_parameters():
+        # extracts top module_name
+        module_name = name.split('.')[0]
+        param_count = param.numel()
+        
+        module_params[module_name]["total"] += param_count
+        total_params += param_count
+        
+        if param.requires_grad:
+            # Only count for real trainable params. (with masks)
+            if name in embedding_masks:
+                trainable_count = embedding_masks[name].sum().item()
+                module_params[module_name]["trainable"] += trainable_count
+                total_trainable_params += trainable_count
+            else:
+                module_params[module_name]["trainable"] += param_count
+                total_trainable_params += param_count
+    
+    print(f"All Params: {total_params:,}")
+    print(f"Trainable Params: {total_trainable_params:,} ({total_trainable_params/total_params*100:.2f}%)")
+    print("\nParams by Module:")
+    
+    for module_name, counts in sorted(module_params.items()):
+        trainable_percentage = counts["trainable"] / counts["total"] * 100 if counts["total"] > 0 else 0
+        total_percentage = counts["total"] / total_params * 100
+        
+        print(f"- {module_name}:")
+        print(f"  Total: {counts['total']:,} ({total_percentage:.2f}% of model)")
+        print(f"  Trainable: {counts['trainable']:,} ({trainable_percentage:.2f}% of module)")
+    
+    return module_params
+
+def create_model(model_name_or_path, revision="main", use_flash_attention = False):
+    model = AutoModel.from_pretrained(
+        model_name_or_path,
+        revision=revision,
+        torch_dtype=torch.bfloat16,
+        device_map="auto",
+        attn_implementation="flash_attention_2" if use_flash_attention else "eager",
+        trust_remote_code=True,
+    )
+    
+    # Set use_cache to False after model loaded
+    model.config.use_cache = False
+
+    # Freeze all parameters
+    for param in model.parameters():
+        param.requires_grad = False
+    
+    model.set_lora_adapter('speech')
+    model.to(torch.bfloat16)
+    
+    # (Optional) unfreeze audio_tower parameters
+    for param in model.audio_tower.parameters():
+       param.requires_grad = True
+
+    # Only unfreeze audio_projector parameters
+    for param in model.audio_projector.parameters():
+        param.requires_grad = True
+
+    # (Optional) unfreeze audio embed_tokens
+    train_embed = True
+    if train_embed:
+        embed_tokens = model.language_model.model.model.embed_tokens
+        
+        embed_tokens.weight.requires_grad = False
+
+        # Added Speech token IDs (only this tokens be trainable)
+        trainable_token_ids = [128253, 128254]
+
+        embed_tokens.weight.requires_grad = True
+        mask = torch.ones_like(embed_tokens.weight, dtype=torch.bool)
+        mask[trainable_token_ids] = False  # Trainable Tokens are False (unfreeze), else True (freeze)
+
+        # backward hook, with gradient masking
+        def embedding_grad_mask_hook(grad):
+            return grad.masked_fill(mask, 0)
+
+        embed_tokens.weight.register_hook(embedding_grad_mask_hook)
+
+        model.language_model.model.model.embed_tokens = embed_tokens
+        
+    count_parameters_by_module(model)
+
+    return model
+
+model_name_or_path = '/mnt/jeff/InCar/LlamaNemotronOmni/Llama-NemotronOmni'
+use_flash_attention = False
+
+output_dir = '../nemotron_tmp3'
+batch_size = 128
+batch_size_per_gpu = 4
+learning_rate = 4.0e-5 # 1.0e-4 for fine-tuning
+wd = 0.01
+num_train_epochs = 10
+
+revision = "main" #"v1.0"
+
+processor = AutoProcessor.from_pretrained(
+    model_name_or_path,
+    revision=revision,
+    trust_remote_code=True,
+)
+
+model = create_model(
+    model_name_or_path,
+    revision=revision,
+    use_flash_attention=use_flash_attention,
+)
+
+train_datasets = []
+
+custom_yating_tw_loc = TWCostumData(processor=processor,
+                      csv_path='/mnt/jeff/InCar/data/tw_data/taiwan_location-srdc_tts-20250505-yating-1-2s-breezyvoice.csv')
+train_datasets.append(custom_yating_tw_loc)  # 9458
+
+custom_tw_loc = TWCostumData(processor=processor,
+                      csv_path='/mnt/jeff/InCar/data/tw_data/taiwan_location-srdc_tts-20250509-common_voice_16_1-TW.csv')
+train_datasets.append(custom_tw_loc) # 1500
+
+custom_tw_loc2 = TWCostumData(processor=processor,
+                      csv_path='/mnt/jeff/InCar/data/tw_data/taiwan_location-srdc_tts-20250529-common_voice_16_1-TW.csv')
+train_datasets.append(custom_tw_loc2) # 9458
+
+# custom_yating_tw_road = TWCostumData(processor=processor,
+#                       csv_path='/mnt/jeff/InCar/data/tw_data/taiwan_road-srdc_tts-20250430-yating-1-2s-breezyvoice.csv')
+# train_datasets.append(custom_yating_tw_road) # 35224
+
+custom_tw_road = TWCostumData(processor=processor,
+                      csv_path='/mnt/jeff/InCar/data/tw_data/taiwan_road-srdc_tts-20250509-common_voice_16_1-TW.csv')
+train_datasets.append(custom_tw_road) # 1500
+
+custom_tw_road2 = TWCostumData(processor=processor,
+                      csv_path='/mnt/jeff/InCar/data/tw_data/taiwan_road-srdc_tts-20250529-common_voice_16_1-TW.csv')
+train_datasets.append(custom_tw_road2) # 35224
+
+
+
+# common voice asr
+commonvoice_speech_tw2 = CommonVoiceDataset(  # 45689
+    processor=processor,
+    source_lang="zh-TW",
+    split="other"
+)
+train_datasets.append(commonvoice_speech_tw2)
+
+commonvoice_speech_cn = CommonVoiceDataset(  #28868
+    processor=processor,
+    source_lang="zh-CN",
+    split="train[:10%]"
+)
+train_datasets.append(commonvoice_speech_cn)
+
+
+commonvoice_speech_tw = CommonVoiceDataset(  #6812
+    processor=processor,
+    source_lang="zh-TW",
+    split="train"
+)
+train_datasets.append(commonvoice_speech_tw)
+
+
+
+
+# Libri Speech Clean ASR mode (English -> English text)
+libri_speech_clean = LibriSpeechDataset(  #103781
+    processor=processor,
+    subset="clean",
+    split="train.360"
+)
+train_datasets.append(libri_speech_clean)
+
+
+# Fleurs ASR mode (English -> English text)
+en_asr_fleurs = FleursDataset(  #2550
+    processor=processor,
+    split="train",
+    source_lang="en_us",  # English
+    mode="asr"
+)
+train_datasets.append(en_asr_fleurs)
+
+ch_asr_fleurs = FleursDataset(  #3152
+    processor=processor,
+    split="train",
+    source_lang="cmn_hans_cn",
+    mode="asr"
+)
+train_datasets.append(ch_asr_fleurs)
+
+
+
+
+
+# en_ch_ast_fleurs = FleursDataset(
+#     processor=processor,
+#     split="train",
+#     source_lang="en_us",
+#     target_lang="cmn_hans_cn",
+#     mode="ast"
+# )
+# train_datasets.append(en_ch_ast_fleurs)
+
+# ch_en_ast_fleurs = FleursDataset(
+#     processor=processor,
+#     split="train",
+#     source_lang="cmn_hans_cn",
+#     target_lang="en_us",
+#     mode="ast"
+# )
+# train_datasets.append(ch_en_ast_fleurs)
+
+print("Count Num of Datasets", len(train_datasets))
+print([len(dataset) for dataset in train_datasets])
+
+# ConcatDataset
+train_dataset = ConcatDataset(train_datasets) if len(train_datasets) > 1 else train_datasets[0]
+print("Count Length of Datas", len(train_dataset))
+
+
+
+# Check GPUs
+num_gpus = torch.cuda.device_count()
+print(f'training on {num_gpus} GPUs')
+
+assert (
+    batch_size % (num_gpus * batch_size_per_gpu) == 0
+), 'Batch size must be divisible by the number of GPUs'
+gradient_accumulation_steps = batch_size // (num_gpus * batch_size_per_gpu)
+
+# hard coded training args
+dp_config = {
+    "fp16": {
+      "enabled": "auto",
+      "loss_scale": 0,
+      "loss_scale_window": 1000,
+      "initial_scale_power": 16,
+      "hysteresis": 2,
+      "min_loss_scale": 1
+  },
+   "zero_optimization": {
+       "stage": 2,
+       "allgather_partitions": True,
+       "allgather_bucket_size": 5e8,
+       "overlap_comm": False,
+       "reduce_scatter": True,
+       "reduce_bucket_size": 5e8,
+       "contiguous_gradients": True,
+       "cpu_offload": True
+   },
+
+   "train_batch_size": "auto",
+    "gradient_accumulation_steps": "auto",
+    "optimizer": {
+        "type": "AdamW",
+        "params": {
+          "lr": "auto",
+          "betas": 'auto',
+          "eps": 'auto',
+          "weight_decay": "auto"
+        }
+    },
+    "scheduler": {
+      "type": "WarmupDecayLR",
+      "params": {
+          "warmup_min_lr": "auto",
+          "warmup_max_lr": "auto",
+          "warmup_num_steps": "auto",
+          "total_num_steps": "auto"
+      }
+  },
+    "gradient_clipping": 1.0,
+    "zero_optimization": {
+        "stage": 0
+    }
+}
+training_args = TrainingArguments(
+    num_train_epochs=num_train_epochs,
+    per_device_train_batch_size=batch_size_per_gpu,
+    gradient_checkpointing=False,
+    gradient_checkpointing_kwargs={'use_reentrant': False},
+    gradient_accumulation_steps=gradient_accumulation_steps,
+    optim='adamw_torch',
+    adam_beta1=0.9,
+    adam_beta2=0.95,
+    adam_epsilon=1e-7,
+    learning_rate=learning_rate,
+    weight_decay=wd,
+    max_grad_norm=1.0,
+    lr_scheduler_type='cosine',
+    warmup_steps=50,
+    logging_steps=10,
+    output_dir=output_dir,
+    save_total_limit=10,
+    save_only_model=True,
+    bf16=True,
+    fp16=False,
+    remove_unused_columns=False,
+    report_to='none',
+    deepspeed=None,
+    disable_tqdm=False,
+    dataloader_num_workers=16,
+    save_strategy='epoch',
+    # save_steps=2500,
+    ddp_find_unused_parameters=True,
+
+)
+
+out_path = Path(training_args.output_dir)
+out_path.mkdir(parents=True, exist_ok=True)
+
+# create optimizer only for trainable params
+optimizer = torch.optim.AdamW(
+    filter(lambda p: p.requires_grad, model.parameters()),
+    lr=learning_rate,
+    weight_decay=wd,
+    betas=(0.9, 0.95),
+    eps=1e-7,
+)
+
+# Trainer Setting
+trainer = Trainer(
+    model=model,
+    args=training_args,
+    data_collator=covost_collate_fn,
+    train_dataset=train_dataset,
+    optimizers=(optimizer, None)
+)
+
+trainer.train()
+
+
+# # 1. Save LoRA Adapter
+model.language_model.model.save_pretrained(output_dir)
+
+# # 1-1. Delete Markdown file
+# markdown_file = os.path.join(output_dir, "README.md")
+# if os.path.exists(markdown_file):
+#     os.remove(markdown_file)
+
+# 2. Save entire model
+model.save_pretrained(output_dir)